Self-Supervision and Weak Supervision for Accurate and Interpretable Chest X-Ray Classification Models

X-rays diagnose numerous thoracic diseases; however, accurate pathology detection requires trained radiologists. The number of available experts may impede population-level patient care, delaying medical action. State-of-the-art (SOTA) machine learning methods categorize chest X-rays across numerous diseases well but do not always account for explainability. Interpretability assessments rarely focus on the conciseness and anatomical correctness of Class Activation Mapping outcomes. These models are not accurate or dependable when tested on different datasets of the same modality. They are not used because their predicted performance is not explainable. This work introduces a self-supervised and weakly supervised pretraining pipeline with an auxiliary loss and supervised fine-tuning that retains performance across datasets. We use the Chest X-ray14 (NIH CXR) dataset for pre-training and CheXpert for fine-tuning. Our model is evaluated on chest X-ray classification tasks but is relevant to other imaging modalities and workloads. Our model outperforms supervised SOTA models on the downstream dataset in categorizing chest X-rays across 14 findings, improving Intersection over Union by 31% on the NIH CXR dataset. Our trained model offers compact, accurate, and interpretable pretrained representations to highlight anatomical sites with limited bounding box annotations.