Topology-Aware Multi-Scale Deep Learning Framework for Automated Lung Histopathological Image Classification

<jats:p>Lung cancer is still one of the most common causes of cancer-related mortality worldwide, and histopathological testing is the clinical gold standard for diagnosis confirmation and differentiation of major cancer subtypes. Despite the fact that deep learning has significantly advanced the automated classification of histopathological images, current methods often fall short in effectively incorporating multi-scale tissue features and preserving higher-order structural relationships, which leads to limited feature discriminability and suboptimal classification performance. To overcome these challenges, in this paper, we offer a novel deep learning framework based on Hierarchical Adaptive Multi-scale Feature Fusion (HAMFF) and Topology-Aware Adaptive Simplicial Representation Learning (TASRL) for automated lung histopathology image classification. First, an Adaptive Stain and Image Normalization (ASIN) module normalizes tissue appearance and reduces staining variability, then tissue region segmentation is performed with a Lightweight Hypercomplex U-Net (LWHU). Then, multi-resolution tissue representations are extracted using a Learnable Wavelet Packet Transform (LWPT). The proposed HAMFF module adaptively fuses the complementary wavelet-domain and texture-domain features to generate a compact and discriminative feature representation.The proposed TASRL module preserves the higher-order spatial and topological relationships via adaptive simplicial representation learning before classification using a Binarized Simplicial Convolutional Neural Network (BSCNN). In addition, the network hyperparameters are optimized by the Triangulation Topology Aggregation Optimizer (TTAO) for better convergence stability and classification performance. The proposed framework is experimentally evaluated on publicly available LC25000 dataset and it achieves overall accuracy of 99.48%, precision of 99.44%, recall of 99.40%, F1-score of 99.42%, specificity of 99.73%, AUC of 0.9987, Matthews correlation coefficient (MCC) of 0.992 and Cohen's Kappa of 0.992. Comprehensive comparative analyses and extensive ablation investigations further show that the proposed HAMFF and TASRL modules consistently boost multi-scale feature representation, topology-aware tissue characterisation, and classification resilience compared to recent state-of-the-art methods. The suggested system offers an accurate, resilient and computation-efficient approach for computer-aided classification of lung histopathology images.</jats:p>