A Mammography Data Management Application for Federated Learning

This study aimed to develop and assess an application designed to enhance the management of a local client database consisting of mammographic images with a focus on ensuring that images are suitably and uniformly prepared for federated learning applications. The application supports a comprehensive approach, starting with a versatile image-loading function that supports DICOM files from various medical imaging devices and settings. It also aims to standardize the labeling and pre-processing of new images, statistical analysis and data visualization of mammographic images across all participating healthcare units. Initial image preprocessing is significantly enhanced through the use of Wiener and CLAHE filters, aimed at reducing noise and improving contrast, respectively, to ensure the highest quality of images for diagnostic purposes. Further refinement in the preprocessing pipeline is achieved with a U-Net model, trained on publicly available databases, which excels in segmenting the breast tissue from images, thereby eliminating irrelevant background and artifacts. This meticulous preparation of images not only standardizes data quality across multiple medical institutions but also facilitates collaborative model training within federated learning frameworks. The program allows for the review of images and their metadata, enables labeling of images with the ability to mark regions of interest (ROI), and utilize a pre-trained model for preliminary BI-RADS classification. A notable addition to the application is the integration of functionalities, thanks to the implementation of Grad-CAM model, designed to elucidate the decision-making processes of deep learning models. This integration further enriches the application's utility in supporting diagnostic and analytical tasks in mammography, providing clear insights into the interpretive reasoning behind model predictions.