A Context-Adaptive Dual Cross-Attention Fusion Framework with Incremental Learning for Robust Lung and Colon Histopathology Classification
TL;DR
The authors' model combines deep spatial features from the VGG13 architecture with multiple handcrafted domain features, such as colour histograms, Local Binary Patterns, and Grey-Level Co-Occurrence Matrices, using a token-based cross-attention module, and outperformed all baseline feature fusion approaches considered.
Abstract
The accurate recognition of lung and colon histopathological images is still one of the most difficult tasks within computational pathology, mainly due to the morphological similarities of different malignancies. This study proposes a novel Context Adaptive Dual Cross Attention Feature Fusion with Incremental Learning framework for efficient multi-class histopathology image classification on the LC25000 dataset. Specifically, the authors' model combines deep spatial features from the VGG13 architecture with multiple handcrafted domain features, such as colour histograms, Local Binary Patterns (LBPs), and Grey-Level Co-Occurrence Matrices (GLCMs), using a token-based cross-attention module. To stabilise attention learning by considering the impact of tissue morphology on affinities, the authors propose a context-adaptive bias modulation strategy. Moreover, a token importance reweighting method is considered to increase the contribution of discriminative features. To address the problem of catastrophic forgetting in multi-class histopathological classification, the incremental learning method is applied. Experimentally, the authors' method achieved an accuracy of 0.9869, an F1-macro score of 0.9862, an AUC-macro score of 0.9924, and a Cohen's kappa of 0.9838. The proposed solution outperformed all baseline feature fusion approaches considered. In particular, the model demonstrated promising results for the lung subtype classification task.