📌 Overview This project detects and classifies lesions in DICOM medical images using DeepLesion (Faster R-CNN) for detection and ResNet50 for classification.

⚡ Features ✅ Lesion Detection – Finds lesion locations using DeepLesion (Faster R-CNN). ✅ Lesion Classification – Identifies lesion type using ResNet50. ✅ False Positive Filtering – Removes misclassified structures (e.g., blood vessels). ✅ Bounding Box Overlays – Draws lesion location & label on the image.

🛠️ Installation 1️⃣ Install Required Dependencies bash Copy Edit pip install torch torchvision pydicom numpy matplotlib opencv-python 2️⃣ Clone the Repository bash Copy Edit git clone https://github.com/YOUR_GITHUB_USERNAME/DICOM-Lesion-Detection.git cd DICOM-Lesion-Detection 3️⃣ Clone DeepLesion (Required for Detection) bash Copy Edit git clone https://github.com/rsummers11/CADLab.git cd CADLab/deep-lesion This contains the pretrained DeepLesion model for detecting lesions.

4️⃣ Run the Script Modify the DICOM file path in Dicome_processing.py:

python Copy Edit filepath = r"C:\Users\ccape\Downloads\Radiology_script\sample.dcm" Then, run:

bash Copy Edit python Dicome_processing.py 📂 Project Structure graphql Copy Edit DICOM-Lesion-Detection/ │── Dicome_processing.py # Main script for detection & classification │── sample.dcm # Example DICOM file (optional) │── README.md # Project documentation │── CADLab/ # DeepLesion (cloned repo) │ ├── deep-lesion/ # Pretrained model & utilities 📊 How It Works 1️⃣ Load & Preprocess DICOM Reads DICOM images and extracts metadata. Converts grayscale images to RGB (for compatibility). 2️⃣ Detect Lesions DeepLesion (Faster R-CNN) locates potential lesions. Filters false positives like thickened blood vessels. 3️⃣ Classify Lesion Type ResNet50 model classifies detected lesions as: Tumor Cyst Hemorrhage Inflammation 4️⃣ Display Results Bounding boxes highlight detected lesions on the DICOM image. Labels & confidence scores annotate each lesion type. 📌 Example Output Detected Lesions: ✅ Tumor (Score: 0.92) – Bounding Box: (120, 80, 200, 160) ✅ Cyst (Score: 0.85) – Bounding Box: (300, 240, 380, 320)

📷 Annotated Image Output (Lesion Bounding Boxes & Labels)

🚀 Next Steps 🔍 Explainability (Grad-CAM) ✅ Generate heatmaps showing which lesion areas influenced classification.

🤖 Multimodal AI (Lesion Detection + Clinical Reports) ✅ Use NLP models to summarize findings automatically.

📄 DICOM Export with Annotations ✅ Save annotated images back into DICOM format for PACS/Radiology.

📢 Contributing Want to improve this project? Contributions are welcome! 🚀

Fork the repo Create a feature branch Commit your changes Open a pull request 🏆 Acknowledgments Built using:

DeepLesion (NIH) – Universal lesion detection dataset. TorchVision Faster R-CNN – Object detection model. RadImageNet – ResNet50 trained on medical images.