RCNet‐V3: Gated Residual Refinement of Coarse Probability Maps for Breast Ultrasound Lesion Segmentation

ABSTRACT

Breast ultrasound probability refinement can recover missed foreground while introducing false positives. We evaluated RCNet‐V3, which refines a frozen ResNet34 U‐Net probability field using a parameter‐compact, soft‐gated signed correction module. The module receives the image, coarse probability and deterministic ambiguity marker (1−); error and boundary targets supervise the gated proposal and gate only during training. Three refinement seeds shared one coarse model and the same 98‐image BUSI test set. At matched Coarse/Fine thresholds of 0.50/0.50, image‐macro Dice increased from 0.757370 to 0.762903 (+0.005533 ± 0.000298), with recall increasing by 0.022718 and precision decreasing by 0.012518. Validation‐selected Fine thresholds of 0.47/0.46/0.47 against Coarse 0.50 gave a Dice increase of +0.006073 ± 0.000232. SDs describe variation among three seed means. Offline image‐paired analysis found larger gates and proposed magnitudes in coarse‐error and reference‐boundary regions, but correct regions also changed. Proposed signs were more often helpful for false negatives than false positives, and matched‐threshold Dice worsened in roughly one third of images. Locked single‐seed transfer to 1875 BUS‐BRA images from 1064 metadata Cases reduced image‐macro Dice by 0.004459 (case‐cluster bootstrap 95% CI, −0.006381 to −0.002456). The module added 102,402 trainable parameters; FLOPs and GPU model‐orward latency increased by 50.95% and 22.36%. These results characterise a modest internal overlap gain with a precision cost and show that spatial concentration does not ensure beneficial refinement or reliable external improvement.