NA-MIC Project Weeks
Back to Projects List
nnUnet - Prostate segmentation on Imaging Data Commons(IDC) data
Key Investigators
- Cosmin Ciausu (Brigham and Women’s Hospital)
- Andrey Fedorov (Brigham and Women’s Hospital)
Project Description
Inference on prostate IDC data using nnUnet segmentation framework.
This framework provides an end to end segmentation pipeline, from pre-processing, data augmentation, hyper-parameter selection to post-processing, with variants of an Unet segmentation model.
The Imaging Data Commons platform provides both labelled and unlabelled prostate scans collections.
Objective
- Augment existing and applicable IDC collections with prostate segmentations obtained from nnUnet pre-trained models.
- Investigate/analyze nnUnet framework generalizability to IDC data.
Approach and Plan
- Use nnUnet models pre-trained on prostate data to get predictions results on labelled IDC collections first, PROSTATEx or qin-prostate-repeatability for example.
- Evaluate performance on labelled data, followed by inference on unlabelled IDC collections.
Progress and Next Steps
- Verification of nnUnet claimed results on prostate decathlon data, used for available pre-trained models.
-
Inference on Qin-Prostate-Repeatability collection using a pre-trained nnUnet model on task05 imaging decathlon data :
- 3d full-res model, T2 and ADC modalities, so there is a need to resample the input.
-
Inference on Qin-Prostate-Repeatability collection using a different pre-trained nnUnet model, task 24 promise.
- Easier to corner the resampling problem since this pre-trained model has only one input modality – T2
- Obtained good dice scores results on the 15 PatientID divided into two studies each using this model.
- Dealt with the Resampling/Converting issue - slice spacing incorrect – use of simpleITK instead of plastimatch
Illustrations
Slicer visualisation of ground truth and predicted whole prostate segmentation mask, on PatientID01. Red is ground truth and green is prediction from nnUnet.
