Breast ultrasound image classification with hard sample generation and semi-supervised learning

Abstract

Ultrasound is a fast and non-invasive imaging technique widely used in the early-stage diagnosis of breast cancer. Ultrasound images are also commonly used in deep learning, but traditional computer algorithms and even ultrasound diagnosticians often struggle to accurately distinguish between benign and malignant cases, as some ultrasound images exhibit mixed characteristics of both. To address the above problems, we defined hard sample mining rules firstly. To address the above problems, we defined rules for hard sample mining and utilized shape descriptors (Concavity, Growth Direction, Compactness, Circle Variance, and Elliptic Variance) to describe the samples. For those samples with overlapped features of both categories, i.e., benign and malignant, we identify them as hard samples. In order to make the network pay more attention to hard samples, we propose a Phased GAN (PGAN) that can generate hard samples to increase the number of such samples in the dataset. However, since the generated hard samples lack labels, they cannot be used in supervised training. To overcome this limitation, we propose a new semi-supervised network (SSN) that trains them with labeled samples. We also add a hard sample loss function to the SSN to regularize the features of the hard samples, which helps the network learn these features better. In summary, we propose a breast ultrasound image classification using PGAN and SSN, which achieves a classification accuracy of 97.7% on a public breast dataset (445 cases of benign and 210 cases of malignant).