notesum.ai

Published at November 26

Scaling nnU-Net for CBCT Segmentation

cs.CV

Released Date: November 26, 2024

Authors: Fabian Isensee¹, Yannick Kirchhoff², Lars Kraemer¹, Maximilian Rokuss², Constantin Ulrich², Klaus H. Maier-Hein²

Aff.: ¹Helmholtz Imaging, German Cancer Research Center (DKFZ), Heidelberg, Germany; ²Division of Medical Image Computing, German Cancer Research Center (DKFZ), Heidelberg, Germany

Arxiv: http://arxiv.org/abs/2411.17213v1

Model	F-cases (n=63)				All cases (n=480)
Model	Dice	HD95	Dice (PP)	HD95 (PP)	Dice	HD95	Dice (PP)	HD95 (PP)
nnU-Net ResEncL default	0.7445	24.99	0.7808	20.23	0.8211	22.135	0.8948	13.682
$+$ no l/r mirr	0.9083	9.758	0.9302	6.212	0.8610	17.417	0.9210	11.091
$+$ no l/r mirr, 1500ep	0.9064	10.354	0.9291	6.262	0.8624	17.392	0.9212	11.038
$+$ no mirr, 1500ep, ps160x320x320	0.8938	12.62	0.9255	7.266	0.8622	17.531	0.9223	11.239
$+$ no l/r mirr, 1500ep, ps160x320x320, pretr	0.9119	9.258	0.9320	5.875	0.8691	16.454	0.9242	11.001
$+$ no l/r mirr, 1500ep, ps160x320x320	0.9098	10.14	0.9344	6.069	0.8686	16.726	0.9271	10.605