An explainable artificial intelligence model for identifying local indicators and detecting lung disease from chest X-ray images

One of the primary responsibilities of radiologists is to diagnose lung illness using chest X-ray images. The radiologist examines the patchy infection in the imaging and makes a rational decision based on their knowledge. Convolutional neural networks work incredibly well in classifying and identifying diseases from medical images. Despite being a promising prediction technology with accuracy equivalent to a person, deep learning (DL) models typically lack explainability, a crucial component for the clinical deployment of DL models in the highly regulated healthcare sector. In this paper, we mimic the radiologist’s decision-making process by identifying local discriminate regions of a chest X-ray image through weekly supervised learning and deriving rules, and explaining why the DL method gives such results. This process is carried out in three phases. Phase one is to train a model on a classification problem to predict lung disease. Phase two is identifying critical regions and training a model on the identified images with critical regions. Phase three combines the local and global features with learning more patterns to classify the diseases. The local and fusion models have shown remarkable improvement in getting 99.6 percent accuracy with fewer epochs.