NeoLUNet

Project for Neonatal MRI Lung Segmentation with U-Nets

Contents

• Introduction
• Overview
• Content
• Usage
• Reference

Introduction

Quantification of lung structural information is crucial to standardize and inform diagnostic processes, enable personalized treatment and monitoring strategies. MRI can provide quantitative information for assessment of the neonatal lung while avoiding radiation exposure.

Overview

We developed an ensemble of deep convolutional neural networks to automatically and with high consistency perform neonatal lung segmentation from MRI sequences. A 3D representation of the lung and a clustering method to separate left and right lobes were implemented to calculate volumetric features.

In addition, ML Models for severity prediction of Bronchopulmonary Dysplasia (BPD) are implemented as an applied example of the use of MRI lung volumetric features for disease prognosis.

Content

1. MRI Lung Segmentation

── 1.1 Source

────── Main Script to Process Single Sequence

────── Functions for Lung Segmentation

────── Functions for Volumetric Feature Extraction

────── 1.1.1 DL Utils

──────────── 2D UNet Model

──────────── Segmentation Metrics

────── 1.1.2 Experiments - Lung Segmentation for Paper

────── Script to Process Paper Dataset

────── Script to Evaluate Paper Dataset

2. BPD Severity Prediction

── 2.1 Experiments for Nested Cross-Validation

3. Paper Resources

── Statistical Analysis and Figures for Paper

Usage

Read Use Instructions instructions

• Clone the repository

git clone '[email protected]:SchubertLab/NeoLUNet.git'

• Install dependencies

conda env create -f neolunet_env.yml

Data and Models

Download the pre-trained 2D-UNet MRI neonatal lung segmentation models here:

Zenodo Link

Segmentation performances and features per MRI-sequence also available in the link.

Reference

Radiology Artificial Intelligence (2023): Link