Comparative Bladder Cancer Tissues Prediction Using Vision Transformer.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Early bladder cancer detection is critical but challenging due to asymptomatic early stages.
• Expert evaluation of endoscopic images is time-consuming and can involve interdisciplinary collaboration.
• Computer-aided decision support systems can optimize expert time for patient treatment.

Purpose of the Study:

• To evaluate the performance of three distinct computational models for bladder tissue classification.
• To develop an accessible computer-aided decision support system for bladder cancer diagnosis.

Main Methods:

• Development and comparison of three models: a convolutional neural network (CNN)-based deep learning (DL) model, a hybrid DL+Machine Learning (ML) model, and a Vision Transformer (ViT) architecture.
• Training and testing models on a bladder tissue dataset.
• Implementation of a graphical user interface (GUI) for the decision support system.

Main Results:

• The Vision Transformer (ViT) architecture demonstrated superior performance metrics compared to the DL and DL+ML models.
• Accuracy scores were 0.9086 (DL), 0.9071 (DL+ML), and 0.9257 (ViT).
• F1 scores were 0.8884 (DL), 0.8496 (DL+ML), and 0.8931 (ViT).

Conclusions:

• The Vision Transformer model shows significant promise for accurate bladder cancer detection from endoscopic images.
• A computer-aided decision support system utilizing ViT can enhance diagnostic efficiency and potentially improve patient outcomes.