Bowel Segmentation using SegFormer

An Automated bowel segmentation and obstruction detection pipeline from 3D MRI scans using SegFormer.

Overview

This project segments small bowel, large bowel, and stomach from MRI scans, then analyzes the segmentation masks to detect potential bowel obstruction by measuring bowel diameter against clinically-verfied bowel diameter thresholds.

Pipeline

• Preprocessing: N4 bias field correction, NLM denoising, and Image augumentations
• Segmentation: Fine-tuned SegFormer model achieving 75% mean IoU across 3 organ classes
• Obstruction Detection: Skeletonization-based diameter measurement to flag dilation risk
• Severity Stratification: 4-tier risk classification (normal, low risk, elevated, high risk)
• Visualization: Slice-level heatmaps and bowel diameter plots

Dataset

• 467 de-identified MRI scans from 107 patients
• ~144 axial slices per scan
• Pixel resolution: 1.5 x 1.5 x 3 mm
• 40,000+ total slices processed

Installation

git clone https://github.com/Rohith-Kumar-S/BowelSegmentation-MRI.git
cd BowelSegmentation-MRI
pip install -r requirements.txt

Requirements

torch
torchvision
transformers
numpy
tqdm
albumentations
matplotlib
ipywidgets
kagglehub
pandas
opencv-python
SimpleITK
antspyx
scikit-learn
scipy
scikit-image

Clinical Thresholds

Diameter thresholds based on peer-reviewed research for small bowel obstruction detection:

Threshold	Small Bowel	Large Bowel	Interpretation
Normal	≤1.7 cm	≤3.0 cm	Rule out obstruction
Low Risk	<2.75 cm	<6.0 cm	Monitor
Elevated	<4.0 cm	<9.0 cm	Possible obstruction
High Risk	≥4.0 cm	≥9.0 cm	Likely obstruction

Reference: https://pubmed.ncbi.nlm.nih.gov/39043061/ https://pubmed.ncbi.nlm.nih.gov/40357737/

Results

Segmentation Performance

Class	IoU
Background	99.76%
Small Bowel	55.08%
Large Bowel	68.93%
Stomach	76.79%
Mean	74.89%

Sample Visualizations

Diameter Profile Plot Bowel diameter across slices with clinical threshold lines, highlighting regions of concern. Following plot is from the results for a test patient from the dataset.

Severity Heatmap Risk levels across all slices for both small and large bowel.

Acknowledgments

• Dataset: [kaggle/happyharrycn/uw-madison-gi-tract-image-segmentation-dataset]
• Clinical thresholds: National Library of Medicine research