Implementing vision transformer for classifying 2D biomedical images.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• The rapid increase in medical imaging data necessitates advanced machine learning algorithms for healthcare applications.
• Accurate classification of biomedical images is critical for disease diagnosis and treatment planning.
• The MedMNISTv2 dataset provides a diverse benchmark for evaluating 2D medical image classification models.

Purpose of the Study:

• To analyze the efficiency of the Vision Transformer (ViT) model on diverse medical imaging modalities within the MedMNISTv2 dataset.
• To assess ViT's capability in capturing intricate patterns for medical image classification.
• To establish new benchmark accuracies for BloodMNIST, BreastMNIST, PathMNIST, and RetinaMNIST datasets using ViT.

Main Methods:

• Selected four subsets from MedMNISTv2: BloodMNIST, BreastMNIST, PathMNIST, and RetinaMNIST, chosen for their diverse modalities and sample sizes.
• Pre-processed input images for model training.
• Trained the ViT-base-patch16-224 model on the selected datasets and evaluated performance using key metrics.

Main Results:

• Achieved new benchmark accuracies: 97.90% for BloodMNIST, 90.38% for BreastMNIST, 94.62% for PathMNIST, and 57% for RetinaMNIST.
• Demonstrated the Vision Transformer model's effectiveness in classifying diverse medical imaging data.
• The model substantially transcended existing benchmark metrics.

Conclusions:

• Vision Transformer models show significant promise for medical image analysis and classification.
• These findings support the adoption of ViT models in healthcare to enhance diagnostic accuracy.
• Further exploration of ViT models can aid medical professionals in clinical decision-making.