딥러닝 기반 유방 초음파 영상의 분할 및 분류 통합 모델

Abstract

This study proposes an integrated multi-task learning model that performs both lesion segmentation and lesion-type classification simultaneously in breast ultrasound images. Conventional single-task approaches handle segmentation and classification independently, often leading to information isolation and a failure to leverage complementary features between the tasks. To overcome these limitations, we adopt a two-stage U-Net architecture with a ResNet-101 backbone and introduce the Hierarchical Gating Module (HGM) as the core component of our framework. HGM hierarchically reuses the coarse segmentation probability map generated in Stage 1 to modulate multi-scale encoder features in Stage 2, thereby maximizing the synergistic interaction between segmentation and classification. Experiments conducted on the BUSI breast ultrasound dataset demonstrate that the proposed HGMNet achieves superior performance compared to existing models, recording a Dice coefficient of 0.7431 for segmentation and an accuracy of 0.8500 for classification. These results indicate that the proposed model can effectively integrate the two tasks within a single unified network, thereby enhancing both the accuracy and reliability of breast ultrasound–based diagnosis.