Data-Efficient and Explainable Deep Learning for Medical Image Processing

Deep learning has demonstrated outstanding performance in medical image processing but its practical translation is inhibited by two recurrent difficulties, extensive annotation requirements and poor interpretability. In this research, we propose a data-efficient and explainable framework for illness identification from chest radiographs with pneumonia screening as a sample medical imaging problem. The pipeline we propose combines self-supervised pre-training, uncertainty-guided active learning, class-balanced augmentation and post-hoc visual explanations. We test our method on a simulated experiment and show that it gets an area under the receiver operating characteristic curve (AUC) of 0.925 from 1000 annotated images and is competitive even with 100 labels. The Grad-CAM style heatmaps for explainability analysis emphasize clinically significant lung regions to aid in model auditability and human-in-the-loop evaluation. Our results indicate that data-efficient learning and explainable artificial intelligence have the ability to simultaneously lower the labeling cost and boost the trust in medical image analysis systems.