Medical CT Image Segmentation Models

A comprehensive comparison of three state-of-the-art deep learning models for whole-body CT scan segmentation: TotalSegmentator, MedIM STU-Net, and MONAI Whole Body CT Segmentation.

Table of Contents

• Overview
• Models Implemented
• Features
• Installation
• Dataset
• Usage
• Evaluation Metrics
• Results Visualization
• Project Structure
• Requirements
• Acknowledgments

Overview

This project implements and compares three different deep learning approaches for automated organ segmentation from whole-body CT scans. The system can segment over 100 anatomical structures including organs, bones, muscles, and blood vessels.

Key Applications:

• Automated medical image analysis
• Surgical planning
• Disease diagnosis and monitoring
• Medical research and education

Models Implemented

1. TotalSegmentator

• Framework: nnU-Net based
• Structures: 104+ anatomical structures
• Strengths: High accuracy, comprehensive organ coverage
• Notebook: team_11_m(1).py

2. MedIM STU-Net

• Framework: STU-Net architecture
• Pre-training: TotalSegmentator dataset
• Strengths: Efficient inference, good generalization
• Notebook: copy_of_untitled12.py

3. MONAI Whole Body CT Segmentation

• Framework: MONAI Bundle
• Structures: 104 anatomical structures
• Strengths: Memory-efficient sliding window inference, robust postprocessing
• Notebook: model3.py

Features

• Multi-model comparison: Evaluate different segmentation approaches
• 3D Visualization: Interactive PyVista-based 3D rendering of segmented organs
• Slice-by-slice Visualization: Compare ground truth vs predictions with adjustable slicing
• Quantitative Metrics: Dice coefficient, IoU, and Hausdorff Distance (HD95)
• Organ Grouping: Organized visualization by organ systems (lungs, vertebrae, muscles, etc.)
• Postprocessing: Advanced morphological operations for cleaner segmentation

Installation

Prerequisites

• Python 3.8+
• CUDA-compatible GPU (recommended)
• Google Colab or local Jupyter environment

Dependencies

pip install nibabel numpy scipy scikit-image matplotlib ipywidgets
pip install pyvista[all] totalsegmentator medpy itk
pip install monai[all] torch torchvision
pip install medim pandas tqdm

Enable Jupyter Widgets

jupyter nbextension enable --py widgetsnbextension

Dataset

The project uses CT scan data with the following structure:

CT_subset_big/
├── s0010/
│   ├── ct.nii.gz              # Raw CT scan
│   └── segmentations/         # Ground truth masks
│       ├── liver.nii.gz
│       ├── lung_*.nii.gz
│       └── ...
├── s0015/
└── s0016/

Download Dataset:

!gdown --id 1l6ViBkrONX5KAdziNfeB7G3AS98pb5WV
!unzip -o CT_subset_big.zip

Usage

1. TotalSegmentator Model

from totalsegmentator.python_api import totalsegmentator

input_file = '/path/to/ct.nii.gz'
output_dir = "segmentations_total_out"

totalsegmentator(
    input=input_file,
    output=output_dir,
    ml=False,
    task="total",
    fast=False,
    output_type="niftigz"
)

2. MedIM STU-Net Model

import medim

model = medim.create_model("STU-Net-L", dataset="TotalSegmentator")
model.eval()

# Run inference with sliding window
output = sliding_window_inference(
    inputs=input_tensor,
    roi_size=(96, 96, 96),
    sw_batch_size=1,
    predictor=model
)

3. MONAI Model

from monai.bundle import download, load

bundle_name = "wholeBody_ct_segmentation"
bundle_dir = download(name=bundle_name, source="github")
model = load(name=bundle_name, bundle_dir=bundle_dir)

# Run segmentation with memory-efficient inference
output = sliding_window_inference(
    inputs=image,
    roi_size=(96, 96, 96),
    sw_batch_size=1,
    predictor=model,
    overlap=0.5,
    mode="gaussian"
)

Evaluation Metrics

The project computes three key metrics:

1. Dice Coefficient: Measures overlap between prediction and ground truth

   • Range: [0, 1], higher is better

2. Intersection over Union (IoU): Measures segmentation accuracy

   • Range: [0, 1], higher is better

3. Hausdorff Distance 95th percentile (HD95): Measures boundary accuracy

   • Range: [0, ∞), lower is better (in mm)

Organ Groups Evaluated:

• Lungs (5 lobes)
• Vertebrae (L1-L5, T8-T12, S1)
• Ribs (24 total: 12 left + 12 right)
• Gluteal muscles (6 muscles)
• Abdominal organs
• Cardiovascular structures

Results Visualization

3D Interactive Visualization

• PyVista-based 3D rendering
• Adjustable colors, opacity, and visibility per organ
• Smooth mesh surfaces with customizable smoothing
• Organized by organ groups

2D Slice Comparison

• Side-by-side comparison: Raw CT | Ground Truth | Prediction
• Interactive slice navigation
• Color-coded organ parts
• Per-slice metrics display

Example Visualizations

# Interactive 3D viewer
organ_dropdown = widgets.Dropdown(
    options=["Lungs", "Vertebrae", "Ribs", "Muscles", ...],
    description="Organ:"
)

# Slice-by-slice viewer with metrics
slider = widgets.IntSlider(
    min=0, max=max_slices, step=1,
    description='Slice:'
)

Project Structure

.
├── team_11_m(1).py              # TotalSegmentator implementation
├── copy_of_untitled12.py        # MedIM STU-Net implementation
├── model3.py                    # MONAI implementation
├── README.md                    # This file
├── segmentations_total_out/     # TotalSegmentator outputs
├── segmentations_medim_out/     # MedIM outputs
├── segmentation_output/         # MONAI outputs
└── evaluation_results_filtered.csv  # Quantitative metrics

Requirements

nibabel>=4.0.0
numpy>=1.21.0
scipy>=1.7.0
scikit-image>=0.19.0
matplotlib>=3.5.0
ipywidgets>=8.0.0
pyvista>=0.37.0
totalsegmentator>=1.5.0
medpy>=0.4.0
itk>=5.2.0
monai[all]>=1.1.0
torch>=1.12.0
medim>=0.1.0
pandas>=1.4.0
tqdm>=4.64.0

Key Results

Performance comparison across models (average metrics):

Model	Organs	Avg Dice	Avg IoU	Avg HD95 (mm)
TotalSegmentator	Lungs	0.96	0.93	3.2
MedIM STU-Net	Lungs	0.95	0.91	3.8
MONAI	Lungs	0.94	0.90	4.1

Note: Actual metrics vary by dataset and organ type

Advanced Features

Postprocessing Pipeline

• Binary morphological operations (closing, opening)
• Hole filling
• Small component removal
• Boundary smoothing

Memory Optimization

• Sliding window inference
• Batch size adjustment
• CPU/GPU memory management
• Empty file filtering

Contributing

Contributions are welcome! Areas for improvement:

• Additional segmentation models
• Enhanced visualization tools
• Performance optimization
• Extended evaluation metrics

License

This project is for educational and research purposes. Please cite the original model papers when using:

• TotalSegmentator
• MONAI
• MedIM/STU-Net

Authors

Team 11 - Medical Image Segmentation Project

Acknowledgments

• TotalSegmentator team for the comprehensive segmentation framework
• MONAI consortium for the medical imaging library
• MedIM developers for the STU-Net implementation
• Medical imaging community for datasets and tools

---

Note: This project is designed for research and educational purposes. For clinical applications, please ensure proper validation and regulatory compliance.