TrueScan is a deep learning-powered system that detects AI-generated or manipulated knee X-rays and classifies them for anomalies. It uses an ensemble of three state-of-the-art CV models served through a full-stack clinical web dashboard.
3rd Place — ML/DL Track | State Level Inter-Collegiate Tech Exhibition at AIT, CKM
AI-generated medical images are increasingly indistinguishable from real scans. A radiologist or clinician relying on a manipulated X-ray for diagnosis could cause serious harm. TrueScan acts as a verification layer by flagging images that are likely synthetic or tampered with before they enter the diagnostic pipeline.
Rather than relying on a single model, TrueScan uses an ensemble of three architecturally distinct networks. Each model votes on whether the image is authentic or AI-generated, and the majority vote determines the final prediction. This approach reduces the risk of any single model's blind spots affecting the outcome.
| Model | Architecture Type | Role |
|---|---|---|
| ResNet50 | Residual CNN | Deep feature extraction |
| VGG19_BN | Classical deep CNN (Batch Norm) | Texture & pattern detection |
| YOLOv8m-cls | YOLO classification head | Fast, generalized classification |
Ensemble voting: A prediction is only marked as confident when at least 2 of 3 models agree.
- Real Knee Scans: Knee Osteoarthritis Severity Grading Dataset
- Fake Knee Scans: Synthetic (DeepFake) Knee Osteoarthritis X-ray Images from GANs
- Ensemble inference — ResNet50 + VGG19_BN + YOLOv8m-cls with majority voting
- Doctor authentication — role-based login and access management
- Automated PDF reports — downloadable reports with prediction confidence and patient details
- Full-stack dashboard — Flask backend with a clean clinical UI
ML: PyTorch, Ultralytics YOLOv8, ONNX
Backend: Flask, Python
Frontend: HTML/CSS/JS
Database: Supabase
Deployment: Local deployment only
git clone https://github.com/prajwalnayaka/TrueScan.git
pip install -r requirements.txt
cd Python_Scripts
python app.pyNote: Trained model weights are not included due to file size. Download from Google Drive and update the model paths in test.py accordingly.
-
Prajwal Nayaka T (GitHub)
- Trained the Core ML models (ResNet50, VGG19_BN, YOLOv8m-cls).
- Developed the Ensemble Voting Mechanism and Inference Pipeline.
- Developed the Report Generation module.
- Integrated above mentioned features into the Flask API.
-
Pragya MV (GitHub)
- Designed and developed frontend files along with styling.
- Developed the Report Generation module.
- Initialized the database structure.
- Built the organized Flask API.
- Train on a larger, more diverse dataset across multiple human anatomical regions
- Add Grad-CAM visualizations so clinicians can see which regions triggered the prediction
- Explore diffusion-model-generated fakes, which are harder to detect than GAN-generated ones
