An Explainable and Uncertainty Aware Hybrid Deep Learning Framework for Prostate MRI Image Classification

Abstract Background Artificial Intelligence systems for prostate Magnetic Resonance Imaging (MRI) require not only strong discrimination but also interpretable outputs and an indication of predictive uncertainty. This study develops an image-level framework that combines engineered MRI features, complementary neural learners, post-hoc explanations, and Monte-Carlo (MC) dropout. Methods DICOM images were represented by 265 features: a 256-bin normalized intensity histogram, five statistical descriptors, and four Gray Level Co-occurrence Matrix (GLCM) descriptors. A Multilayer Perceptron (MLP), one-dimensional Convolutional Neural Network (1D-CNN), and residual dense network produced class probabilities. These six probabilities were concatenated with the original features and processed by a 271-input stacking meta-learner. The supplied experiment used a stratified 60/20/20 training/validation/test split. Performance was assessed with accuracy, precision, sensitivity, F1-score, and Receiver Operating Characteristic-Area Under the Curve (ROC-AUC). LIME and SHAP offered both local and global explanations, and 50 stochastic forward passes estimated MC-dropout uncertainty. Results On the reported 1,200-image test partition, 98.25% of accuracy, 98.23% of F1-score, and 0.9907 of ROC-AUC were achieved by the hybrid model, exceeding the best base-model accuracy by 2.58 percentage points. The confusion matrix contained 596 true-negatives, 4 false-positives, 17 false- negatives, and 583 true-positives, corresponding to 99.32% precision and 97.17% sensitivity. Conclusion Probability-level fusion of heterogeneous learners can improve image-level discrimination while supporting feature attribution and uncertainty inspection. However, the current labels are derived from DICOM study descriptions rather than pathology-confirmed cancer status. The reported performance is therefore methodological and must not be interpreted as clinical cancer-detection performance until the model is rerun with patient-grouped partitions, training-only preprocessing and resampling, verified cancer labels, and independent external validation.