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

Area of Science:

• Medical Imaging
• Artificial Intelligence in Healthcare
• Biomedical Engineering

Background:

• Breast ultrasound lesion segmentation is challenging due to factors like heterogeneous echogenicity, speckle, posterior shadowing, and weak boundaries.
• Accurate delineation of lesion margins is crucial for effective diagnosis and treatment planning.

Purpose of the Study:

• To evaluate a novel CMU-Net-based implementation, RTCMUNet, for improved breast ultrasound lesion segmentation.
• To assess the efficacy of the proposed ML-Retinex block in modulating skip features for enhanced lesion-margin delineation.

Main Methods:

• Developed RTCMUNet, incorporating an ML-Retinex block to modulate skip features before the CMU-Net multi-scale attention gate.
• Evaluated nine models on the BUSI dataset using a three-run internal-validation protocol with 389 cases.
• Assessed performance using seven case-level metrics and compared RTCMUNet against CMU-Net via paired Wilcoxon tests with Holm correction.

Main Results:

• RTCMUNet achieved a Dice score of 0.826 ± 0.220 and IoU of 0.748 ± 0.240.
• The model demonstrated improvements over CMU-Net, with a 0.017 increase in Dice and 0.021 in IoU, and reductions in HD95 and ASSD.
• All seven paired tests remained statistically significant after correction, indicating robust performance.

Conclusions:

• RTCMUNet shows promise as an effective feature-modulation strategy for breast ultrasound lesion segmentation.
• The results support further investigation of the ML-Retinex block within the CMU-Net framework for clinical applications.
• The model maintains efficient processing speeds (42 frames per second) with a modest increase in parameters.