Authors: Rabins Porwal
Abstract: — Deep convolutional networks now match or exceed specialist performance on several image-based diagnostic tasks, yet their adoption in clinical practice remains limited by a problem that accuracy alone cannot solve: a clinician asked to act on a prediction cannot see why it was made. Post-hoc saliency methods offer a partial answer, but different methods applied to the same network routinely disagree, and a map that looks convincing is not necessarily one that reflects the computation the model actually performed. This paper proposes an Explainable Deep Learning (XDL) framework in which interpretability is a training objective rather than an afterthought. An attention-guided refinement module reweights backbone feature maps so that spatial evidence is concentrated before classification. Three complementary attribution methods – Grad-CAM++, integrated gradients and GradientSHAP – are then fused into a single saliency map, and the disagreement among them is quantified as an explanation-consistency score. Finally, a deletion-based faithfulness penalty is added to the loss, so that the network is optimised not only to classify correctly but to concentrate its evidence on regions whose removal genuinely changes the prediction. Evaluation across five public datasets spanning radiography, dermoscopy, fundus photography, magnetic resonance imaging and histopathology gave a mean accuracy of 93.7 per cent, 2.1 percentage points above the strongest baseline. More importantly, faithfulness improved substantially: deletion AUC fell from 0.157 to 0.128 and agreement with expert-annotated lesion masks rose from 0.452 to 0.518 in intersection over union. A blinded review by three clinicians rated the fused maps 4.2 out of 5 for plausibility against 3.6 for the best competing method.
