Enhanced Multilevel Semantic Reasoning with Relation-Aware Fusion and Learned Feature Normalization for Automated Colorectal Cancer Gland Classification on Cell-Graph Representations

Abstract

Accurate classification of colorectal glandular tissue as normal or dysplastic is fundamental to early-stage colorectal cancer (CRC) diagnosis, particularly for pT1-staged adenocarcinomas where invasion depth determines therapeutic strategy. Existing computational pathology approaches rely on either pixel-level convolutional features or simplistic graph neural network embeddings, both of which inadequately capture the hierarchical morphological disruptions characteristic of glandular dysplasia. This paper introduces an enhanced deep learning framework(CRCNet) that operates on cell-graph representations of H&E-stained glandular tissue, integrating three synergistic innovations: (1) a Multilevel Semantic Reasoning (MSR) Encoder with three-tier hierarchical 1D-CNN feature extraction augmented by Squeeze-Excitation channel attention and multi-head self-attention at each resolution level; (2) a Gated Relation-Aware Fusion (G-RAF) module that explicitly models inter-feature dependencies through gated residual transformer blocks; and (3) Learned Feature Normalization (LFN) that replaces batch normalization with sample-wise learned affine transforms, eliminating train-test distribution mismatch. Evaluated on the public pT1-Gland-Graph benchmark (519 glands, 20 patients), CRCNetV2 with 3-seed ensemble achieves 98.08% overall accuracy, 98.04% F1-score, and 99.85% AUC-ROC under Protocol A (stratified 80/20 split), while maintaining 96.34±0.97% accuracy under the clinically rigorous Protocol B (patient-wise 4-fold cross-validation). Comprehensive ablation studies and comparisons against six baseline classifiers confirm the contribution of each architectural component. The proposed framework establishes a new performance benchmark for cell-graph-based glandular tissue classification.