Back to Projects List
HOWTO: Detection of prostate cancer in IDC images using MONAI prostate_mri_anatomy model
- Cosmin Ciausu (Brigham and Women’s Hospital, USA)
- Deepa Krishnaswamy (Brigham and Women’s Hospital, USA)
- Patrick Remerscheid (Brigham and Women’s Hospital, USA and Technical University Munuch, Germany)
- Tina Kapur (Brigham and Women’s Hospital, USA)
- Sandy Wells (Brigham and Women’s Hospital, USA)
- Andrey Fedorov (Brigham and Women’s Hospital, USA)
[MONAI Zoo] has a growing number of pre-trained models for solving a range of image analysis tasks. It is of interest to understand robustness of those models on independent datasets, evaluate their performance.
NCI Imaging Data Commons (IDC) has a growing number of imaging datasets, most of which do not have accompanying annotations, complicating downstream analysis.
In this project we will demonstrate how an existing pre-trained MONAI model packaged as a bundle can be applied to a suitable subset of data from IDC, and how existing annotations can be used to validate results produced by this model.
- Develop an end-to-end documented example demonstrating the use of MONAI bundle on IDC prostae MRI.
- Understand and quantify the performance of the model using ground truth annotations.
- If applicable (results are of good quality), consider sharing the produced annotations within IDC.
Approach and Plan
Develop a Google Colab notebook that contains the following steps:
Install prerequisites and the MONAI bundle https://github.com/Project-MONAI/model-zoo/tree/dev/models/prostate_mri_anatomy.
Select applicable subset of MRI series from IDC (ProstateX and QIN-Prostate-Repeatability collections).
Convert images from DICOM to the format acceptable by the model.
Perform quantitative evaluation of the results.
Convert results into DICOM representation, visualize in OHIF.
Document performance of the model.
Consider sharing analysis results if they are of good quality.
Progress and Next Steps
- Preliminary work applying the model in question to segment prostate anatomy.
- Created bundle segmenting prostate tumors
- Minimum working example on training data sample
- Examination of results on pre-trained model training data : prostate158
- Multi-modal input : T2,ADC, DWI, understand acquisition process of DWI used for training
- Bundle creating thoughts : More extensive documentation about required parameters in inference.json and the relation between anatomy.json and inference.json should be provided.
- Document process of creating bundle, difficulties encountered
- Next steps : Confirm DSC results on prostate158 and evaluate on IDC data(DWI acquisition parameters – QIN Prostate repeatability similar to prostate158 ?)
Background and References