https://github.com/sarathir-dev/multi-scale-vit-for-3d-medical-imaging
A PyTorch implementation of a Multi-Scale Vision Transformer (ViT) for 3D medical image classification using the OrganMNIST3D dataset. This project explores multi-scale attention mechanisms to enhance classification performance in volumetric medical imaging.
https://github.com/sarathir-dev/multi-scale-vit-for-3d-medical-imaging
3d-imaging deep-learning healthcare medical-imaging medmnist multi-scale-model pytorch self-attention tranformers vision-transformer volumetric-data
Last synced: 3 months ago
JSON representation
A PyTorch implementation of a Multi-Scale Vision Transformer (ViT) for 3D medical image classification using the OrganMNIST3D dataset. This project explores multi-scale attention mechanisms to enhance classification performance in volumetric medical imaging.
- Host: GitHub
- URL: https://github.com/sarathir-dev/multi-scale-vit-for-3d-medical-imaging
- Owner: sarathir-dev
- Created: 2025-03-03T03:16:57.000Z (3 months ago)
- Default Branch: main
- Last Pushed: 2025-03-03T04:03:00.000Z (3 months ago)
- Last Synced: 2025-03-03T04:26:13.009Z (3 months ago)
- Topics: 3d-imaging, deep-learning, healthcare, medical-imaging, medmnist, multi-scale-model, pytorch, self-attention, tranformers, vision-transformer, volumetric-data
- Language: Python
- Homepage:
- Size: 3.91 KB
- Stars: 0
- Watchers: 1
- Forks: 0
- Open Issues: 0
-
Metadata Files:
- Readme: README.md
Awesome Lists containing this project
README
# Multi-Scale ViT for 3D Medical Imaging
A PyTorch implementation of a Multi-Scale Vision Transformer (ViT) for 3D medical image classification. This model utilizes self-attention and multi-scale feature extraction to analyze volumetric medical images using the OrganMNIST3D dataset.
## Features
- 3D Vision Transformer (ViT) for volumetric medical imaging
- Multi-Scale Attention Mechanism for feature extraction
- OrganMNIST3D dataset from MedMNIST
- Sinusoidal Positional Embeddings for 3D spatial encoding
- Patch-Based Image Tokenization
- GPU Acceleration (CUDA) for efficient training
## Dataset
This project uses the OrganMNIST3D dataset from MedMNIST, which consists of 3D grayscale volumes of 11 organ classes.
### Download the Dataset
The dataset is automatically downloaded using `medmnist`:
```python
from medmnist.dataset import OrganMNIST3D
train_dataset = OrganMNIST3D(split="train", download=True)
```
## Installation
Clone the repository and install dependencies:
```bash
git clone https://github.com/yourusername/Multi-Scale-ViT-3D.git](https://github.com/sarathir-dev/Multi-Scale-ViT-for-3D-Medical-Imaging.git)
cd Multi-Scale-ViT-3D
pip install -r requirements.txt
```
## Model Architecture
The model follows the Vision Transformer (ViT) architecture, adapted for 3D data:
- Patch Embedding: Converts 3D volumes into smaller patches
- Multi-Head Self-Attention (MHSA): Extracts global features
- Feed-Forward Networks (FFN): Enhances representation learning
- Classification Head: Outputs predictions for 11 organ classes
## Training
Run the training script:
```bash
python training/train.py
```
## Evaluation
To evaluate the model on the test dataset:
```bash
python training/test.py
```
## Results
The model is trained for 10 epochs using CrossEntropyLoss and the Adam optimizer. Performance is measured using accuracy on the test set.
## Contributing
Contributions are welcome. Please open an issue or pull request to improve the project.