Anchor-Free Detection

Instead of regressing offsets from thousands of predefined anchor boxes, anchor-free models directly predict the center point of each object on a heatmap. A peak in the heatmap indicates an object center, and the model simultaneously regresses the width and height of the bounding box from that point. This eliminates the need to predefine aspect ratios and scales, which is particularly advantageous for lesion localization where abnormalities like tumors have highly variable, non-canonical shapes.

Some architectures, like CornerNet, detect objects as pairs of keypoints—the top-left and bottom-right corners—without anchors. A follow-up, CenterNet, adds a third keypoint at the object center to verify that the paired corners belong to the same instance. This triplet approach significantly reduces false positive detections in dense medical scenes, such as overlapping organs in a CT scan, by ensuring geometric consistency before a bounding box is finalized.

Fully convolutional one-stage models like FCOS treat detection as a per-pixel prediction problem. For every pixel in the feature map, the model predicts a 4D vector encoding the distances to the four sides of the bounding box. To suppress low-quality detections far from object centers, a center-ness branch assigns a score that down-weights edge pixels. This dense prediction approach naturally handles objects of any scale, from micro-calcifications to large tumors, without anchor pyramid tuning.

Traditional anchor-based detectors require careful tuning of anchor scales, aspect ratios, and IoU thresholds for positive/negative assignment—hyperparameters that often fail to transfer across medical imaging domains. Anchor-free methods eliminate this combinatorial complexity. The number of design choices drops from dozens to just a few, such as the heatmap peak threshold. This makes the paradigm highly robust when adapting a model from chest X-rays to mammography, where lesion morphologies differ drastically.

DETR and its variants frame detection as a direct set prediction problem using a transformer encoder-decoder. The model predicts a fixed-size set of N objects in parallel, each represented by a class label and a bounding box, without anchors or NMS. Bipartite matching loss uniquely assigns each ground truth to one prediction. For whole slide image analysis, this end-to-end approach removes the need for complex post-processing pipelines, though it requires longer training schedules to converge on small objects.

Anchor-free models assign objects to feature pyramid levels based on their absolute size rather than anchor box IoU. For instance, FCOS defines a scale range for each FPN level—if a lesion's bounding box falls within that range, it is assigned to that level. This direct, geometry-based assignment is more interpretable and avoids the class imbalance problem where extreme aspect ratios cause anchors to go unmatched. It ensures small nodules are consistently routed to high-resolution feature maps.

The foundational concept that anchor-free methods explicitly reject. Anchor boxes are predefined bounding box templates with fixed scales and aspect ratios placed densely over an image. They serve as reference priors for predicting object locations. In medical imaging, defining anchors is problematic because lesion shapes are highly irregular and vary dramatically across modalities. Anchor-free methods eliminate the need to tune these sensitive hyperparameters, simplifying the pipeline and reducing the computational overhead of matching thousands of boxes.

The standard evaluation metric for object detection, used identically for anchor-based and anchor-free models. mAP calculates the area under the precision-recall curve for each class and averages the results across all classes at a given IoU threshold. For medical tasks, FROC (Free-Response ROC) is often preferred as it evaluates sensitivity against the average number of false positives per scan, a more clinically relevant measure than precision for screening applications.