Skip to content

Latest commit

 

History

History
57 lines (37 loc) · 3.11 KB

File metadata and controls

57 lines (37 loc) · 3.11 KB

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

  1. Augment existing and applicable IDC collections with prostate segmentations obtained from nnUnet pre-trained models.
  2. Investigate/analyze nnUnet framework generalizability to IDC data.

Approach and Plan

  1. Use nnUnet models pre-trained on prostate data to get predictions results on labelled IDC collections first, PROSTATEx or qin-prostate-repeatability for example.
  2. Evaluate performance on labelled data, followed by inference on unlabelled IDC collections.

Progress and Next Steps

  1. Verification of nnUnet claimed results on prostate decathlon data, used for available pre-trained models.

  2. 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.
  3. 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
  4. Obtained good dice scores results on the 15 PatientID divided into two studies each using this model.

  5. 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.

Slicer demo

Background and References