A Dual-Stage Noise Training Scheme for Breast Ultrasound Image Classification

Yiming Bian, Arun K. Somani

Abstract: Breast cancer is one of the most common types of cancer and among the top leading causes of death among women all around the world. An early diagnosis is extremely critical, and in-time treatments can greatly help to prevent the cancer cells from spreading. Mammography, MRI, biopsy, ultrasound, etc., are recognized as effective imaging tests for breast cancer. However, the diagnosis entirely depends on the experience and expertise of the radiologist. In the last decade, computer-aided diagnosis (CAD) systems have been developed as a secondary reference to provide an objective analysis of medical images. With the help of deep learning (DL) and convolutional neural networks (CNNs), the accuracy of intelligent systems on image tasks has been perpetually improved. In this work, we specifically focus on the breast ultrasound image classification task. We apply transfer learning to four widely used backbone CNN architectures in the medical image classification field: AlexNet, ResNet-18, ResNet-50, and VGG16. They are fine-tuned on carefully constructed noisy datasets, and the test results suggest that they all acquire remarkable noise resistance, and this immunity is almost invariant to noise intensity. We systematically formalize our methodology as a dual-stage noise training scheme and provide empirical parameter configurations for each backbone CNN. This scheme enjoys being simple, effective, and universal. We believe this study will benefit the development of the robust design of DL models in the medical area.

External IDs:dblp:conf/ic3k/BianS22a