ViT-Glial-Tumor-Classification

ViT-Glial-Tumor-Classification

This repository implements a Vision Transformer (ViT) model for classifying histopathological images of glial tumors into three major categories: Glioblastoma (GBM), Astrocytoma (Astros), and Oligodendroglioma (Oligos) at Diamandis Lab. The pipeline includes dataset preparation, balanced sampling, model fine-tuning, training/validation with Weights & Biases logging, and evaluation with confusion matrix visualization.

Dataset

• Histology images organized into folders by tumor type: GBM, Astros, and Oligos.
• For efficient training, a balanced subset of 18,000 images (6,000 per class) was selected for fine-tuning.

Model

• Pretrained Vision Transformer from Kaiko AI.
• Classification head fine-tuned on the glial tumor dataset.
• Last 5 transformer blocks are also fine-tuned for improved learning.

Training

• Stratified train/val/test split (60/20/20).
• Training includes W&B logging and model checkpointing.
• Early stopping is used to avoid overfitting.

---

Results

The model achieved ** classification performance**:

• Best Validation Accuracy: 94.92% (Epoch 2)
• Final Test Accuracy: 94.64%