Masked R-CNN for Pulmonary Nodule Detection in Chest X-rays

Overview

This repository contains the implementation of a Masked R-CNN model for detecting pulmonary nodules in chest X-ray images. The model is trained and evaluated on the NODE21 dataset, which combines images from PadChest and ChestX-ray14 datasets, along with a specially curated test set from four Dutch hospitals.

Dataset: NODE21

• Dataset Link: Node21.grand-challenge.org/

Training Set

• Image Type: Posterior-anterior (PA) chest X-rays
• Composition:
  • Images with pulmonary nodules (labeled)
  • 1,500 random images without nodules from PadChest and ChestX-ray14

Test Set

• Size: 300 CXR images
• Source: Four different hospitals across the Netherlands
• Positive Case: Contains a pulmonary nodule, confirmed by CT scan after 3 months
• Negative Case: No pulmonary nodule, confirmed by CT scan after 6 months

Model Architecture: Masked R-CNN

Masked R-CNN is a deep learning model designed for medical image analysis. It processes regions of interest within medical scans through the following steps:

1. Generating region proposals via selective search
2. Extracting features using a CNN
3. Classifying anatomical structures or abnormalities
4. Refining locations of detected objects

Training Details

• Optimizer: Stochastic Gradient Descent (SGD)
• Learning Rate: 0.001
• Momentum: 0.9
• Weight Decay: 0.0005
• Epochs: 50

Evaluation Metrics

• Mean Average Precision (mAP)
  • IoU Threshold: 0.6
  • Used during both training and testing phases
• Intersection over Union (IoU)
  • Directly assesses the quality of generated bounding boxes