Vision Transformers for Low-Quality Histopathological Images: A Case Study on Squamous Cell Carcinoma Margin...

Area of Science:

• Digital pathology
• Artificial intelligence in oncology
• Medical imaging analysis

Background:

• Squamous cell carcinoma (SCC) diagnosis is challenging, especially with poor imaging quality in resource-limited areas.
• Accurate SCC margin classification is crucial for effective surgery and preventing recurrence.

Purpose of the Study:

• To develop and evaluate a vision transformer (ViT)-based model for improved SCC margin classification.
• To address limitations of convolutional neural networks (CNNs) with low-quality histopathological images.

Main Methods:

• A customized ViT architecture was employed using transfer learning.
• Performance was assessed via five-fold cross-validation and compared against leading CNN models.
• Ablation studies investigated the impact of architectural choices.

Main Results:

• The ViT model achieved higher accuracy (0.928) and AUC (0.927) compared to the best CNN (InceptionV3: accuracy 0.86, AUC 0.837).
• ViT demonstrated superior performance on low-quality histopathological images.
• Architectural configurations significantly influenced diagnostic performance.

Conclusions:

• ViTs offer transformative potential for histopathological analysis, particularly in resource-limited settings.
• The ViT model enhances diagnostic accuracy and reduces reliance on high-quality imaging.
• This approach presents a scalable solution for global cancer diagnostics.