NSCLC Subtype Classification Using Patch-based CNN with EM Algorithm

Project Overview

This project implements a deep learning approach for classifying Non-Small Cell Lung Carcinoma (NSCLC) into its major subtypes: Adenocarcinoma (ADC) and Squamous Cell Carcinoma (SCC). The system uses a patch-based Convolutional Neural Network (CNN) with an Expectation-Maximization (EM) algorithm to identify discriminative regions in CT scans, enabling accurate cancer subtype classification.

---

Key Features

• Patch-based analysis of high-resolution CT scans
• Automatic identification of discriminative regions using EM algorithm
• Two-level classification approach with decision fusion
• Interactive web interface for clinical use
• High accuracy comparable to expert pathologists

---

Methodology

1. Patch-based CNN with EM Algorithm

This project implements the methodology described in the paper "Patch-based Convolutional Neural Network for Whole Slide Tissue Image Classification." The key insight is that not all regions of a medical image are equally informative for diagnosis.

Approach:

• Patch Extraction: Extracts multiple high-resolution patches from CT scans
• EM-based Discriminative Patch Selection:
  • Initially consider all patches as discriminative
  • Train a CNN model to predict cancer subtypes
  • Apply spatial Gaussian smoothing to probability maps
  • Select patches with higher probability values as discriminative
  • Iterate until convergence

Two-level Classification:

• First level: Patch-level CNN classification
• Second level: Decision fusion using logistic regression or SVM

2. CNN Architecture

The CNN architecture consists of:

• 5 convolutional blocks with batch normalization
• Adaptive pooling to ensure consistent feature map dimensions
• Fully connected layers for final classification

3. Decision Fusion

The system combines patch-level predictions using a Count-based Multiple Instance (CMI) learning approach:

• Creates histograms of patch-level predictions
• Trains a second-level classifier (logistic regression or SVM)
• Makes final image-level predictions

---

Technical Implementation

• Framework: PyTorch
• Dataset: NSCLC-Radiomics dataset
• Web Interface: Streamlit
• Image Processing: OpenCV, scikit-image, pydicom

---

Installation and Usage

Prerequisites

Install required packages:

pip install torch torchvision streamlit opencv-python scikit-image pydicom matplotlib numpy pillow

Running the Application

streamlit run deploy.py

Configuration

Create a .streamlit/config.toml file with the following content:

[server]
runOnSave = false
enableStaticServing = true

[runner]
fastReruns = false

---

Results

The model achieves high accuracy in distinguishing between ADC and SCC subtypes, comparable to the performance of expert pathologists. The patch-based approach with EM algorithm effectively identifies the most discriminative regions in the CT scans, improving classification accuracy.

---

Deployment

The project includes a Streamlit web application that allows medical professionals to:

• Upload DICOM files or CT scan images
• Visualize the uploaded images
• Get cancer subtype predictions with confidence scores
• View probability distributions for different subtypes

---

Future Work

• Extend the model to classify additional NSCLC subtypes
• Implement explainable AI techniques for better interpretability
• Integrate with hospital PACS systems for seamless clinical workflow
• Develop mobile applications for remote diagnosis

---

Disclaimer

This tool is for research purposes only and should not be used for clinical diagnosis without proper validation.

---

License

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

---

Acknowledgments

• Methodology based on the paper "Patch-based Convolutional Neural Network for Whole Slide Tissue Image Classification"
• NSCLC-Radiomics dataset for providing the training and testing data