Automatic Joint Segmentation and Classification of Breast Ultrasound Images via MTL-OCA

Yaling Lu ¹ Fengyuan Sun ^2* Jingyu Wang ³ Kai Yu ^4*

• ¹ Sichuan Tianfu New District People's Hospital, Chengdu, Sichuan Province, China
• ² Guilin University of Electronic Technology, Guilin, Guangxi Zhuang Region, China
• ³ Xi'an Jiaotong University, Xi'an, China
• ⁴ Beijing Smartmore Intelligent Technology Co., Ltd, Beijing, China

Segmentation and classification of breast ultrasound (BUS) images is important for early breast cancer detection, which is also a hot topic in medical image processing and interpretation.Various machine learning and deep learning-based algorithms have been presented for the segmentation and classification of BUS images due to advancements in computer technology and contemporary mathematics. Inspired by these achievements, we suggest a multi-task learning network with an object contextual attention module (MTL-OCA) for segmenting and classifying BUS images. The proposed method employs an object contextual attention module to learn the pixel-region relationships for promoting segmentation masks. The classification task utilizes the high-level information extracted from unenhanced segmentation masks to promote the classification performance. Cross-validation tests are conducted using BUS datasets to evaluate MTL-OCA's performance. MTL-OCA achieves the best classification and segmentation results compared to several state-of-the-art methods.