MIRAM: Masked Image Reconstruction Across Multiple Scales for Breast Lesion Risk Prediction

Self-supervised learning (SSL) has attracted significant interest in machine learning and computer vision communi- ties. Two prominent SSL approaches have been contrastive-based learning and self-distillation with the help of cropping augmenta- tion. Recently, masked image modeling has emerged to be a more powerful SSL approach that uses image inpainting as a pretext task. Following this success, our paper introduces a scalable and practical SSL approach based on more challenging pretext tasks to facilitate the learning of powerful features. Specifically, we use multi-scale image reconstruction from randomly masked input images as the basis for feature learning. We hypothesize that the reconstruction of high-resolution images will help the model attend to smaller but finer spatial details, which is especially useful for recognizing subtle details in medical images. Our method is the first work studying the effect of multi-scale SSL on medical data. The proposed SSL features significantly improve the performance of the Curated Breast Imaging Subset of Digital Database for Screening Mammography (CBIS-DDSM) dataset. Specifically, our method increases the average precision (AP) and the area under the receiver operating characteristic curve (AUC) by 3% and 1% in pathology classification, and by 4% and 2% in mass margins classification when being compared with state- of-the-art (SOTA) algorithms.

Download paper here