Skip to content

Latest commit

 

History

History
76 lines (45 loc) · 3.26 KB

File metadata and controls

76 lines (45 loc) · 3.26 KB

Back to Projects List

MONAI Label lung and airway segmentation

Key Investigators

  • Rudolf Bumm, MD (KSGR)
  • Andres Diaz-Pinto (Nvidia)

Project Description

MONAI Label is a server-client system that facilitates interactive medical image annotation by using AI. It is an open-source and easy-to-install ecosystem that can run locally on a machine with single or multiple GPUs. Both server and client work on the same/different machine. It shares the same principles with MONAI.

The aim of the project is to set up, train and evaluate a lung and airway server model in MONAI Label

Objective

  • set up MONAI Label on a PC with moderate to high-end Nvidea GPU
  • load MONAI Label apps and datasets
  • use Lung CT Segmenter for rapid creation of detailed CT Lung labels in MONAI Label for
    • right lung
    • left lung
    • airways
  • do training with the server model 
  • evaluate the AI´s auto-segmentation performance

Approach and Plan

fine tune the MONAI Label server
provide links

How to set up a MONAI Label in Windows 11

Dataset

This is the dataset we have been using:

Decathlon lung dataset (Task06_lung) 63 cases with lung tumors http://medicaldecathlon.com/ 

It is available for download (8 GB) after installation of MONAI Label and running this command in a powershell or bash: (edited)

monailabel datasets --download --name Task06_Lung --output datasets 

Progress and Next Steps

Illustrations / Results

Fig 1: MONAI Label inference after providing 2 high quality samples and training (50 epochs): Not usable

Fig 2: Status after providing 5 more high-quality labels and  training 1000 epochs /  5 iterations (1 h with RTX 3070 Ti), "deepedit" model:  
ML is able to divide right and left lungs as well as airways, but resolution is low.   

Fig 3: Status after labelling 17 more datasets, training another 1000 epochs /  22 iterations (6 h with RTX 3070 Ti), "segmentation" model: 
Much better resolution.  

Fig 4: Autosegmentation after label correction, 500 epochs / 22 iterations training (1.5h RTX 3070 Ti):
Good result! 

Background and References

https://github.com/Project-MONAI/MONAILabel

https://github.com/rbumm/SlicerLungCTAnalyzer