Using deep learning to construct microcalcification clusters in a mammography prediction model
Po-Yen Hsu, Jia-Lang Xu, Li-Lun Chen, Mu‐Yen Chen · Innovation and Emerging Technologies · 2025
Breast cancer is one of the leading causes of cancer death among women in Taiwan, and it is also the type of cancer with the largest incidence worldwide. Early detection and treatment can effectively reduce mortality. In Taiwan, mammography is widely used to screen for microcalcifications in early breast cancer lesions. However, the characteristics of microcalcifications are difficult to observe, and helping radiologists more quickly and effectively identify microcalcifications is an important challenge. This article proposes a method for automatic lesion prediction and image segmentation using microcalcification cluster labeling data. U-Net and V-Net, two convolutional network architectures, are used for comparison, combining the standardized convolutional network architectures dropout and binary cross-entropy (BCE) with Dice as a loss function to detect microcalcification groups. Other preprocessing methods are used to optimize detection results, including window adjustment, image preprocessing, and data augmentation. Finally, the trained model is applied to mammography images to predict lesions through the image segmentation of microcalcification groups, thereby assisting radiologists in making diagnoses. The experimental results on the private dataset show that U-Net outperforms V-Net with an accuracy rate of 72% when using the original images for training, rising to 81% for training using preprocessed images, and 85% when applying the composite method to preprocessed images. V-Net training with preprocessed images only achieves 70% accuracy, rising to 72% by applying the composite method to preprocessed images.