NEUROXL-CRFNET: A HYBRID TRANSFORMER–GRAPH FRAMEWORK FOR AUTOMATED HEPATOCELLULAR CARCINOMA HISTOPATHOLOGY CLASSIFICATION

Abstract

In this research work, a novel deep learning-based framework for automated hepatocellular carcinoma (HCC) detection from histopathological images is introduced. The pipeline begins with a comprehensive pre-processing stage to enhance image quality and normalize color variability across slides. Macenko color normalization ensures stain-independent intensity profiles, while Gaussian filtering removes artifacts and de-noises tissue sections. Morphological operations segment the tissue regions of interest, after which image patches of size 256 × 256 px are extracted using an overlapping sliding window to capture diverse local features. For feature extraction, a Transformer-based IntentNet backbone models global contextual dependencies within each patch, producing rich attention-aware representations. The features are first ranked by Mutual Information (MI) to identify the most informative descriptors and then refined through a Hybrid Grey Wolf Optimizer (HGWO) that fuses Grey Wolf and Harris Hawk strategies for optimal feature subset selection. Finally, disease classification is performed by the proposed NeuroXL-CRFNet, a hybrid network that integrates a Neuro-Symbolic Graph Convolutional Network (NS-GCN) for structural feature reasoning, XLNet for contextual sequence modeling, and a Conditional Random Field (CRF) decoder for structured prediction. Extensive integration of these modules enables precise HCC detection, offering a robust, end-to-end solution for histopathological image analysis.