NA-MIC Project Weeks
Back to Projects List
Automatic classification of MR scan sequence type
Key Investigators
- Deepa Krishnaswamy (BWH, USA)
- Cosmin Ciausu (BWH, USA)
- Megha Kalia (BWH, USA)
- Andrey Fedorov (BWH, USA)
Project Description
Data curation is a necessary step before using many AI or ML models, but it can be difficult and time-consuming to do manually. For instance in prostate cancer, most tools use multiple types of MR sequences as input to develop models and perform tasks such as segmentation.
In this project, we will develop methods for automatic classification of MR sequences. We had some great discussions and headway last project week, and are continuing this work.
We also made some progress since last project week and developed a few methods for classification of T2 axial, diffusion weighted (DWI), apparent diffusion coefficient (ADC) images, and dynamic contrast enhanced (DCE) images. We used combinations of image data and DICOM metadata as input, and developed a random forest classifier, and also two CNN-based classifiers – see our paper here and code here.
This project week, we’d like to talk to more people, address limitations of our work, and hopefully work on developing a more robust method for classification of scans.
Objective
- We would like to discuss the limitations of our previous work, and brainstorm new ideas for automatic classification of the MR series type.
- We would like to create an easy colab notebook for people to try out the methods
- We would like to think about developing a more robust method
Approach and Plan
- We will talk to people to discuss limitations of our method. For instance, what types of metadata should we use for the classification? Should we have a class for unknown scan type? Should we do a hierarchical classification method? How can we make the model agnostic to the area scanned?
Progress and Next Steps
- WIP Colab notebook - we download data from IDC, and run inference using the three pretrained models.
- WIP HuggingFace space demo - we want the user to choose which data to download from IDC, and then will choose a pre-trained model to run inference
Illustrations
No response