MTU-Net: Multi-Task Convolutional Neural Network for Breast Calcification Segmentation from Mammograms

Computer-Aided Detection (CAD) is a technology that helps radiologists identify malignant microcalcifications (MCs) on mammograms. By minimizing observational oversight, CAD enhances the radiologist's detection accuracy. However, the high incidence of false positives limits the reliance on these technologies. Breast Arterial Calcifications (BAC) are a common source of false positives. Effective identification and elimination of these false positives are crucial for improving CAD performance in detecting malignant MCs. This paper presents a model that can eliminate BACs from positive findings, thereby enhancing the accuracy of CAD. Inspired by the successful outcomes of the UNet model in various biomedical segmentation tasks, a multitask U-Net (MTU-Net) was developed to simultaneously segment different types of calcifications, including MCs and BACs, in mammograms. This was achieved by integrating multiple fully connected output nodes in the output layer and applying different objective functions for each calcification type instead of training different models or using one model with a shared objective function for different classes. The experimental results demonstrate that the proposed MTU-Net model can reduce training and inference times compared to separate multi-structure segmentation problems. In addition, this helps the model converge faster and delivers better segmentation results for specific samples.