MONAI Training Pipeline

This repository contains a complete pipeline for training a UNet model for liver segmentation using the MONAI Task03_Liver dataset, orchestrated with Valohai.

Overview

This project demonstrates how to:

1. Download and preprocess the Task03_Liver dataset
2. Train a UNet model using MONAI
3. Evaluate model performance using Dice Score
4. Run inference on unseen data

Project Structure

• preprocess.py: Downloads and preprocesses the Task03_Liver data
• train.py: Trains the UNet model using MONAI
• evaluate.py: Evaluates model performance on the test set
• inference.py: Performs segmentation inference on new images
• valohai.yaml: Defines the Valohai pipeline and execution steps
• requirements.txt: Core Python dependencies

Pipeline Steps

This pipeline automates the full workflow for medical image segmentation using a U-Net architecture on the Task03_Liver dataset from the Medical Segmentation Decathlon.

1. Preprocess Dataset

• Downloads the Task03_Liver dataset from the Medical Segmentation Decathlon.
• Applies preprocessing transforms such as resampling, cropping, and resizing.
• Splits the dataset into training, validation, and test sets, and generates manifest files.
• Saves the preprocessed volumes and labels as zip using Datasets (check: https://docs.valohai.com/hc/en-us/articles/18704302494481-Creating-datasets)

2. Train Model

• Trains a configurable U-Net model on the preprocessed dataset.

• Supports custom settings for:

  • Number of epochs
  • Learning rate
  • Batch size
  • Input/output channels
  • Number of residual units
  • Channel depth at each stage

3. Evaluate Model

• Evaluates model performance using metrics such as:

  • Dice Similarity Coefficient (DSC)
  • Intersection over Union (IoU)

• Produces detailed logs and segmentation quality visualizations.

4. Run Inference

• Performs sliding window inference on new or unseen volumetric data.

• Input:

  • Raw volume in NIfTI format (.nii.gz) Available test case: https://drive.google.com/drive/u/0/folders/1auCFCixQecS3rcBSbOf-KQXI1-tgDObx

• Outputs:

  • Prediction masks in NIfTI format (.nii.gz)
  • Visualization snapshots for qualitative analysis
  • Example of visualized output

Run the pipeline

vh pipeline run train_and_evaluate

Dataset

This project uses the Task03_Liver dataset from the Medical Segmentation Decathlon. The dataset contains CT volumes of the liver with manual segmentation masks.

Model

The training pipeline uses the UNet architecture implemented in MONAI, optimized for medical image segmentation tasks.

License

This project leverages the MONAI framework and uses datasets from the Medical Segmentation Decathlon. It is distributed under the Apache License 2.0.

Acknowledgments

• Project MONAI
• Valohai for orchestrating the training workflow