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%
