Breast cancer screening with class imbalance datasets of mammogram imaging
C. -M. Liang, Yao‐Chung Chang, C. -Y. Yeh, Sung-Chen Lin, Y. -J. Huang, Wei-Wen Hsu · IET conference proceedings. · 2024
In Taiwan, breast cancer ranks as the fourth leading cause of death among women. Many studies have reported that early detection and treatment are crucial in reducing mortality rates. Mammogram imaging is a key screening tool for the early detection and management of breast cancer. However, the identification and interpretation of breast lesions in mammography is not a trivial task. In addition, the extremely biased ratio of positive and negative case numbers makes it more challenging when it comes to the development of the computer-aided diagnosis (CAD) systems. Nevertheless, if we resample the negative cases to make the training dataset class distribution even with the positive cases, the total number of cases may not be sufficient for training a deep learning model. To avoid such dilemma, our strategy is to guide the deep learning models to learn the correct features of lesions by the annotations of lesions, including masses and micro-calcifications, from the medical physician. Subsequently, the deep features from the SPPF layer in a detection model of YOLOv5m are extracted for both positive and negative cases. Finally, a classifier of Random Forests is trained based on the extracted feature for mammography screening. For our experiments, the source datasets were from the competition of Screening Mammography Breast Cancer Detection held by Radiological Society of North America (RSNA) in late 2022. And the performance of our proposed reached to 0.624, 0.77, 0.596, and 0.672 for Accuracy, Sensitivity, Precision, and F1-Score, respectively.