NA-MIC Project Weeks

Back to Projects List

Evaluation of AI methods for prostate cancer segmentation

Key Investigators

• Deepa Krishnaswamy (Brigham and Women's Hospital, USA)
• Patrick Remerscheid (, Switzerland)
• Cosmin Ciausu (Brigham and Women's Hospital, USA)
• Brianna Burton (3D Side SA, Belgium)
• Laura Levy (, Switzerland)
• Bálint Kovács (DKFZ, Germany)
• Umang Pandey (Universidad Carlos III de Madrid, Spain)
• Andrey Fedorov (Brigham and Women's Hospital, USA)

Presenter location: In-person

Project Description

When it comes to evaluating AI methods, it’s important to have reproducible code and methods. We are interested in evaluating state-of-the-art AI methods for prostate cancer segmentation on data in Imaging Data Commons. Additionally, we have a non-public BWH internal dataset that we would like to use for evaluation.

Objective

1. We first need to identify a set of publicly available AI methods that we can use for prostate cancer segmentation.
2. We then need to identify datasets in IDC that we can use for evaluation, preferably ones with expert delineated segmentations.
3. Then, we will run inference using those methods, convert our output to a standard format (hopefully DICOM SEG) and visualize in OHIF and Slicer.
4. We will make our code and results publicly available in GitHub.

Approach and Plan

1. We will do a literature/code repo search of the methods.
2. We will search for appropriate data in IDC using the portal/BigQuery.
3. We will create a sample set of data from multiple prostate imaging collections, including T2, DWI, ADC and ground truth segmentations.
4. We will run inference using the 4 methods on these sample data sets and visualize in Slicer.
5. If possible we will perform quantitative evaluation.

Progress and Next Steps

1. We have identified two major branches of methods we can use, baseline methods from the PICAI challenge and two methods using MONAI.

   • PICAI has two baseline methods we can run: supervised nnUNet, semi-supervised nnDetection
   • MONAI has two methods we can run: a MONAI bundle and a MONAI Deploy MAP

2. We have identified two datasets in IDC that can be used for evaluation:

   • QIN-Prostate-Repeatability
   • Prostate-MRI-US-Biopsy
3. We have started evaluation of the PICAI supervised nnUNet baseline model and the MONAI Deploy MAP on their training data.
4. We will take Cosmin’s work from PW38 on the MONAI bundle for prostate cancer segmentation. We’ll make sure it works, and evaluate it on publicly available datasets including IDC data (continuation of this notebook).
5. We have evaluated the two PICAI models on almost all of the 3 subsets of data.
6. We’ve run the MONAI bundle on a subset of the 3 collections.
7. We’ve also run the MONAI deploy MAP on a subset of the 3 collections.

Illustrations

** PICAI nnUNet supervised **

ProstateX: The ground truth lesion is in green, and the predicted lesion in red.

** PICAI nnDetection semi-supervised **

ProstateX: The ground truth lesion is in green, and the predicted bounding box in white.

QIN-Prostate-Repeatability: The ground truth lesion is in green, and the predicted bounding box in white.

** MONAI Deploy MAP **

Using the MONAI Deploy MAP pre-trained model for prostate and lesion segmentation on a patient from Prostate-MRI-US-Biopsy. The ground truth lesion segmentation is on the left, and the predicted prostate gland segmentation and lesion segmentations are on the right.

Scrolling through slices of same patient as above:

Patient from QIN-Prostate-Repeatability: The ground truth lesion segmentation is on the left, and the predicted prostate gland segmentation and lesion segmentations are on the right.

Patient from ProstateX: The ground truth lesion segmentation is on the left, and the predicted prostate gland segmentation and lesion segmentations are on the right.

** MONAI bundle **

Using the MONAI bundle and pretrained prostate158 model, on patients from ProstateX. The grountruth lesion is in green, and the predicted in red.

Background and References

Our github repo with notebooks (WIP): https://github.com/deepakri201/prostateSeg

This is a continuation of the work that Cosmin did at PW38: https://projectweek.na-mic.org/PW38_2023_GranCanaria/Projects/MONAI_IDC_PCa_detection/

IDC getting started tutorials: https://github.com/ImagingDataCommons/IDC-Tutorials/tree/master/notebooks/getting_started

No response