Liver Segmentation Using Monai and PyTorch

This repository provides all the necessary code and resources to perform liver segmentation using the Monai framework and PyTorch. The project is structured to allow customization for segmenting other organs as well.

---

Shoutout 🎉

A special shoutout to Mohammed El Amine Mokhtari for his great tutorials on medical imaging and computer vision. His work has been an inspiration and a valuable resource for the community! Youtube Tut

Table of Contents

• Project Overview
• Dataset
• Data Preparation
• Model Architecture
  • UNet
  • Adopt Optimizer
• Training and Logging
  • Mixed Precision Training
  • WandB Integration
• Installation and Usage
• References

---

Project Overview

This project focuses on automatic 3D liver segmentation from medical imaging data using a UNet architecture. The implementation is powered by the Monai framework, an open-source PyTorch-based library for deep learning in healthcare imaging.

Key features include:

• Support for mixed precision training.
• Advanced logging and experiment tracking with Weights and Biases (WandB).
• Incorporation of the Adopt optimizer and Adam optimizer for performance tuning.

---

Dataset

The datasets used for this project are sourced from Kaggle:

1. Liver Tumor Segmentation - Part 1
2. Liver Tumor Segmentation - Part 2

These datasets consist of liver tumor segmentation volumes stored in NIfTI format. For detailed information, please refer to the Kaggle links above.

---

Data Preparation

The data preparation process is automated and outlined in the Datapreparation.ipynb notebook. Key steps include:

1. Conversion of NIfTI files to DICOM format.
2. Grouping DICOM slices: Groups of 74 slices are created for consistency.
3. Reconversion to compressed NIfTI: For training efficiency.
4. Empty segmentation cleanup: Removal of empty segmentation masks and their corresponding volumes.

Ensure you execute the notebook before running the training scripts to prepare the data in the correct format.

---

Model Architecture

UNet

The UNet architecture is utilized for segmentation, featuring:

• Encoder-decoder structure with skip connections.
• Convolutional layers to extract and learn spatial features.
• Compatibility with Monai for medical image processing.

For detailed insights into the UNet design, visit the Monai documentation.

Adopt Optimizer

This project implements the Adopt optimizer, which combines the strengths of:

• Adaptive gradient clipping for robust optimization.
• Momentum-based updates to enhance convergence.

Resources for learning about Adopt:

• Adopt Paper
• Adopt GitHub Repository

---

Training and Logging

Mixed Precision Training

Mixed precision training leverages both float32 and float16 data types to accelerate training while reducing memory usage. It is implemented using PyTorch’s torch.cuda package:

• Forward and backward passes are performed in float16 where possible.
• Critical computations, such as loss scaling, are performed in float32 to maintain numerical stability.

Why Mixed Precision?

• Speeds up training by 1.5–2x.
• Reduces GPU memory usage, enabling larger batch sizes.

Implementation in Code

from torch.cuda import autocast, GradScaler

scaler = GradScaler()
with autocast(device=device):
    outputs = model(inputs)
    loss = loss_function(outputs.float(), label.float())
scaler.scale(loss).backward()
scaler.step(optimizer)
scaler.update()

WandB Integration

Weights and Biases (WandB) is used for experiment tracking, visualization, and hyperparameter tuning.

Setting Up WandB

1. Install WandB:

   pip install wandb

2. Login to WandB:

   wandb login

3. Initialize in your script:

   import wandb
   wandb.init(project="Liver-Segmentation")

Features Logged

• Training and validation losses.
• Dice metrics.
• Checkpoints for model performance.

Example WandB Dashboard

Adopt and mixed precision training vs Adam with no mixed precision

---

Installation and Usage

Prerequisites

Install the required packages:

pip install -r requirements.txt

Clone the Repository

git clone https://github.com/SharifEhab/Automatic-3D-Liver-Segmentation.git
cd Automatic-3D-Liver-Segmentation

Run Data Preparation

Prepare the dataset by running the Datapreparation.ipynb notebook.

Train the Model

Run the training script:

python train.py

---

References

1. Monai Documentation
2. Adopt Optimizer Paper
3. Adopt GitHub Repository
4. Liver Tumor Segmentation Dataset (Kaggle)

---