NA-MIC Project Weeks

Back to Projects List

Automatic classification of MR scan sequence type

Key Investigators

• Deepa Krishnaswamy (Brigham and Women's Hospital/Harvard Medical School, USA)
• Andrey Fedorov (Brigham and Women's Hospital/Harvard Medical School, USA)
• Joost van Griethuysen (The Netherlands Cancer Institute, The Netherlands)

Project Description

Knowing the type of MRI scan is an important data curation step. For instance clinicians and developers need to know if a scan is T1 weighted, T2 weighted, diffusion, etc in order to make a diagnosis or develop an AI model. This curation can take a long time to do manually, especially if the fields in DICOM data are missing or incorrect. Some tools have been developed already, mostly for brain image classification, and only a few are available for abdominal/prostate areas.

The past two project weeks, we’ve made some progress in developing tools for AI/ML classification of prostate MR scans. See our:

• PW 41 page
• PW 40 page
• paper here and code here.

In this project week, we will focus on creating a 3DSlicer module.

Objective

1. We will create a 3DSlicer module to perform the scan type classification on all series in a study.

Approach and Plan

1. We will first allow the user to pick a study from the DICOM database.
2. We will run inference using our pre-trained prostate model on all the series in the study.
3. We will modify the layout automatically.
4. If there is time, we will allow the user to choose a body part and appropriate model - there are models for brain MRI, and chest/abdominal MRI scan classification

Progress and Next Steps

1. We have created an CNN that uses both image+metadata information to classify a scan into T1w, T2w, diffusion and apparent diffusion coefficient maps.

Illustrations

Background and References

Current GitHub repo

PW 41 work

PW 40 work

Related work

- [dcm-classifier](https://github.com/BRAINSia/dcm-classifier)