notesum.ai

Published at December 9

Is Self-Supervision Enough? Benchmarking Foundation Models Against End-to-End Training for Mitotic Figure Classification

cs.CV

Released Date: December 9, 2024

Authors: Jonathan Ganz¹, Jonas Ammeling¹, Emely Rosbach¹, Ludwig Lausser¹, Christof A. Bertram², Katharina Breininger³, Marc Aubreville⁴

Aff.: ¹Technische Hochschule Ingolstadt; ²University of Veterinary Medicine, Vienna; ³Center for AI and Data Science, Julius-Maximilians-Universität Würzburg; ⁴Flensburg University of Applied Sciences

Arxiv: http://arxiv.org/pdf/2412.06365v1

Name	Pretraining Algorithm	Dataset Size		Model Type	Model Size
		WSI	Tiles
Prov-GigaPath [xu2024gigapath]	DINOv2	170K	1.3B	ViT-G	1.1B
H-optimus-0 [hoptimus0]	DINOv2	500K	-	ViT-G	1.1B
Virchow [vorontsov2024virchow]	DINOv2	1.5M	2B	ViT-H	632M
Virchow2 [zimmermann2024virchow2]	DINOv2	3.4M	1.7B	ViT-H	632M
UNI [chen2024uni]	DINOv2	100K	100M	ViT-L	307M
Phikon [filiot2023phikon]	iBOT	6K	43M	ViT-B	86M