October 11, 2022 (Dev Meeting)

Agenda:

#	Owner	Title	Time (min)
1	Behrooz	WG operational items
2	Behrooz	CoNIC dataset license
3	Jonny	NuCLS dataset annotation quality
4	Behrooz	HoVerNet classification accuracy
5	David	Update on interactive labeling for MONAI Label + DSA
6	All	Discussion on active learning for histopathology patches
7	Markus	Presentation on dicomslide

Attendees

Behrooz Hashemian, Erich R. Bremer, Ziyue Xu, Jonny Hancox, Lee Alex Donald Cooper, David Manthey, Tahsin Kurc, Gregory Lee, Shan Raza, Michael Boone, Andres Diaz-Pinto, Gigon Bae, Markus D. Herrmann

Notes

• Please join Project MONAI slack and pathology-microscopy channel: https://docs.google.com/forms/d/e/1FAIpQLScOhMoUk6qbsBcbuBIlZJRoqP6idi5cSMOpw4Loo9qTsHxZcg/viewform
• CoNIC dataset license: Shan mentioned that they don't have a plan to release it on a more permissive license but it should be discussed with Nasir.
• NuCLS dataset annotation quality : Jonny explained the quality concerns about NuCLS. Lee mentioned that this dataset would be nice to be used with NuClick. Behrooz explained why using NuCLS might be a risky choice for the next release due to the time limitation and lack of point of reference for model performance on this dataset.
• HoVerNet classification accuracy: discussed the claim about low accuracy of HoVerNet classification but without knowing the training and testing datasets, it is not possible to evaluate it.
• Update on interactive labeling for MONAI Label + DSA: David has started a local branch, and experienced the time DSA took for container and communicate cost about a second. He is also working to improve this and looking into efficiency of json emitting and ingestion.
• Active learning: Discussed about using variance as a selecting criteria for patches to be annotated, where multiple inferences are run with dropouts to calculate the variance in the inference results. We agree that it is a good algorithm for the next release and anything beyond that can be more complicated. Lee's work on active learning with a team in Spain will need few more weeks of work to be published in arXiv.
• Conflicting patch annotations: discussed about how to resolve the discrepancy between patches for the same region (overlapping) or non-matching borders of patches. We discussed that there are many ways and no trivial way to handle them. David mentioned that DSA shows all the annotation and do not consolidate them, and it is better to be done on the data store side than the viewer. For MONAI hovernet pipeline (without the viewer), we are planning to average any inference results for the same physical location. Markus also mentioned that they have worked on this for overlapping patches to make a preference for the results of the patch where ROI is at the center to the results that lies at the border on another patch.
• Markus has presented their work on dicomslide as well as their related work on High Dicom.

Action items:

PIC	Item
Behrooz	Follow up with Nasir on the CoNIC dataset license
Shan	Check with Simon about HoVerNet classification accuracy
Lee	Talk to MSKCC people at PathVisions about their claim on HoVerNet accuracy