Prostate Cancer Simulation and Prediction Using Diffusion Models and Grad-CAM++ on MRI Data
Abstract
Prostate cancer is the second most common cancer in men around the globe and reliable detection from multiparametric MRI (mpMRI) remains a challenge. There are three continuing challenges in deploying deep-learning models in this field: lack of manual annotations for training, model interpretability and non-stability of generative augmentation pipelines. In this paper, we introduce an end-to-end approach that can solve these three problems. A diffusion model is trained on the public PROSTATEx benchmark to generate realistic MRI slices and adding to the training set without manual annotation additional manual annotation. A simple Convolutional Neural Neural Network (CNN) on the real and synthetic data. corpus for binary classification of malignancy. Finally, Grad-CAM++ is used to obtain class-discriminative activation maps, localizing diagnostically regions and making model predictions explainable to clinicians. Experiments produce an accuracy of 76.9%, a sensitivity of 0.68, a specificity of 0.83, and a macro F1-score of 0.76 on the PROSTATEx test set. Qualitative saliency analysis shows that the network focuses anatomically relevant areas of the prostate. The proposed system provides a robust, data-efficient, interpretable pipeline for AI-assisted prostate MRI diagnosis.