Lung Cancer Detection🔬

A deep learning system for classifying lung cancer from CT scans using CNN and Xception transfer learning. 📊 Clones dataset from GitHub, preprocesses images, trains a 4-class model (normal, adenocarcinoma, large cell, squamous cell), and predicts with visualizations. 🚀 Ideal for medical image analysis!

📋 Overview

This project implements a lung cancer prediction system using Convolutional Neural Networks (CNN) with transfer learning based on the Xception model pre-trained on ImageNet. It classifies chest CT scan images into four categories:

• Normal
• Adenocarcinoma
• Large Cell Carcinoma
• Squamous Cell Carcinoma

The codebase is a Jupyter notebook designed for Google Colab, leveraging GPU acceleration for efficient training. It includes data preprocessing, model training, evaluation, and prediction with random test image selection or user-uploaded images.

🛠️ Features

• Dataset Handling: Clones dataset from your GitHub repository and ensures correct folder structure for 4 classes.
• Data Augmentation: Applies transformations (flips, rotations, zooms) to enhance training data.
• Transfer Learning: Uses Xception model with frozen pre-trained layers for robust feature extraction.
• Model Training: Includes callbacks for early stopping, learning rate reduction, and model checkpointing.
• Evaluation & Visualization: Plots accuracy/loss curves and supports predictions with true/predicted label display.
• Flexible Prediction: Allows random test image selection or user-uploaded images for inference.

📂 Dataset

The dataset is sourced from your GitHub repository: YOUR_REPOSITORY_URL. It is organized into three directories:

• dataset/train/: Training images
• dataset/valid/: Validation images
• dataset/test/: Test images

Each directory contains subfolders for the four classes:

• normal
• adenocarcinoma_left.lower.lobe_T2_N0_M0_Ib
• large.cell.carcinoma_left.hilum_T2_N2_M0_IIIa
• squamous.cell.carcinoma_left.hilum_T1_N2_M0_IIIa

Note: If the dataset is incomplete (e.g., empty folders), you may need to populate it with images matching the class names. Ensure images are in PNG/JPG format.

🚀 Getting Started

Prerequisites

• Google Colab with GPU runtime
• Google Drive account (for saving models)
• Python 3.x with required libraries (see below)

Installation

1. Open the Jupyter notebook Lung_Tumor_Detection.ipynb in Google Colab.

2. Install dependencies by running the first cell:

   !pip install pandas numpy seaborn matplotlib scikit-learn tensorflow

Setup

1. Clone the Repository: Run the cell to clone the dataset and code from your repository:

   !git clone https://github.com/shahin-ro/Lung-Cancer-Detection.git
   %cd https://github.com/shahin-ro/Lung-Cancer-Detection.git

2. Mount Google Drive: Mount your Google Drive to save the trained model:

   from google.colab import drive
   drive.mount('/content/drive')

3. Verify Dataset: Ensure the dataset in ./dataset/ has the correct structure. The code automatically removes unexpected folders and ensures the four class directories exist.

Usage

1. Run the Notebook: Execute the cells sequentially in Google Colab:

   • Cell 1: Installs dependencies.
   • Cell 2: Clones your repository.
   • Cell 3: Mounts Google Drive.
   • Cell 4: Verifies and cleans dataset structure.
   • Cell 5: Loads and preprocesses data with ImageDataGenerator.
   • Cell 6: Defines and compiles the Xception-based model.
   • Cell 7: Trains the model with callbacks and saves it to Google Drive.
   • Cell 8: Evaluates the model and plots accuracy/loss curves.
   • Cell 9: Randomly selects a test image or allows uploading an image for prediction.

2. Prediction:

   • For random test image prediction, select option 1 in the prediction cell.
   • For custom image prediction, select option 2 and upload a PNG/JPG CT scan image via the Colab file picker.

Example Prediction

# Output for a random test image
Selected image: ./dataset/test/adenocarcinoma_left.lower.lobe_T2_N0_M0_Ib/image1.png
True class: adenocarcinoma_left.lower.lobe_T2_N0_M0_Ib
Predicted class: adenocarcinoma_left.lower.lobe_T2_N0_M0_Ib (Probability: 0.9273)
[Image displayed with true and predicted labels]

📈 Results

• Training Accuracy: ~93% (as reported in the original repository, varies with dataset size).
• Validation Accuracy: Monitored during training with early stopping to prevent overfitting.
• Test Accuracy: Evaluated on the test set, displayed in Cell 8.
• Visualizations: Training/validation accuracy and loss curves are plotted for performance analysis.

🛠️ Troubleshooting

• Empty Dataset Folders: If ./dataset/test/ or other directories are empty, populate them with images matching the class names. Check folder contents with !ls -l dataset/test/*.
• Model Not Found: Ensure the trained model is saved at /content/drive/MyDrive/trained_lung_cancer_model.h5. Re-run training if needed.
• Class Mismatch: The code ensures only 4 classes are used. Verify folder names match expected_classes.
• Image Format: Use PNG/JPG images for compatibility. Update the prediction code if other formats are used.
• Colab Memory Issues: Reduce BATCH_SIZE (e.g., to 4) or IMAGE_SIZE (e.g., to 224x224) if out-of-memory errors occur.

🙏 Acknowledgements

• Thanks to hallowshaw for the original project inspiration and dataset structure.
• The Xception model is based on TensorFlow/Keras implementations pre-trained on ImageNet.

📜 License

This project is licensed under the MIT License. See the LICENSE file for details.

📬 Contact

For issues or suggestions, open an issue on this repository or contact the maintainer.