MR ProGradeNet A Multi-Resolution Deep Learning Framework for Automated Prostate Cancer Grading from MRI.

Prostate cancer remains a leading cause of cancer-related morbidity in men, and its accurate aggressiveness grading is critical for optimizing treatment strategies. While multiparametric MRI (mpMRI) has revolutionized prostate cancer diagnostics through non-invasive visualization of anatomical and functional tissue contrasts, current automated grading frameworks face significant limitations due to inter-scanner heterogeneity, inadequate modeling of tumor microarchitecture, and insufficient clinical reliability. This study introduces MR-ProGradeNet (Multi-Resolution Prostate Grading Network), a novel deep learning architecture that systematically addresses these challenges via three synergistic modules. First, the A-HASH Preprocessing Module unifies modality-specific intensity profiles and spatial representations using learned histogram normalization and prostate-specialized UNet segmentation, ensuring scanner-invariant input harmonization. Second, the FRDF (Fractal Radiomic Deep Fusion) Module captures fine-grained angiogenic textures and high-level semantic abstractions through dual-path processing of fractal dimension-enhanced DCE MRI and EfficientNet-encoded T2/ADC modalities, fused via gated attention. Finally, the ProMMGrader Classification Module incorporates label distribution learning and confidence-weighted ensembling across modalities, reflecting Gleason-grade uncertainty and reducing misclassification risk. A comprehensive evaluation on the public PROSTATEx dataset demonstrates that MR-ProGradeNet achieves 97.5% accuracy, a 0.974 F1-score, and an AUC of 0.999, substantially outperforming traditional CNNs and state-of-the-art prostate grading methods. This work represents a clinically aligned and interpretable advancement in AI-based mpMRI analysis, capable of enhancing diagnostic confidence and reproducibility in real-world multi-center settings.