brain-tumor-vit-candor

Brain Tumor Classification ViT (Candor Dust)

This is a Vision Transformer (ViT) fine-tuned for classifying brain MRI scans into four categories:

1. glioma
2. meningioma
3. notumor
4. pituitary

Model Description

The model is based on google/vit-base-patch16-224-in21k and fine-tuned on a dataset of brain MRI scans. It is used as the core classification engine in the Candor Dust application.

Intended Use

• Primary Use Case: Research prototype for AI-assisted brain tumor classification.
• Out of Scope Use: This model is NOT a certified medical device and should not be used for direct clinical diagnosis without human oversight.

Uncertainty & Explainability

This model deployment includes:

• Monte Carlo (MC) Dropout: During inference, the model runs 20 passes with dropout enabled to estimate predictive uncertainty. High uncertainty triggers a "Requires Human Review" alert.
• Grad-CAM (Gradient-weighted Class Activation Mapping): Generates heatmaps indicating which regions of the MRI scan contributed most to the prediction.

Usage

You can load the model directly via the transformers library:

from transformers import ViTForImageClassification, ViTImageProcessor
import torch
from PIL import Image

model_name = "diwaskunwar10/brain-tumor-vit-candor"
processor = ViTImageProcessor.from_pretrained(model_name)
model = ViTForImageClassification.from_pretrained(model_name)

image = Image.open("path_to_mri.jpg").convert("RGB")
inputs = processor(images=image, return_tensors="pt")
with torch.no_grad():
    outputs = model(**inputs)
    probs = torch.nn.functional.softmax(outputs.logits, dim=-1)
    predicted_class = probs.argmax(-1).item()
    print(model.config.id2label[predicted_class])

Disclaimer

⚠️ RESEARCH PROTOTYPE ONLY — NOT a certified medical device. Not for clinical use. Always consult qualified healthcare professionals for medical decisions.