notesum.ai

Published at October 30

Revisiting MAE pre-training for 3D medical image segmentation

cs.CV

cs.AI

cs.LG

Released Date: October 30, 2024

Authors: Tassilo Wald¹, Constantin Ulrich¹, Stanislav Lukyanenko², Andrei Goncharov², Alberto Paderno³, Leander Maerkisch², Paul F. Jäger, Klaus Maier-Hein¹

Aff.: ¹Division of Medical Image Computing, German Cancer Research Center (DKFZ), Heidelberg, Germany; ²FLOY, Munich, Germany; ³Department of Biomedical Sciences, Humanitas University, IRCCS Humanitas Research Hospital, Rozzano, Italy

Arxiv: http://arxiv.org/abs/2410.23132v1

Pretraining	D1	D2	D3	D4	D5	Avg. D1-D5
No Dyn.	45.56	72.26	88.80	60.44	82.61	69.93
No Fixed	49.37	69.13	88.78	60.74	81.33	69.87
VoCo	50.35	67.20	88.22	57.82	80.29	68.77
VF	49.93	69.58	88.83	61.75	81.48	70.31
MG	50.50	71.14	88.83	63.29	82.15	71.18
S3D-B (ours)	51.49	74.01	88.83	62.39	81.54	71.65
S3D-L (ours)	51.42	72.84	89.09	63.30	82.15	71.76