GR-HDUNET: GAN-refined hybrid dense U-net with ensemble for colorectal cancer classification

Colorectal cancer is a leading cause of cancer-related mortality; early, accurate histopathological assessment is pivotal. We present a two-stage framework—GR-HDUNET plus a transformer ensemble—that unifies boundary-aware segmentation with multi-model classification. The segmentation stage combines a DenseNet encoder, multi-scale attention, residual decoding, and cGAN refinement to produce anatomically faithful masks, achieving class-average Dice = 0.895, Boundary F1 ≈ 0.82, HOG-similarity ≈ 0.92 on EBHI-Seg. These masks guide a three-branch classifier (Vision Transformer, Swin Transformer, ConvNeXt) whose predictions are fused by weighted averaging. On the six-class EBHI-Seg dataset, weighted averaging attains 95.1% accuracy, 94.2% F1, and micro-AUC = 0.95, surpassing the best single backbone by ≥1.0 percentage point. Qualitative overlays confirm crisp gland delineation, and ROC curves show clear class separation. Importantly, segmentation is trained with pixel-level masks, while classification uses slide-level labels. The pipeline is computationally tractable (≤1.6 s per 224×224 image on a single RTX 4090) and readily reproducible. By coupling cGAN-refined, boundary-aware segmentation with complementary transformer features, our approach delivers robust colorectal-tissue segmentation and classification on EBHI-Seg and offers a transferable template for digital pathology.