Lung-Nodule segmentation & classification

Deep learning pipeline for lung nodule detection and classification using LIDC-IDRI dataset. Includes preprocessing, UNet-based segmentation, ResNet-based malignancy classification, and full training/visualization workflows.

🧠 Model Overview

• Segmentation: 2D U-Net trained on CT slices with Dice + BCE loss.
• Classification: ResNet18 fine-tuned on extracted nodule patches.

📦 Requirements

To run this project, make sure you have the following packages installed:

pip install -r requirements.txt

🧩 Step 0: Preprocess LIDC-IDRI to Extract 2D Slices

Before training or running any models, you must extract 2D image slices and segmentation masks from the original LIDC-IDRI dataset.

⚠️ Important: Follow the instructions in
LIDC-IDRI-Preprocessing-master/PREPROCESSING_README.md

This step uses the pylidc library to:

• Parse raw DICOM scans
• Segment lungs
• Extract nodule masks based on consensus
• Save 2D .npy slices and their masks

Once completed, the following directories will be created:

LIDC-IDRI-Preprocessing-master/ ├── data/ │ ├── Image/ # 2D CT slices as .npy │ ├── Mask/ # Corresponding binary nodule masks │ └── Meta/ # CSV with metadata for each slice

or just replace the entire data folder in LIDC-IDRI-Preprocessing-master with this folder you can download here

📓 Step 1: Running the Pipeline

After installing requirements and running the preprocessing script, open the main Jupyter notebook to run the complete pipeline:

The notebook includes:

🔹 Segmentation Model Training (U-Net)

• Trains on 2D slices from preprocessed CT scans.
• Uses Dice + BCE loss with a positive pixel weighting strategy.
• Saves the best model as best_model.pth.

🔹 Patch Extraction

• Extracts 64×64 patches around nodules from the images using bounding boxes.
• Patches are saved under:
  • patches/benign/
  • patches/malignant/

🔹 Classification Model Training (ResNet18)

• Classifies each patch as benign or malignant.
• Best model saved as best_classifier.pth.

🔹 Joint Visualization

• Randomly picks a nodule slice and patch.
• Displays:
  • Original CT image
  • Ground truth segmentation
  • U-Net prediction
  • Classification result vs. ground truth

🏋️ Model Weights

• You can download the weights for the trained segmentation model here

• You can download the weights for the trained classifier here

📊 Results

• Segmentation Dice Score: ~0.75 (on validation set of 30 patients)
• Classification Accuracy: ~85% (benign vs malignant)
• Sample output:

🗂️ Dataset

• Source: LIDC-IDRI Dataset
• We used a subset of 150 patients for training and validation.
• Preprocessing handled using pylidc.

📚 References

• MONAI: Medical Open Network for AI - https://monai.io/
• Jaeho Lee. LIDC-IDRI-Preprocessing. GitHub repository.
• Matthew Hancock. pylidc. GitHub repository.
• The Cancer Imaging Archive. LIDC-IDRI. National Biomedical Imaging Archive.