Skip to content
Open access

An attention-guided EfficientNetV2S–vision transformer framework for robust histopathological cancer classification

2026 · Journal of Computational Mathematics, Intelligent and Secure Systems · 0 citations

Abstract

Automatic detection of metastatic cancers from histopathological images still poses great difficulty due to complex cellular structure and contextual dependencies exhibited by cancer, as well as variability in tissue appearance. these problems, This research presents a new attentionguided hybrid framework for binary cancer classification from histopathological images through local feature extraction using EfficientNetV2S and contextual feature representation learning via transformers. The obtained convolutional feature maps are transformed into token representations and analyzed through multihead selfattention in order to capture long-term dependencies in image patches. A thorough data preprocessing and augmentation pipeline is introduced to make the system more robust to different variabilities. The framework is benchmarked on the Histopathologic Cancer Detection dataset and is compared against traditional deep learning methods. The ViT-Hybrid framework using TTA resulted in 95.49% accuracy, 95.31% precision, 93.48% recall, and 94.39% F1 score on the held-out validation set. It can be seen that incorporating the convolutional feature extraction with the attention mechanism to model the context proves more beneficial for histopathological image classification than the use of traditional CNN-based frameworks. This framework provides a strong deep learning solution for automatic cancer detection from the histopathological image patches. Traditional CNN-based frameworks provide good local morphological features extraction but fail to understand the long-range context of tissues Transformer models overcome this constraint but necessitate efficient feature representation. In this research work, we fill this void by presenting an attention-based EfficientNetV2S–Vision Transformer model which utilizes both local and global feature extraction for reliable cancer classification in histopathology images. Our model can be applied in computer aided diagnosis systems to detect cancer at an early stage.

Read PDF

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.