An interactive web application for brain tumor detection using a custom-trained CNN model with transfer learning. Users can upload MRI scans and get predictions in real time through a Streamlit frontend powered by a FastAPI backend using ONNX inference.
- Custom-trained model using DenseNet and transfer learning.
- Data augmentation applied to improve generalization
- High performance: ~89% accuracy and recall**
- Interactive web interface: Upload MRI scans and get instant predictions
- Fully Dockerized: Easy deployment with Docker and Docker Compose
- Lightweight: No database required. No need to have Tensorflow installed. Get inference using ONNX with a pre-trained model.
The model was trained on the Brain Tumor Classification MRI dataset:
https://www.kaggle.com/datasets/sartajbhuvaji/brain-tumor-classification-mri
The training process and model evaluation can be found in the Kaggle notebook:\ (subject to change) https://www.kaggle.com/code/arsalanjafari/tumor-vs-no-tumor-transfer-learning-cnn-acc-90
Clone the repository and run the app using Docker Compose:
git clone https://github.com/Arsalanjdev/CortiScan
cd CortiScan
docker-compose up --build- FastAPI backend: http://localhost:8000/docs
- Streamlit frontend: http://localhost:8501
Upload MRI scans via the Streamlit interface to get instant predictions of tumor vs no tumor.
- Python 3.11
- FastAPI – REST API backend
- Streamlit – Interactive frontend
- DenseNet + Transfer Learning – Model architecture
- Pillow – Image processing
- Docker & Docker Compose – Containerized deployment
- ONNX = Model Inference
Accuracy: 0.8883
Precision: 0.9169
Recall: 0.8883
| Class | Precision | Recall | F1-Score | Support |
|---|---|---|---|---|
| tumor | 0.7103 | 0.9810 | 0.8240 | 105 |
| no_tumor | 0.9920 | 0.8547 | 0.9182 | 289 |
| accuracy | 0.8883 | 0.8883 | 0.8883 | 0.8883 |
| macro avg | 0.8512 | 0.9178 | 0.8711 | 394 |
| weighted avg | 0.9169 | 0.8883 | 0.8931 | 394 |
This project is open-source under the MIT License.