RTCMUNet: retinex-inspired feature modulation for breast ultrasound lesion segmentation

Abstract

Objective. Breast ultrasound lesion segmentation remains challenging because heterogeneous echogenicity, speckle, posterior shadowing, and weak boundaries impair lesion-margin delineation. Approach. We evaluate a CMU-Net-based implementation, termed RTCMUNet, in which the proposed ML-Retinex block modulates skip features before the inherited CMU-Net multi-scale attention gate. ML-Retinex combines learnable multi-scale projections, fixed Gaussian low-pass context estimates, clipped feature-ratio cues, and residual gating. It operates on learned skip features and neither estimates nor corrects scanner time-gain compensation. Nine models were evaluated on BUSI under a common three-run internal-validation protocol whose mutually disjoint validation subsets formed a pooled set of 389 cases. We assessed seven case-level metrics, paired Wilcoxon tests against CMU-Net with Holm correction, exploratory design-stage analyses, diagnostic controls, a secondary historical BUS/BUSI/TNSCUI benchmark, and a synthetic brightness-gradient stress proxy. Main results. RTCMUNet achieved Dice 0.826±0.220, IoU 0.748±0.240, HD95 20.65±32.94 px, and ASSD 7.14±14.04 px. Sensitivity was 0.848±0.212, precision 0.838±0.239, and BF1 0.570±0.291. Relative to CMU-Net, Dice and IoU increased by 0.017 and 0.021, while HD95 and ASSD decreased by 3.11 and 2.23 px. All seven paired tests remained significant after Holm correction (largest adjusted p=0.0104). RTCMUNet added 5.9% parameters and maintained 42 frames per second. Significance. Within retrospective internal validation, and without scanner-level gain data or an independent test set, the results support further evaluation of ML-Retinex as a compact feature-modulation strategy in the tested CMU-Net/BUSI setting.