Towards Next-Generation Sustainable Cancer Diagnosis: A Hybrid Transformer-Enhanced Texture-Attention Framework with Multi-Objective Optimization

Sustainable cancer diagnosis with histopathological images has not yet been achieved as high computational cost, class imbalance, and lack of model interpretability are present. The currently available CNN and transformer-based models are computationally expensive, resulting into high performance, but limit their clinical applicability. The proposed paper suggests using Hybrid Transformer-Enhanced Texture Attention (TETA) framework, which combines the handcrafted Histogram Gray-Level Co-occurrence Matrix (HGLCM) texture features with transformer-based global attention. Enzyme Action Optimization (EAO) and Multi-Objective Optimization (MOO) are used in the optimization of the model to compromise between the diagnostic accuracy, energy consumption, and the speed of inferences. The novel TextureAware Attention Module is a biologically significant region detector and enhances the interpretability and diagnostic reliability. The CHAVI histopathology dataset was evaluated experimentally with 98.8% accuracy and 98.5% F1-score and used 60 J of energy and 18 ms of inference time - better than CNN, ResNet-50, ViT, and hybrid baselines. The suggested framework identifies a scalable, explainable, and energy-efficient cancer detection method based on AI, which is in line with the vision of Green AI of sustainable healthcare.