Multi-Scale Contextual Segmentation for Early Breast Carcinoma Detection in Ultrasound

Abstract

Identifying early-stage breast carcinoma is crucial for improving patient outcomes, yet automated systems frequently struggle with accurate segmentation of small tumors in ultrasound images. This paper introduces a novel deep learning framework that leverages diverse receptive fields and multiscale feature integration to overcome the inherent limitations of fixed-kernel architectures. Our approach enhances the capture of fine-grained tumor localization and contextual information, significantly improving the segmentation of subtle breast lesions. Validated on public breast ultrasound datasets, this method demonstrates superior performance in isolating small carcinomas compared to existing segmentation techniques, marking a significant advancement for computer-aided early cancer diagnosis.