Enhancing Injury Segmentation in Breast Mammograms Through Semi-Supervised Learning

In this study, we aim to evaluate whether using information from medical reports for a large private dataset alongside small labeled datasets can lead to improved model performance. First, we introduce a pipeline to obtain a large imprecisely labeled dataset from private images and their associated expert reports for the task of detecting injuries in mammograms. We then evaluated several different methods for using imprecise labels in training AI models and compared the results to when the model was trained using only a small accurate dataset. This research brings the following achievements. Introducing a new pipeline for labeling mammograms in order to identify damage without the need for labels in the initial training data. This study is a comprehensive effort to evaluate the effectiveness of different solutions for using imprecisely labeled datasets in lesion segmentation in mammograms.