-
Notifications
You must be signed in to change notification settings - Fork 1.1k
July 19, 2022 (Dev Meeting)
Behrooz Hashemian edited this page Jul 19, 2022
·
9 revisions
| # | Owner | Title | Time (min) |
|---|---|---|---|
| 1 | All | New people introduction | 30 |
| 2 | Behrooz | GitHub Project for MONAI Pathology | 15 |
| 3 | All | MetaTensor for digital pathology | 15 |
Behrooz Hashemian, Lee Alex Donald Cooper, Ziyue Xu, Michael Boone, David Manthey, Gigon Bae, Shan Raza , Jeff Baumes, Marcos Novaes, Jason Klotzer, Gregory Lee
- Marcos and Jason explained their work in GCP for DICOM in radiology and how they are supporting viewer functionality around OHIF with DICOM store. They like to take the same approach for digital pathology, and interested in MONAI Pathology with DICOM support. They are looking into Slim as the viewer on their platform since it has a native dicom support: https://github.com/herrmannlab/slim
- The main challenge for digital pathology would be annotation and a subgroup of DICOM WG-16 has worked to standardize them for pathology and with some interoperabilities with radiology: https://www.dicomstandard.org/activity/wgs/wg-26
- Lee was concerned about the scalability of annotations since there can be millions of annotations in one slide.
- Google Cloud Platform WSI to DICOM Converter: https://github.com/GoogleCloudPlatform/wsi-to-dicom-converter
- Required metadata:
- magnification: most important property to include in MetaTensor. Although magnification levels are commonly used, they are not standardized and may vary between different vendors.
- level: the information can be derived from magnification but from practical point of view, it might make sense to include
- location: required for patches
- Resampling in digital pathology is not commonly needed but there are use cases that even working with the same magnification level, their physical units do not match and need to bring them into the same physical unit with resampling.
- In digital pathology (unlike fluorescence microscopy) usually the transformation are 2D. There use cases dealing with the serial section that requires a 3D representation (adding layers as the third dimension) but they are not very common and maybe less than 5% according to Shan.
- Also discussed about moving cucim object along the extracted image in AI pipelines but this requires the cucim object be pickleable: https://github.com/rapidsai/cucim/issues/326
| PIC | Item |
|---|---|