Lung Segmentation using Medical Transformer (MedT)

Automated lung segmentation from chest X-ray images using a Medical Transformer (MedT) deep learning model.
This project demonstrates the use of transformer-based architectures for medical image segmentation, achieving superior results compared to traditional CNN models like U-Net or SegNet.

---

Team Members

• Janapareddy Vidya Varshini – 230041013
• Korubilli Vaishnavi – 230041016
• Mullapudi Namaswi – 230041023

---

Problem Statement

Develop an automated system to segment lung regions from chest X-ray (CXR) images using the Medical Transformer (MedT) model.
The goal is to build an accurate and efficient segmentation model capable of delineating lung regions with minimal supervision.

Lung segmentation is a crucial preprocessing step in detecting and analyzing pulmonary diseases.
Manual segmentation is time-consuming and prone to human error — hence, an automated deep-learning solution is proposed.

---

Dataset Structure

Dataset Info

• Total images: 800
• Matched image-mask pairs: 704
• Image dimensions: 2300–3000 px range (resized to 128×128)
• Data split:
  • Train: 70%
  • Validation: 15%
  • Test: 15%

---

⚙️ Data Preprocessing & Augmentation

Implemented through the ChestXrayDataset class:

• Resize: All images & masks → 128×128
• Binarize masks: Pixel >127 → 1 (lung), else 0
• Normalize images: Using ImageNet mean & std
• Augmentation (for training):
  • Random horizontal flips
  • Random brightness/contrast (0.8–1.2)
  • Geometrically consistent image-mask transforms
• Conversion: NumPy → PyTorch tensors (C, H, W)

---

Model Architecture — Medical Transformer (MedT)

MedT addresses the limitations of CNNs (poor long-range context) and vanilla Transformers (poor local detail).

Key Components

1. Gated Axial Attention (GAA):

   • Decomposes 2D attention into two 1D attentions (height & width).
   • Introduces learnable gates for positional encoding.
   • Efficient and ideal for medical imaging tasks.

2. Local-Global (LoGo) Training Strategy:

   • Global Branch: Operates on a downsampled image (128×128 → 32×32) for contextual understanding.
   • Local Branch: Processes small 4×4 patches for detailed local features.
   • Combined outputs capture both fine and global context.

---

Experimental Setup

Parameter	Value
Framework	PyTorch
Device	CUDA (GPU) if available
Image Size	128×128
Epochs	40
Batch Size	4
Optimizer	Adam
Learning Rate	0.0001
Scheduler	ReduceLROnPlateau (factor 0.5, patience 10)
Loss	Binary Cross-Entropy (BCE) + Dice Loss
Metrics	Accuracy, Precision, Recall, IoU, Dice Score

---

Results & Observations

• Best model achieved after 19th epoch
• Evaluated on the test set
• High Dice coefficient and IoU confirming segmentation quality
• Model generalized well on unseen X-rays

Metric	Description
Dice Coefficient (F1)	2TP / (2TP + FP + FN) — measures overlap
IoU (Intersection over Union)	TP / (TP + FP + FN)
Precision	TP / (TP + FP) — avoids over-segmentation
Recall	TP / (TP + FN) — measures detection completeness

---

Web Application Deployment

A Flask-based web interface was built to make the segmentation model accessible.

Features

• Upload chest X-ray image (.png)
• Model performs segmentation and returns mask
• Displays original + segmented output side by side
• Responsive UI built with HTML, CSS, JavaScript
• Backend uses Flask and pre-trained MedT model

Run the Web App

# Clone the repository
git clone https://github.com/Namaswi24/Lung-Segmentation.git
cd Lung-Segmentation

# Create virtual environment (recommended)
python -m venv venv
source venv/bin/activate    # For Linux/Mac
venv\Scripts\activate       # For Windows

# Install dependencies
pip install -r requirements.txt

# Run Flask app
python app.py