Jia-Jen Lee

Artificial Intelligence and Robotics Innovation CenterChina Medical University Hospital, Taichung, Taiwan

Shih-Yen Hou

Artificial Intelligence and Robotics Innovation CenterChina Medical University Hospital, Taichung, Taiwan

Kee Koon Ng

Division of Cardiovascular Medicine, Department of Internal MedicineChina Medical University Hospital, Taichung, Taiwan

Wei-Chun Wang

Artificial Intelligence and Robotics Innovation CenterChina Medical University Hospital, Taichung, Taiwan

Department of Neurology, China Medical University HospitalChina Medical University, Taichung, Taiwan

Neuroscience and Brain Disease Center, China Medical University, Taichung, Taiwan

Shih-Sheng Chang

Artificial Intelligence and Robotics Innovation CenterChina Medical University Hospital, Taichung, Taiwan

Division of Cardiovascular Medicine, Department of Internal MedicineChina Medical University Hospital, Taichung, Taiwan

School of Medicine, China Medical University, Taichung, Taiwan

Abstract

Invasive coronary angiography (CAG) is the gold standard for diagnosing coronary artery disease, but interpretation varies substantially among observers. Existing AI systems can improve consistency but lack auditable decision processes and are limited in comprehensive open-ended assessment, undermining clinician trust and clinical adoption readiness. We developed CARDEA, a unified large vision-language model that serves as the inference core of a CAG pipeline. It was trained solely on public datasets and closed-ended tasks in three stages: visual feature alignment, a self-distilled Chain-of-Box (CoB) cold start, and reinforcement learning with verifiable rewards (RLVR) with a CoB reward encouraging bounding-box use in the reasoning trace. We assessed its two study-level diagnoses, dominance classification and complexity assessment, against a dedicated classifier and two interventional cardiologists. Report generation was excluded from training and evaluated zero-shot across stages on an external cohort using vessel-severity macro-$F_1$. CARDEA trailed the classifier on in-distribution dominance but drew level under domain shift (accuracy, 0.91 [95% confidence interval (CI), 0.86 to 0.95]) and was comparable to the cardiologists on complexity assessment (accuracy, 0.90 [CI, 0.82 to 0.97]). Only RLVR improved zero-shot report generation, raising its vessel-severity macro-$F_1$ (0.686 [CI, 0.664 to 0.707]) above the untuned base model (0.513) and over twice the always-normal floor (0.312). CARDEA runs an end-to-end CAG pipeline from raw multi-view videos through keyframe selection to study-level diagnosis while exposing auditable spatial evidence behind its conclusions. RLVR on verifiable closed-ended tasks surfaced open-ended reporting ability that supervised imitation did not. Clinical use requires prospective validation against expert cardiologists.

Keywords: Coronary Angiography, End-to-End Pipeline, Large Vision-Language Models, Reinforcement Learning with Verifiable Rewards, Auditable Reasoning, Visual Grounding.

1 Introduction

Invasive coronary angiography (CAG) remains the gold standard for diagnosing coronary artery disease [1]. However, projecting 3D coronary structures onto a 2D plane introduces geometric distortions such as vessel overlap and foreshortening, which can lead to underestimation of lesion severity [2]. Because acquiring additional projections increases contrast exposure and radiation, clinicians must mentally integrate the available views of each lesion [3]. Visual interpretation remains subject to inter-observer variability, with a recent study reporting 77.4% overall agreement among three experienced cardiologists reading the same angiograms [4].

Deep learning has been applied to reduce this variability, with early single-view models analyzing individual angiographic projections [5, 6]; cascaded pipelines like CathAI [7] and DeepCoro [8] chained several such modules, but remain vulnerable to compounding errors across them. To address this, foundation models such as DeepCORO-CLIP [9] pretrain on video-text pairs and fuse multiple views for study-level analysis. Yet these models are discriminative, outputting only closed-ended labels or spatial coordinates. Recent work including Nakamura et al. [10] and Jiang et al. [11] has begun applying large vision-language models (LVLMs) [12] to CAG for free-text diagnosis or reporting. However, such open-ended drafts are clinically useful only when easy to verify; pairing a diagnosis with spatial anchors lowers verification cost [13] while raising clinician trust [14]. In practice, neither does this within the report. Nakamura et al.’s model cannot output bounding boxes, and Jiang et al.’s appear only in sub-tasks separate from the report. Existing systems thus leave a gap: they either remain discriminative or generate open-ended narratives without auditable traces that let clinicians inspect a model’s logic.

Reinforcement learning with verifiable rewards (RLVR), exemplified by DeepSeek-R1 [15], elicits reasoning traces that add interpretability. But an ordinary text-only trace does not reveal which anatomical structures the model attends to; grounded-reasoning work therefore embeds bounding boxes as spatial anchors within the reasoning trace [16]. A related study brings this to medical imaging and terms this mechanism Chain-of-Box (CoB) [17], letting a human audit how each conclusion was reached.

We present CARDEA, a unified LVLM that runs an end-to-end CAG pipeline, from raw multi-view sequences through keyframe selection to study-level diagnosis, coupling competitive diagnostic performance with auditable CoB reasoning. Trained only on closed-ended tasks from public datasets [18, 19, 20, 21] through three stages—visual feature alignment, a self-distilled CoB cold start, and RLVR with a CoB reward—we evaluate whether this competence generalizes zero-shot to a fully held-out cohort, including open-ended report generation. We also examine whether policy optimization surfaces clinical reasoning that supervised imitation does not. The model weights and inference code are released for reproducibility.

2 Methods

CARDEA interprets a full CAG study through a two-pass pipeline (Figure 1). A first pass runs single-view inference on each angiographic video to select representative keyframes and classify their view; a second pass performs study-level multi-image CoB reasoning over the curated keyframes. The first pass filters raw frames because processing them all with a large model is computationally prohibitive, and many are non-diagnostic owing to poor cardiac alignment or insufficient contrast. It therefore discards these frames and retains a compact set covering the major left and right coronary views.

媒体内容 · 前往原文查看

2.1 Datasets and Tasks

To equip CARDEA to perform every task of the two-pass pipeline itself and to ground its reasoning in CoB evidence, we reformulated five public datasets into instructional tasks; full preprocessing and splits are in Appendix S1.1. To align the model with foundational CAG features and let it emit meaningful bounding boxes within its reasoning, ARCADE [19] (3,000 keyframes from 1,500 patients with diverse equipment) supplies two single-view tasks, Vessel Detection (25 coronary segments based on the SYNTAX score [22]) and Stenosis Detection ($\phantom{}\ge 50\%$ diameter stenosis). For the single-view first pass, which curates diagnostic frames, CADICA [18] (multi-view videos from 42 patients) supplies Keyframe Selection. The first pass then performs View Classification, for which ARCADE’s vessel annotations supply the left coronary artery (LCA) and right coronary artery (RCA) classes, while non-diagnostic CADICA frames and non-CAG medical images from PubMedVision [23] supply the OTHER class. For the second pass, which produces the study-level diagnoses, CoronaryDominance [20] (1,574 studies) supplies Dominance Classification (Left vs. Right, by the SYNTAX definition), and CardioSyntax [21] (1,844 studies) supplies Complexity Assessment, which we define by discretizing the continuous SYNTAX score into normal-to-intermediate ($0$–$32$) vs. high ($\phantom{}>32$) complexity.

AngioCAD [24] is excluded from all training and serves as our held-out zero-shot benchmark. It supplies Multi-Frame View Classification (LCA vs. RCA), Multi-Frame RCA Binary Stenosis Classification (Lesion vs. Non-lesion), and open-ended Report Generation over the four major branches: left main (LM), left anterior descending (LAD), left circumflex (LCX), and RCA.

2.2 Model and Training

CARDEA is built on Qwen3-VL-30B-A3B-Thinking [25] and trained in three stages: two of supervised fine-tuning (SFT), followed by RLVR. Configuration and hyperparameters are in Appendix S1.3. The first stage aligns the model from the general domain to CAG. We fine-tune it on the single-view tasks, teaching it to localize vessels and stenoses with bounding boxes. The second stage builds on the first, extending the model’s perception into its reasoning. However, CoB is not native to the base model, and hand-annotating reasoning traces is costly. We therefore synthesize the cold-start data by self-distillation [26] on dominance classification. The untuned base model serves as the teacher, generating the CoB reasoning traces from the Stage 1 model’s boxes and a textual dominance decision guide (Figure 2). The third stage uses RLVR to push accuracy on study-level tasks and to strengthen the CoB behavior seeded by the cold start. Reward functions are defined in Appendix S1.4. For study-level tasks, CoB behavior earns an extra reward conditional on a correct final answer. We restrict it to the study level, because these diagnoses draw a single conclusion from several views and therefore need CoB to explain how they reach that conclusion from local features across the views. Conversely, a single-view task involves no cross-view synthesis and needs no such explanation. We deliberately exclude open-ended tasks like report generation from RLVR. Such tasks have no rule-based verifier, so they would need a reward model, which is prone to reward hacking [15].

媒体内容 · 前往原文查看

2.3 Metrics and Baselines

We choose evaluation metrics by task type. Classification uses accuracy and Macro $F_1$, with a target-class $F_1$ for the binary AngioCAD tasks to match with the baselines’ report; detection uses instance-level $F_1$@IoU$\ge$0.5. Two tasks use custom metrics: keyframe selection by a mean frame distance ($D_k$), and report generation by Vessel Severity Macro-$F_1$ (VS-$F_1$) in two-class and three-class forms, where MedGemma-27B-IT [27] parses each free-text report into per-vessel severity labels. ROUGE-L [28] and BERTScore [29] serve as reference-based text-similarity metrics. Metric definitions are in Appendix S1.2.

Baselines are published dedicated models evaluated on the same test sets [30, 8, 31, 20, 24], named per task in Tables 1 and 2, with details in their footnotes. The exception is complexity assessment, where we compare against two interventional cardiologists with 10 and 3 years of experience [21]. Keyframe selection has no published baseline, so we report CARDEA’s value alone.

Some tasks are reported under several settings. For stenosis detection, we additionally report $F_1$@(IoU$\ge$0.5 or IoP$\ge$0.6), the relaxed matching criterion of Jiang et al.’s LVLM [11], which reduces sensitivity to differences in box extent and makes our value comparable with theirs (Appendix S1.2.2). For vessel detection, we also report an 11-segment result matching DeepCoro’s original convention. For view classification, CARDEA trains on three classes (including OTHER), but the baseline was evaluated only on LCA and RCA frames, so we add a two-class result to match. For dominance classification, we report on two official subsets: in-distribution Real Distribution (clinical class imbalance) and out-of-distribution Domain Shift (distinct imaging equipment). On held-out AngioCAD, to gauge the pipeline’s view filtering, we report the full cohort ($n=412$) and a valid-views subset ($n=326$) retaining studies with both LCA and RCA views identified by CARDEA; the subset is for reference only because the baselines were not evaluated on it.

For every diagnostic metric we report a 95% bootstrap confidence interval (CI), whereas for text-similarity metrics we report point estimates only. Two estimates are distinguishable when their CIs do not overlap, or a bare baseline estimate falls outside CARDEA’s CI, and otherwise comparable, though not necessarily equivalent. Because these CIs are not adjusted for multiple comparisons, all distinctions are exploratory. Two further caveats apply. First, two test sets are small (complexity, $n=60$; keyframe selection, 48 videos from 5 patients) and may be underpowered. Second, two comparisons are not strictly head-to-head: DeepCoro’s Algorithm 4 outputs segmentation masks that we converted to bounding boxes, possibly understating its detection score, and the cardiologists’ binary labels come from a post-hoc binarization of continuous SYNTAX scores, not a task they performed natively.

3 Results

3.1 Closed-Ended Diagnostics

Throughout, CARDEA denotes our final model in thinking mode, which enables explicit CoB reasoning. On the tasks it was trained on (Table 1), CARDEA was competitive with specialized baselines. In single-view detection it reached 0.37 on stenosis (vs. 0.36 for DCA-YOLOv8) and 0.50 on vessel detection (vs. 0.47 for DeepCoro’s Algorithm 4). It reached a Macro $F_1$ of 0.99 on both the 3-class and 2-class view tasks (the latter vs. 1.00 for YOLOv8x-cls), and a mean frame distance of 0.60 on keyframe selection, placing its predicted optimal frame within one frame of the expert-annotated usable range on average, though this split (48 videos, 5 patients) is preliminary. At the study level, dominance classification was lower than the 2D ConvNeXt on the Real Distribution subset (accuracy, 0.94 vs. 0.97; Macro $F_1$, 0.88 vs. 0.94) but comparable under Domain Shift (accuracy, 0.91 vs. 0.89; Macro $F_1$, 0.89 vs. 0.88). On complexity assessment CARDEA matched two cardiologists in accuracy (0.90 vs. 0.90 and 0.88), though its Macro $F_1$ point estimate was slightly lower (0.80 vs. 0.83 and 0.81). Across these closed-ended tasks, CARDEA’s only statistically distinguishable shortfall was Real Distribution dominance.

In contrast to the task-specific baselines, we additionally compared CARDEA with Jiang et al.’s LVLM [11]. For stenosis detection, CARDEA’s $F_1$@(IoU$\ge$0.5 or IoP$\ge$0.6) of 0.60 matched their 0.60, up from its $F_1$@IoU$\ge$0.5 of 0.37. For vessel detection, its $F_1$@IoU$\ge$0.5 of 0.50 exceeded their 0.46.

媒体内容 · 前往原文查看

Level	Task	Metric	CARDEA	Baseline / Expert
Frame	Stenosis Det.	$F_1$@IoU$\ge$0.5	0.37 (0.32–0.41)	$0.36\pm 0.08$ a
		$F_1$@(IoU$\ge$0.5 or IoP$\ge$0.6)	0.60 (0.56–0.65)	0.60 f
	Vessel Det.	$F_1$@IoU$\ge$0.5, 25 seg.	0.50 (0.48–0.53)	0.47 (0.45–0.50) b, $†$
		$F_1$@IoU$\ge$0.5, 25 seg.	0.50 (0.48–0.53)	0.46 f
		$F_1$@IoU$\ge$0.5, 11 seg.	0.58 (0.54–0.61)	0.57 (0.54–0.61) b, $†$
	View Cls. (3-class)	Accuracy	0.99 (0.98–1.00)	—
		Macro $F_1$	0.99 (0.98–1.00)	—
	View Cls. (2-class)	Accuracy	0.99 (0.98–1.00)	1.00 c
		Macro $F_1$	0.99 (0.98–1.00)	1.00 c
	Keyframe Sel.	$D_k$ ($\downarrow$)	0.60 (0.23–1.08)	—
Study	Dominance (Real)	Accuracy	0.94 (0.92–0.96)	0.97 d
		Macro $F_1$	0.88 (0.82–0.92)	0.94 d
	Dominance (Shift)	Accuracy	0.91 (0.86–0.95)	0.89 d
		Macro $F_1$	0.89 (0.83–0.94)	0.88 d
	Complexity Assess.	Accuracy	0.90 (0.82–0.97)	E1: 0.90 (0.82–0.97) e, $‡$ E2: 0.88 (0.80–0.95)
		Macro $F_1$	0.80 (0.63–0.93)	E1: 0.83 (0.68–0.94) e, $‡$ E2: 0.81 (0.66–0.92)

Entries are point estimates with 95% confidence intervals in parentheses, obtained by recomputing each metric across 5,000 case-level bootstrap resamples. A baseline is shown as a point estimate alone where the source reported no interval and its per-case predictions were unavailable, so no interval could be computed here; the one exception is DCA-YOLOv8, whose source-reported uncertainty is given as $\pm$ (footnote a).

Macro $F_1$, unweighted mean of per-class $F_1$; $F_1$@IoU$\ge$0.5, instance-level detection $F_1$ at an IoU threshold of 0.5; IoP, intersection over prediction; 25 seg. and 11 seg., the 25 ARCADE-defined coronary segments and DeepCoro’s original 11-segment convention; $D_k$, mean distance (in frames) from the predicted keyframe to the expert-annotated usable range, lower being better. OTHER, non-diagnostic or non-CAG frames; View Cls. (3-class) covers LCA/RCA/OTHER and the 2-class variant LCA/RCA only. Real and Shift are the official in-distribution (clinical class imbalance) and out-of-distribution (distinct imaging equipment) CoronaryDominance test subsets.

^a DCA-YOLOv8, tuned by CMA-ES, the best optimizer on the large backbone [30]. The uncertainty is the source-reported 95% CI from stratified 3-fold cross-validation, not a case-level bootstrap CI.

^b DeepCoro’s Algorithm 4, an ensemble of seven vessel-segmentation models [8].

^c YOLOv8x-cls, a YOLO model specifically trained for view classification [31].

^d 2D ConvNeXt, classifying frames independently with majority-vote aggregation [20].

^e E1 / E2 = Expert 1 (10 yrs) / Expert 2 (3 yrs), interventional cardiologists [21].

^f Jiang et al.’s large vision-language model [11]. Their relaxed criterion additionally accepts a predicted box whose intersection over prediction with an unmatched ground-truth box is $\phantom{}\ge 0.6$ (Appendix S1.2.2); following their rationale, we apply it to stenosis only, where the location of a box matters more than its exact extent. Their vessel score uses the standard $F_1$@IoU$\ge$0.5.

^$†$ Algorithm 4’s masks were converted to bounding boxes for this detection score, possibly understating it. In both the 25- and 11-segment comparisons, CARDEA and DeepCoro were evaluated over the same segments.

^$‡$ Derived by binarizing the experts’ continuous SYNTAX scores at the $\phantom{}\le 32$ vs. $\phantom{}>32$ threshold.

3.2 Zero-Shot Generalization on AngioCAD

On the closed-ended AngioCAD tasks (Table 2), CARDEA was evaluated zero-shot against two baselines developed on this held-out cohort [24]. It surpassed the Adaptive Feature Fusion baseline on RCA binary stenosis classification (Lesion $F_1$, 0.85 vs. 0.81) and trailed the VGG19+LSTM model on view classification (RCA $F_1$, 0.90 vs. 0.95). Restricting to the valid-views subset raised both scores (Lesion $F_1$, 0.86; RCA $F_1$, 0.95); as this quality filter is not applied to the baselines, these results are shown for reference only.

媒体内容 · 前往原文查看

Task & Model	Accuracy	Target $F_1$	Macro $F_1$
RCA Stenosis Classification — Full
Adaptive Feature Fusiona	0.72	0.81$†$	—
CARDEA	0.80 (0.77–0.82)	0.85 (0.83–0.87)$†$	0.77 (0.73–0.80)
RCA Stenosis Classification — Valid-views
CARDEA	0.81 (0.78–0.84)	0.86 (0.84–0.88)$†$	0.78 (0.75–0.81)
View Classification — Full
VGG19 + LSTMb	0.97	0.95$‡$	—
CARDEA	0.94 (0.93–0.95)	0.90 (0.88–0.92)$‡$	0.93 (0.92–0.94)
View Classification — Valid-views
CARDEA	0.97 (0.96–0.98)	0.95 (0.94–0.96)$‡$	0.96 (0.96–0.97)

Entries are point estimates with 95% confidence intervals in parentheses, obtained by recomputing each metric across 5,000 case-level bootstrap resamples. Neither baseline reported an interval and their per-case predictions were unavailable, so both are shown as point estimates alone.

Both tasks operate on 5-frame clips and are binary: RCA stenosis classification (Lesion vs. Non-lesion) and view classification (LCA vs. RCA). Full comprises all 412 studies and excludes none, so both models are scored on the same cohort; the valid-views subset ($n=326$) keeps only studies in which CARDEA identified both an LCA and an RCA view, a filter the baselines do not apply, and is therefore shown for reference only (Section 3.2). Each clip is anchored on a keyframe CARDEA selects itself, whereas the baselines follow their original frame-sampling protocol, a difference intrinsic to comparing an end-to-end system against modular baselines.

^a Dynamic weighting of dual-backbone features (ResNet101 and VGG16) [24].

^b Frame-wise feature extraction (VGG19) with temporal sequence modeling (LSTM) [24].

^$†$ Target metric represents the lesion-positive class (Lesion) $F_1$-score.

^$‡$ Target metric represents the minority class (RCA) $F_1$-score.

3.3 Report Generation Across Training Stages

A core question is whether closed-ended training alone can improve an LVLM’s open-ended report generation; in our pipeline this ability rose only after RLVR, not under SFT (Table 3). Since text-similarity metrics need not solely reflect diagnostic agreement, we prioritize VS-$F_1$. An always-normal report in the reference format served as a stress test for text-similarity metrics and as the VS-$F_1$ floor. With 74% of graded AngioCAD sub-segments normal, it outscored our final model on ROUGE-L (0.81 vs. 0.35) and BERTScore (0.77 vs. 0.74), despite a two-class VS-$F_1$ of 0.312. Because VS-$F_1$ scores extracted discrete severity labels and macro-averages across classes, wording overlap and normal-class dominance affect it less. The untuned base model already reached a two-class VS-$F_1$ of 0.513, well above the naive floor, but the supervised stages eroded it (Stage 1, 0.452; Stage 2, 0.373). Only RLVR reversed the decline, raising it to 0.644 with thinking disabled and 0.686 with native thinking, surpassing the base model and more than doubling the always-normal floor; the three-class score mirrored this trajectory. On the valid-views subset the final two-class VS-$F_1$ reached 0.716.

媒体内容 · 前往原文查看

	Diagnostic		Text similarity
Stage	VS-$F_1$ (2-class)	VS-$F_1$ (3-class)	ROUGE-L	BERTScore
Full cohort ($n=412$)
Naive (always-normal)	0.312 (0.298–0.325)	0.208 (0.199–0.217)	0.81	0.77
Base Model (Qwen3-VL) [25]	0.513 (0.481–0.544)	0.361 (0.335–0.389)	0.22	0.66
Stage 1 (Align)	0.452 (0.427–0.478)	0.309 (0.287–0.332)	0.33	0.73
Stage 2 (Cold start)	0.373 (0.351–0.396)	0.253 (0.235–0.274)	0.34	0.69
Stage 3 (RLVR w/o Think)	0.644 (0.619–0.668)	0.480 (0.454–0.505)	0.35	0.73
CARDEA (Final)	0.686 (0.664–0.707)	0.508 (0.485–0.531)	0.35	0.74
Valid-views subset ($n=326$)
CARDEA (Final)	0.716 (0.692–0.739)	0.549 (0.522–0.575)	0.36	0.74

Entries in the two VS-$F_1$ columns are point estimates with 95% confidence intervals in parentheses, obtained by recomputing each metric across 5,000 case-level bootstrap resamples. The text-similarity columns are point estimates only.

All rows except the last use the full cohort of 412 studies, which excludes none; the last row reports CARDEA on the valid-views subset ($n=326$), which keeps only studies in which both an LCA and an RCA view were identified.

Always-normal is a fixed dummy prediction (every vessel labeled normal) scored directly against the ground truth. VS-$F_1$ (Vessel Severity Macro-$F_1$) is the primary diagnostic metric, reported in 2-class (lesion/non-lesion) and 3-class (normal/mild–moderate/severe) forms. ROUGE-L and BERTScore are supplementary reference-based text-similarity metrics whose scores need not solely reflect diagnostic agreement (Section 3.3). Stage 3 (RLVR w/o Think) and CARDEA (Final) use the same Stage 3 model with thinking disabled and enabled, respectively.

Across five independent MedGemma label-extraction runs over the same fixed reports (sampling temperature 0.5), the standard deviation of VS-$F_1$ was below 0.003 for every entry, indicating that label extraction is stable under stochastic decoding.

3.4 Ablation Studies

CoB reasoning is not native to the base Qwen3-VL. Against the selected configuration (cold start with a conditional CoB reward), we compared three variants, each altering one design choice: no cold start, no reward, or an unconditional reward (Table 4).

Removing the cold start slowed CoB adoption (about 130 steps to near-full usage vs. about 30 with it; Figure 3A) but was not decisive. The reward eventually pulled usage to the same level, and accuracy stayed comparable. The cold start’s real effect was on box scale, as its traces inherit the fine-grained vessel and stenosis boxes distilled from Stage 1. The reward asks only for a box and is silent on its size, so without that demonstration the model keeps whatever coarse boxes still earn the reward. In our runs these spanned nearly the whole frame (median 0.76 of the frame area vs. about 0.01 with a cold start; Figure 3B), running counter to the interpretability purpose.

Removing the CoB reward let grounding collapse to about 5% (Figure 3A), yet accuracy stayed comparable. CoB is therefore not what drives accuracy; rewarding it does no harm and is what sustains the grounding.

媒体内容 · 前往原文查看

A

B

Making the reward unconditional left accuracy statistically indistinguishable on our test splits. In a related tool-use setting, DeepEyes [32] reported a clearer gap, with the conditional reward converging to higher accuracy. We nonetheless retain the conditional form: it had the highest validation point estimates and is the stricter rule, never rewarding a box on a wrong answer.

媒体内容 · 前往原文查看

RLVR initialization	Align	Cold start	Cold start	Cold start
CoB reward	Cond.	None	Uncond.	Cond.
Validation
Dominance	0.80	0.82	0.77	0.84
Complexity Assess.	0.71	0.75	0.71	0.76
Test
Dominance (Real)	0.81 (0.75–0.86)	0.81 (0.75–0.86)	0.73 (0.68–0.79)	0.76 (0.70–0.81)
Dominance (Shift)	0.78 (0.71–0.85)	0.86 (0.80–0.91)	0.79 (0.72–0.86)	0.86 (0.80–0.91)
Complexity Assess.	0.73 (0.60–0.84)	0.82 (0.69–0.92)	0.81 (0.69–0.91)	0.85 (0.73–0.94)
Report (2-class)	0.661 (0.633–0.689)	0.613 (0.587–0.639)	0.665 (0.640–0.691)	0.680 (0.654–0.706)

Test entries are point estimates with 95% confidence intervals in parentheses, obtained by recomputing each metric across 5,000 case-level bootstrap resamples. Validation entries are monitor-only and are shown as point estimates alone.

The four variants were each trained for 400 steps under a fixed budget and are distinct from the full CARDEA model. Validation rows are used for model selection and are the mean over the final five validation checkpoints (steps 360–400, evaluated every 10 steps), from a single run and on different populations from the test rows. Test rows are the official CoronaryDominance/CardioSyntax test splits and zero-shot AngioCAD report generation.

Align denotes initialization from the Stage 1 aligned model and Cold start from the Stage 2 cold-start model; Cond., conditional CoB reward; Uncond., unconditional CoB reward; and None, no CoB reward. Real and Shift denote the Real Distribution and Domain Shift test sets, respectively. Dominance classification and complexity assessment use Macro $F_1$; report generation uses the two-class vessel-severity macro-$F_1$.

4 Discussion

CARDEA is a generalist that broadly matches its specialized baselines, but its aim is not to win every subtask: it is to be the single inference core of a two-pass pipeline that delivers study-level diagnoses and grounds them in auditable bounding boxes within the reasoning trace, which no prior CAG system does [7, 8, 9, 10, 11]. The pipeline’s final diagnosis lies in the study-level tasks, where it matched two cardiologists’ accuracy on complexity assessment and drew level with the dedicated classifier under domain shift, its only distinguishable shortfall being Real Distribution dominance. Its view filtering also raised all three zero-shot point estimates, suggesting view completeness improves information density beyond its computational savings.

Jiang et al. [11] supervised report generation directly and it still performed poorly, which they attribute to pairing one comprehensive report with no intermediate reasoning to link findings to statements. Our results suggest a different route: report generation was excluded from training, yet the two supervised stages eroded its zero-shot quality, and only RLVR reversed the decline. We hypothesize that each verified answer forces CARDEA to run its own multi-view synthesis, which the reward repeatedly refines; the RLVR model surpasses the untuned base even with native thinking disabled, indicating this synthesis is internalized rather than emitted in the trace. By contrast, imitation copies only the output form, consistent with the decline under supervision. Although this interpretation remains hypothetical and rests on a single run, it echoes DeepSeek-R1 [15], where policy optimization surfaced capabilities that imitation did not reach.

CARDEA’s auditable grounding rests on three design choices. The cold-start distillation drives box quality: without it, boxes tend to be coarse and less interpretable (Figure 3B). The CoB reward sustains the grounding at no cost to accuracy: on every split, each rewarded configuration remained comparable to or above the variant trained without it (Table 4). Conditioning that reward on a correct answer yielded higher validation point estimates than the unconditional form, but the two forms remained comparable at test.

We also weigh the trade-offs CARDEA would face in deployment. First, its autoregressive architecture yields a median inference time of about 6 seconds per study even under tensor parallelism on $4\times$NVIDIA B200 GPUs (Appendix S3), so it cannot serve applications that need a sub-second response. Second, although CARDEA performed CoB on nearly every in-distribution diagnosis (usage 99.5% on Real Distribution dominance and 100% on complexity assessment), its usage falls on out-of-distribution imaging or tasks (65% on Domain Shift dominance and 75% on report generation). Repeated sampling narrowed both gaps, raising Domain Shift dominance usage from 65% to 92% and report-generation usage from 75% to 98.5% across seven rollouts without materially changing task performance (Appendix S2.1). Deployment should therefore be tested against the expected data distribution first, and closing the gap fully may require further SFT on rejection-sampled trajectories from out-of-distribution data.

These conclusions are subject to several limitations. Above all, CARDEA has not been validated in a clinical setting, and its reports have not undergone blinded comparison with physician-authored reports; prospective validation against expert cardiologists is a prerequisite for clinical use. Training limitations include cross-stage comparisons based on one run per stage; training data focused mainly on vessels and stenoses, leaving other findings insufficiently learned; and study-level training using at most 10 static keyframes per study without the temporal dynamics of contrast flow. Evaluation limitations include CoB assessment based only on box frequency, box area, and final-answer correctness, not box faithfulness; the report metric’s lack of a human-expert reference, requiring report scores to be interpreted relative to the naive floor and across stages rather than as absolute measures of report quality; and two small test sets (complexity assessment, $n=60$; keyframe selection, 48 videos from 5 patients) that may be underpowered. Additional limitations and technical considerations are discussed in Appendix S3.

5 Conclusion

We presented CARDEA, a unified LVLM for end-to-end CAG interpretation developed without human-annotated reasoning traces. Its central contribution is an auditable decision process in which CoB evidence links conclusions to specific regions in the source images, allowing clinicians to inspect the spatial basis of each diagnosis. After three-stage training on closed-ended tasks, CARDEA remained competitive with task-specific baselines and performed comparably to two interventional cardiologists on complexity assessment. It also generalized zero-shot to the fully held-out AngioCAD cohort, surpassing the published baseline for RCA stenosis classification. Although report generation was excluded from training, the RLVR stage reversed the decline under SFT and raised zero-shot performance above the untuned base model. Together, these results establish the feasibility of auditable end-to-end CAG interpretation but do not yet establish readiness for clinical use. Prospective validation against expert cardiologists remains essential before deployment.

Ethics Approval

The research protocol encompassing this work was approved by the Research Ethics Committee of China Medical University and Hospital, Taichung, Taiwan (IRB No. CMUH114-REC2-012), with a waiver of informed consent. All analyses reported in this manuscript used only publicly available, de-identified datasets; no patient data from China Medical University Hospital were analyzed.

Funding

This work was supported by the National Science and Technology Council, Taiwan (Grant No. NSTC 114-2314-B-039-073), and China Medical University Hospital (Grant Nos. DMR-115-087 and DMR-115-108).

Use of Artificial Intelligence

This study used ChatGPT (OpenAI) and Claude (Anthropic) to improve the manuscript’s language, grammar, and style. The authors conceived and drafted the manuscript. These tools also assisted in debugging and optimizing portions of the experimental code. The authors reviewed all AI-generated suggestions and finalized all modifications. They assume full responsibility for the scientific accuracy and integrity of the work. No protected health information was used or disclosed because all datasets were publicly available and de-identified.

Data Sharing Statement

All datasets used in this study are publicly available [18, 19, 23, 20, 21, 24]. The model weights are available at https://huggingface.co/benbayibaurba/cardea-v0, and the inference code at https://github.com/benbayibaurba/cardea.

References

• [1] E. Oikonomou, P. Theofilis, S. Lampsas, O. Katsarou, K. Kalogeras, G. Marinos, A. Tsatsaragkou, A. Anastasiou, A. Lysandrou, M. Gounaridi, I. Gialamas, M. Vavuranakis, D. Tousoulis, M. Vavuranakis, and G. Siasos (2022) Current concepts and future applications of non-invasive functional and anatomical evaluation of coronary artery disease. Life 12 (11), pp. 1803. External Links: Document Cited by: §1.
• [2] S. Çimen, A. Gooya, M. Grass, and A. F. Frangi (2016) Reconstruction of coronary arteries from x-ray angiography: a review. Medical Image Analysis 32, pp. 46–68. External Links: Document Cited by: §1.
• [3] P. Green, P. Frobisher, and S. Ramcharitar (2016) Optimal angiographic views for invasive coronary angiography: a guide for trainees. British Journal of Cardiology 23, pp. 110–113. External Links: Document Cited by: §1.
• [4] S. Shivaie, H. Tohidi, P. Loganathan, M. Kar, H. Hashemy, and M. A. Shafiee (2024) Interobserver variability of coronary stenosis characterized by coronary angiography: a single-center (toronto general hospital) retrospective chart review by staff cardiologists. Vascular Health and Risk Management 20, pp. 359–368. External Links: Document Cited by: §1.
• [5] T. Chen, J. Yap, Y. K. Keong, et al. (2025) Computational methods for analysing x-ray-guided coronary angiography. Journal of Asian Pacific Society of Cardiology (JAPSC) 4, pp. e12. Cited by: §1.
• [6] J. H. Moon, W. C. Cha, M. J. Chung, K. Lee, B. H. Cho, J. H. Choi, et al. (2021) Automatic stenosis recognition from coronary angiography using convolutional neural networks. Computer methods and programs in biomedicine 198, pp. 105819. Cited by: §1.
• [7] R. Avram, J. E. Olgin, Z. Ahmed, L. Verreault-Julien, A. Wan, J. Barrios, S. Abreau, D. Wan, J. E. Gonzalez, J. Tardif, et al. (2023) CathAI: fully automated coronary angiography interpretation and stenosis estimation. npj Digital Medicine 6 (1), pp. 142. Cited by: §1, §4.
• [8] É. Labrecque Langlais, D. Corbin, O. Tastet, A. Hayek, G. Doolub, S. Mrad, J. Tardif, J. Tanguay, G. Marquis-Gravel, G. H. Tison, et al. (2024) Evaluation of stenoses using ai video models applied to coronary angiography. NPJ digital medicine 7 (1), pp. 138. Cited by: §1, §2.3, Table 1, §4.
• [9] S. Harrabi, Y. Wu, M. Vukadinovic, et al. (2026) DeepCORO-clip: a multi-view foundation model for comprehensive coronary angiography video-text analysis and external validation. arXiv preprint arXiv:2603.17675. Cited by: §S3.1, §1, §4.
• [10] Y. Nakamura, S. Kodera, H. Settai, H. Shinohara, M. Tamura, T. Noguchi, T. Furusawa, R. Takizawa, T. Kabayama, and N. Takeda (2025) CAG-vlm: fine-tuning of a large-scale model to recognize angiographic images for next-generation diagnostic systems. arXiv preprint arXiv:2505.04964. Cited by: §1, §4.
• [11] Q. Jiang, Y. Ke, L. G. Sinisterra, K. Elangovan, Z. Li, K. K. Yeo, Y. Jonathan, and D. S. W. Ting (2026) Vision language model for coronary angiogram analysis and report generation: development and evaluation study. medRxiv, pp. 2026–04. Cited by: 2nd item, §S1.2.2, §1, §2.3, §3.1, Table 1, §4, §4.
• [12] C. Li, C. Wong, S. Zhang, N. Usuyama, H. Liu, J. Yang, T. Naumann, H. Poon, and J. Gao (2023) Llava-med: training a large language-and-vision assistant for biomedicine in one day. Advances in Neural Information Processing Systems 36, pp. 28541–28564. Cited by: §1.
• [13] S. Bannur, K. Bouzid, D. C. Castro, A. Schwaighofer, A. Thieme, S. Bond-Taylor, M. Ilse, F. Pérez-García, V. Salvatelli, H. Sharma, et al. (2024) Maira-2: grounded radiology report generation. arXiv preprint arXiv:2406.04449. Cited by: §1.
• [14] N. Yildirim, H. Richardson, M. T. Wetscherek, J. Bajwa, J. Jacob, M. A. Pinnock, S. Harris, D. Coelho De Castro, S. Bannur, S. Hyland, et al. (2024) Multimodal healthcare ai: identifying and designing clinically relevant vision-language applications for radiology. In Proceedings of the 2024 CHI Conference on Human Factors in Computing Systems, pp. 1–22. Cited by: §1.
• [15] D. Guo, D. Yang, H. Zhang, J. Song, R. Zhang, R. Xu, Q. Zhu, S. Ma, P. Wang, X. Bi, et al. (2025) Deepseek-r1: incentivizing reasoning capability in llms via reinforcement learning. arXiv preprint arXiv:2501.12948. Cited by: §S3.3, §1, §2.2, §4.
• [16] Y. Fan, X. He, D. Yang, K. Zheng, C. Kuo, Y. Zheng, S. J. Narayanaraju, X. Guan, and X. E. Wang (2025) GRIT: teaching mllms to think with images. arXiv preprint arXiv:2505.15879. Cited by: §1.
• [17] H. Xu, Y. Nie, H. Wang, Y. Chen, W. Li, J. Ning, L. Liu, H. Wang, L. Zhu, J. Liu, et al. (2025) Medground-r1: advancing medical image grounding via spatial-semantic rewarded group relative policy optimization. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pp. 391–401. Cited by: §1.
• [18] A. Jiménez-Partinen, M. A. Molina-Cabello, K. Thurnhofer-Hemsi, E. J. Palomo, J. Rodríguez-Capitán, A. I. Molina-Ramos, and M. Jiménez-Navarro (2024) CADICA: a new dataset for coronary artery disease detection by using invasive coronary angiography. Expert Systems 41 (12), pp. e13708. Cited by: 1st item, 2nd item, §1, §2.1, Data Sharing Statement.
• [19] M. Popov, A. Amanturdieva, N. Zhaksylyk, A. Alkanov, A. Saniyazbekov, T. Aimyshev, E. Ismailov, A. Bulegenov, A. Kuzhukeyev, A. Kulanbayeva, et al. (2024) Dataset for automatic region-based coronary artery disease diagnostics using x-ray angiography images. Scientific data 11 (1), pp. 20. Cited by: 1st item, §1, §2.1, Data Sharing Statement.
• [20] I. Kruzhilov, G. Mazanov, A. Ponomarchuk, G. Zubkova, A. Shadrin, R. Utegenov, P. Blinov, and I. Bessonov (2025) CoronaryDominance: angiogram dataset for coronary dominance classification. Scientific Data 12 (1), pp. 341. Cited by: 1st item, §S1.1.2, §1, §2.1, §2.3, Table 1, Data Sharing Statement.
• [21] A. Ponomarchuk, I. Kruzhilov, G. Mazanov, R. Utegenov, A. Shadrin, G. Zubkova, I. Bessonov, and P. Blinov (2025) CardioSyntax: end-to-end syntax score prediction-dataset, benchmark and method. In 2025 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pp. 5873–5883. Cited by: 2nd item, §S1.1.2, §1, §2.1, §2.3, Table 1, Data Sharing Statement.
• [22] G. Sianos, M. Morel, A. P. Kappetein, M. Morice, A. Colombo, K. Dawkins, M. Van Den Brand, N. Van Dyck, M. E. Russell, F. W. Mohr, et al. (2005) The syntax score: an angiographic tool grading the complexity of coronary artery disease. EuroIntervention 1 (2), pp. 219–227. Cited by: §2.1.
• [23] J. Chen, C. Gui, R. Ouyang, A. Gao, S. Chen, G. H. Chen, X. Wang, R. Zhang, Z. Cai, K. Ji, et al. (2024) Huatuogpt-vision, towards injecting medical visual knowledge into multimodal llms at scale. arXiv preprint arXiv:2406.19280. Cited by: 1st item, §2.1, Data Sharing Statement.
• [24] M. S. Hosseini, A. R. Naghsh-Nilchi, M. Safayani, M. Sadeghi, E. Shirvani, M. Danesh, and S. A. Miramirkhani (2026) AngioCAD: a public x-ray angiography dataset and an adaptive fusion framework for stenosis detection. Computer Methods and Programs in Biomedicine, pp. 109368. Cited by: 3rd item, §S1.1.2, §2.1, §2.3, §3.2, Table 2, Table 2, Data Sharing Statement.
• [25] S. Bai, Y. Cai, R. Chen, K. Chen, X. Chen, Z. Cheng, L. Deng, W. Ding, C. Gao, C. Ge, W. Ge, Z. Guo, Q. Huang, J. Huang, F. Huang, B. Hui, S. Jiang, Z. Li, M. Li, M. Li, K. Li, Z. Lin, J. Lin, X. Liu, J. Liu, C. Liu, Y. Liu, D. Liu, S. Liu, D. Lu, R. Luo, C. Lv, R. Men, L. Meng, X. Ren, X. Ren, S. Song, Y. Sun, J. Tang, J. Tu, J. Wan, P. Wang, P. Wang, Q. Wang, Y. Wang, T. Xie, Y. Xu, H. Xu, J. Xu, Z. Yang, M. Yang, J. Yang, A. Yang, B. Yu, F. Zhang, H. Zhang, X. Zhang, B. Zheng, H. Zhong, J. Zhou, F. Zhou, J. Zhou, Y. Zhu, and K. Zhu (2025) Qwen3-vl technical report. External Links: 2511.21631, Link Cited by: §2.2, Table 3.
• [26] K. Chen, P. Shi, H. Qiu, Z. Zeng, S. Yang, W. Mao, and L. Ma (2025) Metis-specs: decoupling multimodal learning via self-distilled preference-based cold start. arXiv preprint arXiv:2510.25801. Cited by: §2.2.
• [27] A. Sellergren, S. Kazemzadeh, T. Jaroensri, A. Kiraly, M. Traverse, T. Kohlberger, S. Xu, F. Jamil, C. Hughes, C. Lau, et al. (2025) Medgemma technical report. arXiv preprint arXiv:2507.05201. Cited by: §S1.2.3, §2.3.
• [28] C. Lin (2004) Rouge: a package for automatic evaluation of summaries. In Text summarization branches out, pp. 74–81. Cited by: §S1.2.3, §2.3.
• [29] T. Zhang, V. Kishore, F. Wu, K. Q. Weinberger, and Y. Artzi (2019) Bertscore: evaluating text generation with bert. arXiv preprint arXiv:1904.09675. Cited by: §S1.2.3, §2.3.
• [30] M. Pascual-González, A. Jiménez-Partinen, E. J. Palomo, E. López-Rubio, and A. Ortega-Gómez (2025) Hyperparameter optimization of yolo models for invasive coronary angiography lesion detection and assessment. Computers in Biology and Medicine 196, pp. 110697. Cited by: §2.3, Table 1.
• [31] D. Tran, A. Huynh, A. Huynh, and T. Nguyen-Thoi (2025) Anatomy-specific two-stage yolov8 approach for improved coronary segmentation using the arcade dataset. Optics Continuum 4 (2), pp. 303–317. Cited by: §2.3, Table 1.
• [32] Z. Zheng, M. Yang, J. Hong, C. Zhao, G. Xu, L. Yang, C. Shen, and X. Yu (2026) DeepEyes: incentivizing “thinking with images” via reinforcement learning. In International Conference on Learning Representations, Vol. 2026, pp. 126775–126798. External Links: Link Cited by: §3.4.
• [33] H. Moalla, A. Ghrab, A. Bahloul, B. B. Hamed, and L. Abid (2023) Deep angiokey: a novel approach for objective keyframe extraction in coronary angiography analysis. Research Square preprint. External Links: Document Cited by: §S1.2.1.
• [34] Y. Zheng, R. Zhang, J. Zhang, Y. Ye, Z. Luo, Z. Feng, and Y. Ma (2024) LlamaFactory: unified efficient fine-tuning of 100+ language models. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 3: System Demonstrations), Bangkok, Thailand. External Links: Link Cited by: §S1.3.
• [35] Y. Zheng, J. Lu, S. Wang, Z. Feng, D. Kuang, Y. Xiong, and R. Zhang (2025) EasyR1: an efficient, scalable, multi-modality rl training framework. Note: https://github.com/hiyouga/EasyR1 Cited by: §S1.3.
• [36] G. Sheng, C. Zhang, Z. Ye, X. Wu, W. Zhang, R. Zhang, Y. Peng, H. Lin, and C. Wu (2024) HybridFlow: a flexible and efficient rlhf framework. arXiv preprint arXiv:2409.19256. Cited by: §S1.3.
• [37] E. J. Hu, Y. Shen, P. Wallis, Z. Allen-Zhu, Y. Li, S. Wang, L. Wang, and W. Chen (2021) Lora: low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685. Cited by: Table S1.
• [38] I. Loshchilov and F. Hutter (2017) Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101. Cited by: Table S1.
• [39] C. Gao, C. Zheng, X. Chen, K. Dang, S. Liu, B. Yu, A. Yang, S. Bai, J. Zhou, and J. Lin (2025) Soft adaptive policy optimization. arXiv preprint arXiv:2511.20347. Cited by: Table S1.
• [40] S. Rajbhandari, J. Rasley, O. Ruwase, and Y. He (2020) Zero: memory optimizations toward training trillion parameter models. In SC20: international conference for high performance computing, networking, storage and analysis, pp. 1–16. Cited by: Table S1.
• [41] Y. Zhao, A. Gu, R. Varma, L. Luo, C. Huang, M. Xu, L. Wright, H. Shojanazeri, M. Ott, S. Shleifer, et al. (2023) Pytorch fsdp: experiences on scaling fully sharded data parallel. arXiv preprint arXiv:2304.11277. Cited by: Table S1.
• [42] W. Kwon, Z. Li, S. Zhuang, Y. Sheng, L. Zheng, C. H. Yu, J. E. Gonzalez, H. Zhang, and I. Stoica (2023) Efficient memory management for large language model serving with pagedattention. In Proceedings of the ACM SIGOPS 29th Symposium on Operating Systems Principles, Cited by: §S3.4.

Supplementary Appendix

Appendix S1 Supplementary Methods

S1.1 Datasets and Preprocessing

CARDEA is designed to perform both single-view and study-level reasoning within our end-to-end pipeline. We categorize the datasets accordingly and describe the preprocessing and derived tasks for each.

S1.1.1 Single-View Datasets

• ARCADE [19]: A dataset providing 3,000 single-view keyframes compiled from 1,500 distinct patients using diverse imaging equipment. It includes expert annotations for 25 distinct coronary segments (defined by the SYNTAX score) and localizations for stenotic lesions ($\phantom{}\ge 50\%$ diameter stenosis). We utilize these annotations to construct three tasks: (1) Vessel Detection, outputting spatial bounding boxes for specific vessel segments; (2) Stenosis Detection, outputting bounding boxes for stenotic lesions; and (3) View Classification for left coronary artery (LCA), right coronary artery (RCA), or OTHER. The LCA and RCA labels are derived from the segment annotations. To formulate the “OTHER” class, we manually curated a small set of non-coronary-angiographic medical images from PubMedVision [23] alongside non-diagnostic, low-quality frames from CADICA [18], enabling the model to train for view recognition and image usability simultaneously; this patient diversity supports generalization. The segment and stenosis annotations form two separate 1,500-image subsets, each with its own official split of 1,000 training, 200 validation, and 300 test samples.

• CADICA [18]: A dataset comprising 668 multi-view angiographic videos from 42 patients, among which 382 videos were expertly curated as diagnostic. Inspired by Jiang et al. [11], who used CADICA to train a standalone ViT-based frame selector, we use these expert-labeled frames to train our unified LVLM for the Keyframe Selection task. As a preprocessing step, each frame is resized from $512\times 512$ to $256\times 256$; videos exceeding 50 frames are subsampled to 50 evenly spaced frames to bound computational cost. Beyond keyframe selection, CADICA also carries multi-level stenosis annotations (bounding boxes labeled across seven severity levels, from sub-50% to total occlusion); we include these as an auxiliary localization signal during feature alignment, exposing the model to lesions milder than ARCADE’s 50% stenosis threshold. We do not benchmark this task in isolation. With only 42 patients, highly correlated intra-video frames, and labels split across seven severity levels, the per-level data is too limited to support a separate localization benchmark. Since no official split is provided, we partition the 42 patients by stratified sampling on maximum stenosis severity into 32 training, 5 validation, and 5 test patients.

S1.1.2 Preliminary LVLM for Study-Level Dataset Preparation

To construct the study-level datasets, we use the single-view tasks from ARCADE and CADICA to train a preliminary LVLM that serves purely as an offline data-preparation utility. For each raw video, it selects the optimal diagnostic keyframe, classifies it as LCA, RCA, or OTHER, discards non-diagnostic images, and ensures each study retains both valid LCA and RCA views. To bound the visual token budget, we cap each study at 10 keyframes. For the training and validation splits we additionally keep only those with at least four to maintain quality, whereas the official test sets are retained in full to preserve comparability with the published baselines. This preliminary LVLM is used only to build the training datasets (CoronaryDominance [20] and CardioSyntax [21]), as the final CARDEA model does not yet exist at that stage; on the held-out AngioCAD [24], all steps—keyframe selection, view filtering, and diagnostic reasoning—are instead performed end-to-end by the fully trained CARDEA, as in its intended deployment.

S1.1.3 Study-Level Datasets

• CoronaryDominance [20]: A dataset comprising 1,574 multi-view angiographic studies. Each study is strictly labeled as either Left or Right dominance according to the SYNTAX score definitions (co-dominance is excluded). This dataset is used for the study-level task of Dominance Classification. After the offline preprocessing (Section S1.1.2), the Main subset yields 924 training and 100 validation studies. For evaluation, we utilize the two official independent test sets: the Real Distribution subset (400 studies reflecting clinical class imbalance) and the Domain Shift subset (149 studies acquired with distinct imaging equipment, an out-of-distribution cohort).

• CardioSyntax [21]: A dataset encompassing 1,844 angiographic studies with expert-evaluated continuous SYNTAX scores, officially split into 1,784 training and 60 test studies. Given the difficulty of predicting continuous SYNTAX scores from images alone, we formulate a binary Complexity Assessment task by discretizing the scores into normal-to-intermediate complexity ($0$–$32$) and high complexity ($\phantom{}>32$). After applying the same offline preprocessing as in Section S1.1.2, 1,433 studies remain for training and 38 for validation. Evaluation is conducted on the official independent test set of 60 studies, for which the dataset additionally provides labels from two other cardiologists that serve as expert baselines.

• AngioCAD [24]: A dataset of 413 angiographic studies annotated with stenosis grades for each coronary segment (defined by the American Heart Association coronary segment model). Excluded from all training stages, it serves exclusively as a zero-shot evaluation benchmark. We derive three tasks: (1) Multi-Frame View Classification (LCA vs. RCA): following the original AngioCAD setup, CARDEA’s single-view inference extracts the keyframe and samples 5 consecutive frames to predict the view; (2) Multi-Frame RCA Binary Stenosis Classification (Lesion vs. Non-lesion): using the same 5-frame sampling restricted to RCA views, with each clip labeled Lesion ($\phantom{}>0\%$ stenosis) or Non-lesion; and (3) Report Generation: the model generates a JSON clinical report covering the four major branches (LM, LAD, LCX, RCA), each field a free-text description with explicit segment references and stenosis-percentage estimates. Of the 413 studies, one is missing from the released data,¹¹ 1 The missing study is case 157. leaving 412 available for evaluation. To examine the effect of the end-to-end view filtering, we evaluate all three tasks on two cohorts: all 412 studies, and a valid-views subset of 326 studies retaining both an LCA and an RCA view.

S1.2 Evaluation Metrics

Standard classification metrics including Macro $F_1$, Micro $F_1$, Accuracy, and Balanced Accuracy follow conventional definitions and are not elaborated here.

S1.2.1 Keyframe Selection

We evaluate keyframe selection using $D_k$, defined as the minimum absolute index distance between the model’s predicted optimal frame $id{x}_{best}^{pred}$ and the set of expert-annotated diagnostically usable frames $S_{gt}$:

$D_k=\underset{j\in S_{gt}}{\min }\left|id{x}_{best}^{pred}-j\right|$

(S1)

A lower $D_k$ indicates better performance, with a value of zero indicating that the predicted frame falls within the expert-annotated usable range. This metric adapts the frame distance metric proposed in Deep AngioKey [33], which computes the distance from a single designated ground-truth frame. Our adaptation instead measures the distance from the full set of annotated usable frames because CADICA provides a usable frame set rather than a single optimal index.

S1.2.2 Detection

We evaluate detection at the instance level using one-to-one matching between predicted bounding boxes $B_{pred}$ and ground-truth boxes $B_{gt}$. A predicted box $b$ and a ground-truth box $g$ are eligible for matching if they have the same semantic label and an Intersection over Union (IoU) of at least 0.5. The set of eligible pairs is

$E_{0.5}^{\text{IoU}}=\{(b,g)\in B_{pred}\times B_{gt}|b_{label}=g_{label},\mathrm{IoU}(b_{box},g_{box})\ge 0.5\}.$

(S2)

The eligible pairs in $E_{0.5}^{\text{IoU}}$ are matched one-to-one, with each predicted and ground-truth box appearing in at most one match. The resulting pairs form $M_{\text{IoU}}$.

The resulting true-positive (TP), false-positive (FP), and false-negative (FN) counts are

$TP=\left|M_{\text{IoU}}\right|,FP=\left|B_{pred}\right|-TP,FN=\left|B_{gt}\right|-TP.$

(S3)

Precision, recall, and $F_1$ are then computed as

$P=\frac{TP}{TP+FP},R=\frac{TP}{TP+FN},F_1=2\frac{PR}{P+R}.$

(S4)

We denote this instance-level score as $F_1$@IoU$\ge$0.5.

However, an IoU threshold of 0.5 can understate stenosis localization performance on CAG [11] because the criterion depends strongly on box extent. Since stenosis extent can be difficult to annotate consistently owing to the interpretive difficulty and inter-observer variability noted in the Introduction, the same diffuse lesion may be represented as one large box or several smaller boxes. A prediction that correctly localizes a lesion but differs in extent can therefore fall below the threshold and be counted as both a false positive and a false negative.

We therefore use Jiang et al.’s predicted-area overlap criterion for stenosis detection to facilitate comparison. For a predicted box $b$ and a ground-truth box $g$, intersection over prediction (IoP) is defined as

$\mathrm{IoP}(b,g)=\frac{\mathrm{area}\left(b_{box}\cap g_{box}\right)}{\mathrm{area}\left(b_{box}\right)}.$

(S5)

After the IoU-based matching is completed, $U_{pred}$ and $U_{gt}$ denote the sets of predicted and ground-truth boxes that remain unmatched, respectively. The pairs eligible under the IoP criterion are

$E_{0.6}^{\text{IoP}}=\{(b,g)\in U_{pred}\times U_{gt}|b_{label}=g_{label},\mathrm{IoP}(b,g)\ge 0.6\}.$

(S6)

The eligible pairs in $E_{0.6}^{\text{IoP}}$ are matched one-to-one. The resulting pairs form $M_{\text{IoP}}$.

The final counts under the relaxed criterion are

$T{P}_{rel}=\left|M_{\text{IoU}}\right|+\left|M_{\text{IoP}}\right|,F{P}_{rel}=\left|B_{pred}\right|-T{P}_{rel},F{N}_{rel}=\left|B_{gt}\right|-T{P}_{rel}.$

(S7)

This relaxation is applied only to stenosis detection; substituting the relaxed counts into the definitions above yields $F_1$@(IoU$\ge$0.5 or IoP$\ge$0.6).

S1.2.3 Report Generation

We evaluate report generation using vessel-severity macro-$F_1$ (VS-$F_1$), a task-specific application of macro-$F_1$ to severity labels for the major coronary branches. Figure S1 provides a worked example of the evaluation representations described in this section.

AngioCAD provides segment-level stenosis grades for 15 coronary segments. For evaluation, these segments are grouped into four major branches: LM; LAD (proximal, mid, and distal LAD and the first and second diagonal branches); LCX (proximal, mid, and distal LCX and the obtuse marginal); and RCA (proximal, mid, and distal RCA, the posterior descending artery, and the posterolateral branch).

We consolidate AngioCAD’s seven original segment-level categories at two granularities. For the primary two-class evaluation, the categories are consolidated into non-lesion (NL) and lesion (1–100%). For the finer-grained three-class evaluation, they are consolidated into the ordered classes normal (NL), mild–moderate (1–50%), and severe (51–100%). At both granularities, the ground-truth label for each major branch is determined by the highest severity among its sub-segments. Panels A and B of Figure S1 show the resulting two- and three-class labels for one held-out test case, respectively.

The report-generation prompt requires a structured JSON report with one free-text field for each major branch. Within each field, findings are reported at the segment level using numerical stenosis percentages or ranges rather than qualitative severity terms alone. Vessels with no diseased segments are reported as normal, whereas anatomically absent or unvisualized vessels are reported as nan. Panel C of Figure S1 shows a CARDEA-generated report produced under these instructions.

For each test case, only the generated free-text report is presented to MedGemma-27B-IT [27], which extracts a branch-level severity label for each major branch. A rule-based routine then compares the extracted labels with the corresponding ground-truth labels, pools the counts across all major branches and test cases, and computes the macro-averaged $F_1$:

$\text{VS-}{F}_1=\frac{1}{\left|C\right|}\sum _{c\in C}2\frac{P_c{R}_c}{P_c+R_c},$

(S8)

where $C$ is the set of severity classes at the selected granularity, and $P_c$ and $R_c$ are the conventional precision and recall for class $c$. Because VS-$F_1$ is macro-averaged, each severity class contributes equally, preventing the dominant normal class from overwhelming the aggregate score.

To assess the stability of MedGemma label extraction under stochastic decoding, we repeat the extraction five times at a sampling temperature of 0.5 using the same fixed set of generated reports. The standard deviation across runs is below 0.003 for every VS-$F_1$ entry reported in the paper. A complementary consistency check on the full cohort ($n=412$) shows that MedGemma correctly extracts the label normal for every major branch from the always-normal reports, yielding VS-$F_1$ scores identical to those obtained by directly comparing fixed normal labels with the ground truth (Table 3). Panel D of Figure S1 shows the fixed always-normal report used for this check.

Alongside VS-$F_1$, we report ROUGE-L [28] and BERTScore [29] as supplementary reference-based text-similarity metrics. Because AngioCAD provides no reference reports, we construct a reference for each test case by grouping the segment-level labels under their corresponding major-branch fields without aggregating them into branch-level labels (Figure S1E). The fixed always-normal report stress-tests these metrics’ sensitivity to wording overlap and normal-segment prevalence (Section 3.3).

媒体内容 · 前往原文查看

A. Two-class VS-$F_1$ ground truth LM lesion LAD non-lesion LCX lesion RCA lesion

B. Three-class VS-$F_1$ ground truth LM mild–moderate LAD normal LCX mild–moderate RCA severe

C. CARDEA-generated report { "lm": "30-50% stenosis", "lad": "normal", "lcx": "normal", "rca": "Severe 70-90% stenosis in proximal and mid segments, mild 0-50% in distal" }

D. Fixed always-normal report { "lm": "normal/nan", "lad": "proximal normal/nan; mid normal/nan; distal normal/nan; 1st diagonal normal/nan; 2nd diagonal normal/nan", "lcx": "proximal normal/nan; mid normal/nan; distal normal/nan; om normal/nan", "rca": "proximal normal/nan; mid normal/nan; distal normal/nan; pda normal/nan; plb normal/nan" }

E. ROUGE-L/BERTScore reference { "lm": "0-50% stenosis", "lad": "proximal normal/nan; mid normal/nan; distal normal/nan; 1st diagonal normal/nan; 2nd diagonal normal/nan", "lcx": "proximal normal/nan; mid 0-50% stenosis; distal normal/nan; om normal/nan", "rca": "proximal 0-50% stenosis; mid 50-100% stenosis; distal 0-50% stenosis; pda normal/nan; plb 0-50% stenosis" }

S1.3 Training Configuration

We use Qwen3-VL-30B-A3B-Thinking as the backbone and conduct training on a single node equipped with eight NVIDIA B200 GPUs. The Align and Cold Start stages are implemented with LlamaFactory [34], whereas RLVR is implemented with EasyR1 [35, 36]. Table S1 reports the principal hyperparameters and distributed configuration for each stage.

媒体内容 · 前往原文查看

Setting	Value
SFT: Align and Cold Start
Adaptation	LoRA (rank 32, $\alpha =64$, dropout 0)
Optimizer	AdamW
Learning rate	$1\times {10}^{-4}$ for Align; $5\times {10}^{-5}$ for Cold Start
Global batch size	64
Training duration	2 epochs per stage
Distributed configuration	DeepSpeed ZeRO Stage 2
RLVR
Algorithm	SAPO
Optimizer	AdamW
Learning rate	$1\times {10}^{-6}$
Rollout batch size	32
Responses per prompt	13
Rollout temperature	0.8
Distributed configuration	FSDP for policy updates

SFT, supervised fine-tuning; LoRA, low-rank adaptation [37]; AdamW, Adam with decoupled weight decay [38]; RLVR, reinforcement learning with verifiable rewards; SAPO, Soft Adaptive Policy Optimization [39]; ZeRO, Zero Redundancy Optimizer [40]; FSDP, Fully Sharded Data Parallel [41].

S1.4 Reward Functions

S1.4.1 Total Reward

During RLVR, we assign each rollout a format reward, a task-specific accuracy reward, and a CoB reward. We denote these terms by $R_{\mathrm{fmt}}$, $R_{\mathrm{acc}}$, and $R_{\mathrm{CoB}}$. Their weights are $w_{\mathrm{fmt}}$, $w_{\mathrm{acc}}$, and $w_{\mathrm{CoB}}$. The total reward is

$R_{\mathrm{total}}=w_{\mathrm{fmt}}{R}_{\mathrm{fmt}}+\mathrm{𝟙}\left(R_{\mathrm{fmt}}>0\right)\left[w_{\mathrm{acc}}{R}_{\mathrm{acc}}+w_{\mathrm{CoB}}{R}_{\mathrm{CoB}}\right].$

(S9)

In our configuration, the format weight is 0.2 and the accuracy weight is 0.8. The CoB weight is 1.2 for study-level tasks and 0 for single-view tasks. The format reward also gates the other components. If it is zero, neither accuracy nor CoB contributes to the total reward.

S1.4.2 Format Reward

The format reward is granted when the response follows the required output format:

$R_{\mathrm{fmt}}=\mathrm{𝟙}\left(\text{reasoning and answer are both present and separately parseable}\right).$

(S10)

S1.4.3 Task-Specific Accuracy Reward

The accuracy reward is defined separately for classification, detection, and keyframe selection.

For classification, let $\hat{y}$ and $y_{gt}$ denote the predicted and ground-truth classes, respectively. The reward is binary accuracy per sample:

$R_{\mathrm{acc}}^{cls}=\mathrm{𝟙}\left(\hat{y}=y_{gt}\right).$

(S11)

For detection, the reward reuses the one-to-one matching procedure of Section S1.2.2 but replaces the binary match outcome with a graded score. A roughly correct box therefore receives partial credit. For each predicted box $b$, let $g$ denote its matched ground-truth box when a match exists. The localization score $s\left(b\right)$ is

$s\left(b\right)=\{\begin{array}{cc}0,& b\text{ unmatched, or }\mathrm{IoU}(b_{box},g_{box})\le 0.2,\\ \frac{\mathrm{IoU}(b_{box},g_{box})-0.2}{0.7-0.2},& 0.2<\mathrm{IoU}(b_{box},g_{box})<0.7,\\ 1,& \mathrm{IoU}(b_{box},g_{box})\ge 0.7\end{array}.$

(S12)

Summing these scores gives fractional counts:

$TP=\sum _{b\in B_{pred}}s\left(b\right),FP=\left|B_{pred}\right|-TP,FN=\left|B_{gt}\right|-TP.$

(S13)

These counts define precision and recall for every rollout:

$P=\{\begin{array}{cc}\frac{TP}{TP+FP},& TP+FP>0,\\ 0,& TP+FP=0,\end{array}R=\{\begin{array}{cc}\frac{TP}{TP+FN},& TP+FN>0,\\ 0,& TP+FN=0.\end{array}$

(S14)

The detection accuracy reward is

$R_{\mathrm{acc}}^{det}=\{\begin{array}{cc}1,& B_{pred}=B_{gt}=\varnothing ,\\ 2\cdot \frac{P\cdot R}{P+R},& P+R>0,\\ 0,& \text{otherwise}\end{array}.$

(S15)

The detection scores in Table 1 use the hard $F_1$@IoU$\ge$0.5 of Section S1.2.2 and pool boxes over the complete test set. The graded reward above is used only during training and is computed separately for each rollout.

For keyframe selection, the prompt asks the model to return a JSON object containing a list of diagnostically usable frame indices and one optimal-frame index. We denote the predicted and ground-truth usable-frame sets by $S_{pred}$ and $S_{gt}$, and the predicted optimal-frame index by $id{x}_{best}^{pred}$. The precision and recall of the predicted usable-frame set are

$P=\frac{\left|S_{pred}\cap S_{gt}\right|}{\left|S_{pred}\right|},R=\frac{\left|S_{pred}\cap S_{gt}\right|}{\left|S_{gt}\right|}.$

(S16)

The keyframe-selection accuracy reward adds an optimal-frame bonus to their harmonic mean:

$R_{\mathrm{acc}}^{key}=2\frac{PR}{P+R}+\mathrm{𝟙}\left(id{x}_{best}^{pred}\in S_{pred}\cap S_{gt}\right).$

(S17)

The bonus is awarded when the predicted optimal frame belongs to both usable-frame sets.

S1.4.4 Chain-of-Box Reward

The CoB reward encourages the model to include spatial evidence in its reasoning trace. For a reasoning trace $\tau$, let $g\left(\tau \right)$ indicate whether at least one valid bounding box can be parsed from it:

$g\left(\tau \right)=\mathrm{𝟙}\left(\text{at least one valid bounding box can be parsed from }\tau \right).$

(S18)

The CoB reward is gated on answer correctness through the classification reward defined in Section S1.4.3:

$R_{\mathrm{CoB}}=g\left(\tau \right)\cdot R_{\mathrm{acc}}^{cls}.$

(S19)

The unconditional ablation of Section 3.4 removes the gate and rewards the box by itself:

$R_{\mathrm{CoB}}^{\mathrm{uncond}}=g\left(\tau \right).$

(S20)

Appendix S2 Supplementary Results

S2.1 Repeated-Sampling CoB Coverage

The incomplete CoB coverage noted in the Discussion is a property of a single pass and is partly recoverable by sampling. Exploiting the model’s decoding stochasticity (thinking enabled, sampling temperature 0.5), we drew seven independent rollouts per study for the three trained-task test sets and for zero-shot report generation on the full AngioCAD cohort ($n=412$). Among the first $k$, we select the first rollout whose reasoning trace contains a valid CoB box, and keep the first rollout if none of the first $k$ does. CoB usage@$k$ is the fraction of selected rollouts containing a box. Task performance@$k$ is exact-match accuracy for the classification tasks and two-class VS-$F_1$ for report generation; at $k=1$, both measures reproduce the corresponding single-pass values reported in the main text. As Figure S2 shows, Real Distribution usage rose from 99.5% to 100.0% with accuracy unchanged at 94.3%, whereas Complexity remained at 100.0% usage and 90.0% accuracy. By contrast, Domain Shift usage rose from 65.1% at $k=1$ to 92.0% at $k=7$, while accuracy changed only from 90.6% to 91.3%. For report generation, CoB usage increased from 75.0% at $k=1$ to 98.5% at $k=7$, while two-class VS-$F_1$ changed only from 68.6% to 69.1%. The small fluctuations should not be interpreted as evidence of improved or degraded report quality because they may reflect both rollout-sampling and judge variability. Thus, repeated sampling improved CoB coverage on the lower-coverage evaluations without materially changing task performance.

媒体内容 · 前往原文查看

A

B

Appendix S3 Extended Limitations

The Discussion summarizes the key limitations of the study. A complete account follows.

S3.1 Training and Data Limitations

Public training datasets such as ARCADE and CADICA mainly annotate the location and severity of coronary stenoses. However, comprehensive diagnostic interpretation also requires characterization of additional lesion features (e.g., calcification, thrombus, dissection, and bifurcation involvement). These features may affect lesion assessment and subsequent interventional planning. Recognizing procedural equipment (e.g., catheters and guidewires) and implanted devices (e.g., coronary stents) may also be important for understanding the context in which images were acquired. For example, the DeepCORO-CLIP pipeline [9] identifies non-coronary structures and stent presence. It also uses the first appearance of interventional equipment to retain only diagnostic videos. Such information could help an operational system select diagnostically relevant content from raw procedural recordings during deployment. CARDEA and its current evaluation may therefore not fully capture the range of angiographic findings and procedural contexts encountered in routine coronary angiography.

The CoronaryDominance and CardioSyntax training sets were constructed with a preliminary LVLM that selected keyframes and filtered views (Appendix S1.1). Each study was limited to 10 static keyframes, so the training data support cross-view reasoning but not the temporal dynamics of contrast flow. Any systematic preference for particular projection angles or contrast densities may therefore be reflected in those sets, but we did not measure its downstream effect.

The AngioCAD RCA stenosis task labels any visible narrowing ($\phantom{}>0\%$), whereas ARCADE uses a $\phantom{}\ge 50\%$ threshold. CADICA supplied sub-50% lesion labels during feature alignment, but we did not test whether this prior exposure affected performance on AngioCAD.

CARDEA’s training used one run per stage, so the bootstrap CIs capture test-set sampling variability but not variation across training runs.

S3.2 Evaluation Limitations

The evaluation of trained closed-ended tasks is limited by small test sets. Keyframe selection includes 48 videos from 5 patients, while complexity assessment includes 60 studies. Both may be underpowered and should be considered preliminary. Direct comparability is also limited by two output conversions. We converted DeepCoro’s Algorithm 4 segmentation masks to bounding boxes, possibly understating its detection score. For complexity assessment, we binarized the cardiologists’ continuous SYNTAX scores rather than using predictions produced natively for the binary task. Although the CIs overlap, CARDEA’s Macro $F_1$ point estimate is lower than either cardiologist’s, possibly reflecting reduced sensitivity to the minority high-risk class, where missed high-risk studies are the main clinical concern. Statistically non-distinguishable results do not establish equivalence.

The valid-views subset retains only AngioCAD studies in which CARDEA identified both an LCA and an RCA view. Although this provides potentially more complete view coverage, it may introduce selection bias by excluding more difficult studies. Results on the complete cohort are therefore primary.

The report-generation evaluation uses MedGemma-27B-IT to extract per-vessel severity labels from each generated report. The ground-truth labels and final $F_1$ are derived deterministically, so only label extraction depends on MedGemma (Appendix S1.2.3). The parser exactly reproduces the naive floor on always-normal inputs, but its agreement with human annotators on ambiguous free-text reports was not measured.

VS-$F_1$ also coarsely measures a single dimension of report quality. It scores branch-level diagnostic agreement after the 15 annotated segments are collapsed into four branches by maximum severity, so a report that assigns a lesion to the wrong segment within a branch, or omits a second lesion in the same branch, is still scored as correct. Expression, including fluency, readability, and terminological precision, enters none of our metrics. The report metric also lacks a human-expert reference. Report scores should therefore be interpreted relative to the naive floor and across stages rather than as absolute measures of report quality.

S3.3 Model and Reasoning Limitations

CoB coverage was lower on the Domain Shift test set and on zero-shot report generation than on the in-distribution trained tasks. Repeated sampling partly recovered coverage without materially changing task performance (Appendix S2.1). Before deployment, CoB coverage should therefore be evaluated on data representative of the intended setting.

Box presence does not establish box faithfulness. Our evaluation measures box frequency and area and relates box presence to final-answer correctness, but it does not test whether a generated box accurately localizes the structure it names. The boxes expose spatial claims for human inspection but are not verified annotations. Their faithfulness requires clinically supervised assessment.

Near-universal coverage may require rejection sampling to select correct-answer trajectories with valid CoBs, followed by further SFT [15]. The same offline route could also serve open-ended tasks, filtering repeated CARDEA rollouts by complementary measures of report-quality dimensions that VS-$F_1$ leaves unscored, or by expert review, and fine-tuning on the survivors. Because this filtering is performed offline rather than used as an optimization reward, it avoids directly optimizing the policy against an open-ended reward model, the setting whose hacking risk motivated excluding open-ended tasks from RLVR.

S3.4 Clinical and Operational Readiness

Served with vLLM [42] on four NVIDIA B200 GPUs with tensor parallelism, end-to-end processing from raw multi-view videos through report generation took a median of roughly 6 seconds per study across 30 randomly sampled AngioCAD studies. In the first pass, component tasks run in parallel, so its wall-clock time is determined by the slowest task. These values are hardware-dependent. On less capable hardware, autoregressive decoding of a full CoB reasoning trace may be substantially slower than a single discriminative forward pass. CARDEA is therefore unsuitable for settings that require sub-second responses, including intra-procedural real-time guidance. Latency should be evaluated on the intended hardware before deployment.

CARDEA has not been validated in a clinical setting. Prospective evaluation against domain experts is a prerequisite for clinical use and should include blinded comparison with physician-authored reports.

Appendix S4 Qualitative Inference Examples

For each task, the example shown was drawn at random from the corresponding test set: the CoronaryDominance Real Distribution subset for dominance classification, the CardioSyntax test set for complexity assessment, and the held-out AngioCAD cohort for report generation. Sampling was restricted to cases whose reasoning trace contained at least one bounding box; for the two classification tasks, it was further restricted to cases in which CARDEA produced the correct final answer.