Coronary Angiography Stenosis Detection: A Practical High-Recall-First Approach

From YOLO's Cross-Domain Failure: Annotation Format Is the Culprit

In medical imaging AI, stenosis detection in coronary angiography is a high-risk, high-value task. Unlike conventional object detection, in clinical settings the cost of a missed detection far outweighs that of a false positive—as long as false positives don't overwhelm the entire image, physicians are fully capable of quickly reviewing them. A developer recently shared their project approach on Reddit: abandoning precision-first in favor of high recall, while exploring core pain points such as inconsistent data annotation and distorted evaluation metrics. This article draws on that practical experience to outline technical paths and evaluation traps worth attention in coronary stenosis detection.

The developer initially tried training a YOLO model on an internal dataset, where in-domain performance was acceptable, but performance dropped sharply when switching to another dataset. Investigation revealed that the problem wasn't entirely about model generalization, but rather that the two datasets had completely different annotation conventions: one used fixed-size bounding boxes around lesion points, while the other used bounding boxes that tightly fit the lesion contour.

This reveals an easily overlooked fact: the same disease may be defined as two different detection tasks at the annotation level. A fixed-box model learns "the approximate location of the lesion center," while a tight-box model learns "the precise extent and shape of the lesion." When the model is deployed across datasets, both the input distribution and annotation semantics shift simultaneously, making performance collapse almost inevitable. For slender, morphologically variable lesions like coronary stenosis, unifying annotation conventions takes priority over model selection.

Two common paradigms for bounding box annotation: In object detection, bounding boxes typically follow one of two annotation orientations. One is the fixed-size box, which assigns a preset square region to each lesion point—fast to annotate and highly consistent, but encoding only the center location. The other is the tight box, which requires box edges to closely follow the lesion's outer contour—richer annotation information, but costly to annotate and heavily subject to subjective influence. At the loss function level, the two are equivalent to having the model learn different spatial distributions: the regression target for a fixed-box model is a smoothed location heatmap, while a tight-box model must simultaneously fit size and shape. Therefore, when training and testing come from different annotation paradigms, the model is effectively performing two different tasks. For coronary stenosis, where morphology varies along the vessel's long axis, the size variance of tight boxes is enormous, and the bias of fixed boxes is even more pronounced—cross-domain performance degradation is nearly unavoidable. Unifying annotation conventions or introducing annotation-semantics-adaptive domain adaptation methods is a more fundamental solution than mere hyperparameter tuning.

Vessel Segmentation + Radius Profile: An Interpretable Candidate Generation Path

The developer already had prerequisite capabilities in vessel segmentation and centerline extraction. Their first instinct was to compute the radius profile along the vessel centerline, flag candidate stenoses through abrupt local narrowing of the radius, and then use a neural network to filter the candidates. This approach has inherent clinical interpretability: the essence of stenosis is a reduction in luminal radius.

From a technical implementation perspective, the radius profile method has several advantages. It directly leverages the geometric structure of vessels and can generate candidates without relying on large amounts of annotated data—the recall ceiling is determined by segmentation quality rather than the detection model. After candidate generation, the second-stage classification network only needs to determine "whether this location is a true stenosis," greatly reducing task difficulty and increasing tolerance for annotation noise. Similar "rules-first, learning-second" cascaded strategies have mature applications in industrial defect detection, and migrating them to coronary angiography scenarios is reasonable.

However, the radius profile method has high requirements for vessel segmentation accuracy. Bifurcations, overlapping vessel segments, and diffuse stenosis (as opposed to focal narrowing) are areas where this method tends to fail, requiring specially designed rules in post-processing or additional features.

Computing centerlines and radius profiles: Vessel centerline extraction is typically based on distance transforms of segmentation masks or skeletonization algorithms, yielding the topological structure of the vessel tree. By computing the shortest distance to the vessel wall point by point along the centerline, the radius profile is obtained. Stenoses appear as local minima in the radius curve, and candidates can be flagged by setting thresholds for the magnitude of decline and window length. The recall ceiling of this method depends on segmentation completeness and centerline continuity: if segmentation is missing at bifurcations, the radius profile is interrupted, causing missed detections. To mitigate this, multi-branch centerline tracking can be employed in bifurcation regions, or elliptical fitting of vessel cross-sections can be introduced to correct radius estimation in overlapping segments. Additionally, the radius decline in diffuse stenosis is gradual, with no obvious local minimum, so the cumulative radius change rate over a sliding window should be used as a supplementary feature.

The Evaluation Dilemma: When the Gold Standard Itself Is Incomplete

The sharpest question the developer raised is: physician annotations do not mark every stenosis, so recall calculated from these labels is necessarily optimistic. This is a common challenge in medical imaging evaluation—incomplete gold standards make high recall look better than it actually is.

A pragmatic approach is multi-level evaluation. The first level calculates recall against lesions explicitly annotated by physicians, serving as a lower-bound reference; the second level calculates precision on a randomly sampled candidate set reviewed by experts; the third level introduces pseudo-recall estimation on unannotated data, for example, estimating the miss rate through cross-validation of two physicians reading independently. A more rigorous approach is designing a "Readers study," where radiologists judge each candidate output by the model one by one, yielding a recall estimate close to the true value.

Moreover, the annotation format differences mentioned by the developer also directly affect evaluation. If the matching rules used during evaluation (IoU thresholds, center-point distance) are inconsistent with the semantics of training annotations, recall figures lose comparability. It is recommended to fix a matching protocol aligned with clinical reporting conventions early in the project.

Recall bias under incomplete gold standards: In medical imaging evaluation, the denominator of recall is the total number of true positives. If physician annotations cover only a portion of them (i.e., missed labels), the denominator is underestimated and recall is overestimated. This bias is particularly significant in scenarios with many lesions and reading fatigue. A commonly used correction approach is the capture-recapture method: using annotation sets from two independent reading physicians, the number of missed labels is estimated from the ratio of intersection to union, thereby correcting recall. More rigorous is the multi-reader multi-case (MRMC) study design, where multiple physicians perform blinded judgments on model outputs and original annotations, with statistical comparisons using metrics such as the area under the ROC curve (AUC). Defining "acceptable miss rates" and matching rules (such as center-point distance thresholds) jointly with clinical experts early in the project is more meaningful than pursuing a single number after the fact.

Data and Resources: Options Beyond ARCADE

In terms of public data, ARCADE is currently the most commonly used benchmark dataset for coronary stenosis detection, but its coverage is limited. Developers can consider the following directions:

• Multi-center private data collaboration: Establish data-sharing agreements with hospital radiology departments to obtain real angiography data with structured reports.
• Vessel segmentation dataset transfer: Although retinal vessel datasets such as DRIVE and STARE are not coronary, they can be used to pretrain segmentation and centerline extraction modules.
• Synthetic data augmentation: Generate synthetic angiography images with known stenosis locations based on vessel tree models, to supplement samples of rare morphologies.

It should be noted that when transferring across datasets, annotation semantic alignment must be done first, otherwise the mistakes of the YOLO experiment will be repeated.

Is the Approach Off Track?

Returning to the developer's final question: is this path reasonable? Starting from the goal of high recall priority, a cascaded architecture separating candidate generation from candidate filtering is reasonable, and is easier to debug and interpret than end-to-end detection. What truly warrants vigilance is self-deception in the evaluation phase—if the gold standard is incomplete, any recall figure can only serve as a relative reference. It is recommended to involve clinical experts in designing the evaluation protocol early in the project, rather than considering how to quantify performance only after model training is complete.

For clinical assistive scenarios like coronary stenosis detection, a high-recall-first strategy aligns with actual workflow: the model screens, and the physician confirms. Delegating false-positive control to the second-stage filter and the physician's review efficiency, rather than sacrificing recall to pursue end-to-end high precision, is a wiser engineering trade-off.