An integrated model for early breast cancer prediction using microcalcifications and patient risk factors

Microcalcifications on mammograms are one of the earliest radiographic indicators of breast cancer and patient outcomes are significantly improved by early detection. The recent advances in deep learning have improved mammography interpretation to overlook the crucial context provided by patient specific data. This paper presents a novel dual-stream deep learning model that combines mammography image processing with clinical and patient demographic data to improve the accuracy and specificity of breast cancer risk prediction. Our method isolates and characterizes microcalcified areas in mammograms using perceptual saliency and junction tensor analysis, spotting minute patterns that could point to cancer. Segmentation is performed using U-Net, ResUNet and AttentionU-Net models, enabling precise delineation of microcalcified regions. Furthermore, the algorithm incorporates image attributes and patient data, such as age, breast density, tumor size and shape. This model offers a thorough and context-aware evaluation for early breast cancer prediction by combining clinical and image-derived variables into a single predictive framework. Segmentation evaluation on the MIAS and DDSM datasets demonstrates superior performance of ResUNet, particularly when using a combined loss function incorporating Dice score, Intersection over Union (IoU), and Hurdroff Loss. ResUNet achieves combined loss scores of 0.92 and 0.86 on MIAS and DDSM, respectively. Model performance is greatly enhanced by the incorporation of saliency, junction tensor, segmentation using ResUNet and classification using ResNet50, which attains the best accuracy on both datasets. On the MIAS, ResNet50 shows an accuracy of 96.6% and 98.2 % for the DDSM dataset.