ViSwNeXtNet Deep Patch-Wise Ensemble of Vision Transformers and ConvNeXt for Robust Binary Histopathology...

Area of Science:

• Computational pathology
• Artificial intelligence in medicine
• Digital pathology

Background:

• Intestinal metaplasia (IM) is a precancerous gastric condition requiring accurate diagnosis for early intervention.
• Traditional histopathological evaluation of H&E slides is labor-intensive and subject to interobserver variability.
• Deep learning, especially transformer models, shows promise for improving diagnostic accuracy in pathology.

Purpose of the Study:

• To develop and evaluate ViSwNeXtNet, a novel patch-wise ensemble framework using transformer-based models for diagnosing intestinal metaplasia.
• To assess the performance of ViSwNeXtNet on both custom-collected and public datasets.

Main Methods:

• Proposed ViSwNeXtNet framework integrates ConvNeXt-Tiny, Swin-Tiny, and ViT-Base transformer models for feature extraction.
• Features were concatenated, dimensionality reduced using iterative neighborhood component analysis (INCA), and classified with a quadratic SVM.
• Evaluated on a custom dataset (516 IM, 521 control) and the public GasHisSDB dataset (20,160 normal, 13,124 abnormal patches).

Main Results:

• ViSwNeXtNet achieved 94.41% accuracy, 94.63% sensitivity, and 94.40% F1 score on the custom dataset.
• On the GasHisSDB dataset, performance reached 99.20% accuracy, 99.39% sensitivity, and 99.16% F1 score.
• The model outperformed individual backbone models and demonstrated strong generalizability.

Conclusions:

• ViSwNeXtNet effectively combines local, regional, and global tissue features using an ensemble of transformer models.
• INCA-based feature selection significantly improved classification and reduced dimensionality.
• The findings support ViSwNeXtNet's potential for integration into clinical pathology workflows for improved gastric cancer prevention.