Explainable Pneumonia Triage from Chest X-Rays: Grad-CAM and Monte Carlo Dropout for Uncertainty-Aware Selective Referral

Explainable Pneumonia Triage from Chest X-Rays explores how deep learning can be made more transparent and reliable for pneumonia detection. The project uses a transfer-learned DenseNet-121 model trained on the public Chest X-Ray Pneumonia dataset. Grad-CAM is used to show which parts of an X-ray influenced the model’s prediction, while Monte Carlo Dropout is used to estimate how uncertain the model is about each case. The model achieved an AUC-ROC of 0.9975, an accuracy of 96.76%, and an F1-score of 0.9774. Uncertainty was also closely linked to prediction errors: correctly classified images had a mean predictive entropy of 0.038 nats, compared with 0.414 nats for misclassified cases. A simulated triage experiment was then used to test whether uncertain cases could be referred for human review. Deferring the 30% most uncertain cases increased accuracy on the remaining cases to 99.76%, while 40% deferral resulted in perfect accuracy on the retained set. The project also shows why uncertainty alone is not always enough. Some incorrect predictions were made with very high confidence, while Grad-CAM still revealed that the model was focusing on less relevant image regions. This highlights how visual explanations and uncertainty estimates can complement each other in medical AI. This work is a research prototype and retrospective evaluation, not a clinically validated diagnostic system or medical device.