Novel application of attention-enhanced hybrid efficient net for multi-scale histopathological cancer classification.

Examination of high-resolution whole-slide images requires an analysis of the histopathological images, which is essential in the precise diagnosis of cancer; nevertheless, manual analysis is time-consuming and is highly susceptible to inter-observer variability. This paper will address these shortcomings by developing a new Attention-Enhanced Hybrid EffNet (AEH-EffNet) system to automatically classify lung and colon cancer histopathology using AEH-EffNet. In contrast to traditional that directly use existing EfficientNet architectures, AEH-EffNet presents multi-scale patch-based learning, attention-based feature optimization, and adaptive backbone optimization of EfficientNet B0-B7 to effectively learn cellular-level textures and tissue-level contextual patterns and strike a balance between accuracy and computed cost. It consists of comprehensive preprocessing of data, such as stain normalization, noise elimination and resolution standardization to enhance the quality of data. The proposed framework is tested on a benchmark dataset of 25,000 histopathology images consisting of benign lung tissue, lung adenocarcinoma, lung squamous cell carcinoma, colon adenocarcinoma and benign colon tissue. As per the experimental findings, AEH-EffNet can reach a top 97.5% testing accuracy with testing loss of 0.10 loss on high-resolution input and 96% with lower training time on the adaptively chosen EfficientNet-B2 architecture, which represents a trade-off between performance and efficiency. These results confirm that AEH-EffNet is a strong, scalable, and clinically dependable system of automated cancer histopathology classification, which helps in more effective and quicker detection of diagnosis based on digital pathology.