notesum.ai

Published at October 18

Less is More: Selective Reduction of CT Data for Self-Supervised Pre-Training of Deep Learning Models with Contrastive Learning Improves Downstream Classification Performance

cs.CL

cs.AI

cs.LG

Released Date: October 18, 2024

Authors: Daniel Wolf¹, Tristan Payer¹, Catharina Silvia Lisson², Christoph Gerhard Lisson², Meinrad Beer², Michael Götz, Timo Ropinski¹

Aff.: ¹Visual Computing Research Group, Institute of Media Informatics, Ulm University; ²Experimental Radiology Research Group, Department for Diagnostic and Interventional Radiology, Ulm University Medical Center

Arxiv: https://arxiv.org/abs/2410.14524v1

Pre-Training		Downstream Results
Dataset	Method	COVID-19		OrgMNIST		Brain
		AUC	F1	AUC	F1	AUC	F1
PET-CT	ALL	0.775 $\pm$ 0.009	0.719 $\pm$ 0.010	0.968 $\pm$ 0.003	0.752 $\pm$ 0.003	0.727 $\pm$ 0.042	0.534 $\pm$ 0.073
	EveryN 20 %	0.801 $\pm$ 0.006	0.735 $\pm$ 0.009	0.972 $\pm$ 0.003	0.782 $\pm$ 0.003	0.781 $\pm$ 0.035	0.665 $\pm$ 0.070
	EveryN 10 %	0.810 $\pm$ 0.007	0.740 $\pm$ 0.016	0.973 $\pm$ 0.002	0.793 $\pm$ 0.002	0.798 $\pm$ 0.031	0.674 $\pm$ 0.074
LIDC	ALL	0.807 $\pm$ 0.006	0.744 $\pm$ 0.013	0.972 $\pm$ 0.003	0.769 $\pm$ 0.003	0.734 $\pm$ 0.046	0.609 $\pm$ 0.072
	EveryN 20 %	0.810 $\pm$ 0.004	0.751 $\pm$ 0.010	0.977 $\pm$ 0.005	0.792 $\pm$ 0.003	0.739 $\pm$ 0.044	0.610 $\pm$ 0.046
	EveryN 10 %	0.812 $\pm$ 0.006	0.756 $\pm$ 0.010	0.979 $\pm$ 0.002	0.800 $\pm$ 0.003	0.740 $\pm$ 0.041	0.614 $\pm$ 0.046