Measuring the bias of incorrect application of feature selection when using cross-validation in radiomics

Insights into Imaging

Table 1 Overview of the datasets

Dataset	N	d	Dimensionality (#Samples/#Features)	Outcome balance (%)	Modality	Tumor type	DOI
Carvalho2018 [30]	262	117	2.22	59	FDG-PET	NSCLC	https://doi.org/10.1371/journal.pone.0192859
Hosny2018A (HarvardRT) [31]	293	1004	0.29	54	CT	NSCLC	https://doi.org/10.1371/journal.pmed.1002711
Hosny2018B (Maastro) [31]	211	1004	0.21	28	CT	NSCLC	https://doi.org/10.1371/journal.pmed.1002711
Hosny2018C (Moffitt) [31]	183	1004	0.18	73	CT	NSCLC	https://doi.org/10.1371/journal.pmed.1002711
Ramella2018 [32]	91	242	0.37	55	CT	NSCLC	https://doi.org/10.1371/journal.pone.0207455
Toivonen2019 [33]	100	7105	0.01	60	MRI	Prostate Cancer	https://doi.org/10.1371/journal.pone.0217702
Keek2020 [34]	273	1322	0.21	40	CT	HNSCC	https://doi.org/10.1371/journal.pone.0232639
Li2020 [35]	51	396	0.13	63	MRI	Glioma	https://doi.org/10.1371/journal.pone.0227703
Park2020 [36]	768	940	0.82	24	US	Thyroid Cancer	https://doi.org/10.1371/journal.pone.0227315
Song2020 [37]	260	264	0.98	49	MR	Prostate Cancer	https://doi.org/10.1371/journal.pone.0237587

Overview of all radiomics datasets used. Only publicly available datasets were included to allow for easy reproducibility. N denotes the sample size, while d denotes the number of features (corresponding to the dimension of the data). The outcome balance measures the number of events in the outcome. DOI denotes the identifier of the publication corresponding to the dataset