Comparative performance of fully-automated and semi-automated artificial intelligence methods for the detection of clinically significant prostate cancer on MRI: a systematic review

Insights into Imaging

Table 3 Predictive modelling characteristics of studies using deep learning-based fully-automated AI methods

Study	No. of patients	Training set	Validation set	Test set	Algorithm	MRI input	Image registration	Image segmentation	Outcome	Zone	Analysis	Evaluation strategy
Wang [20]	346	204	fivefold CV	142	CNN (MISN)	ADC, BVAL, DWI₀, DWI₁, DWI₂, K^trans, T2WI-Cor, T2WI-Sag, T2WI-Tra	NR	Open data	csPCa vs iPCa or benign lesions	PZ or TZ	Per lesion	Internal hold-out
Fernandez-Quilez [21]	200	NR^a	NR^a	NR^a	CNN (VGG16)	T2WI, ADC	NR	Open data	csPCa vs iPCa or benign lesions	WP	Per lesion	Internal hold-out
Schelb [22]	312	250	No	62	CNN (U-Net)	T2WI, DWI	SimpleITK, non-rigid Bspline with Mattes mutual information criterion	Automated (U-Net)	csPCa vs iPCa or benign lesions	WP	Per lesion, per patient	Internal hold-out
Deniffel [23]	499	324	75	50^b	CNN (3D)	T2WI, ADC, DWI	Static, affine	Manual bounding boxes	csPCa vs iPCa or benign lesions	WP	Per patient	Internal hold-out
Seetharaman [24]	424	102	fivefold CV	322	CNN (SPCNet)	T2WI, ADC	Manual	Registration from pathology images	csPCa vs iPCa or benign lesions	WP	Per pixel, per lesion	Internal hold-out

ADC, apparent diffusion coefficient; CNN, convolutional neural networks; csPCa, clinically significant prostate cancer; CV, cross-validation; DWI, diffusion-weighted imaging; iPCa, indolent prostate cancer; MISN, multi-input selection network; MRI, magnetic resonance imaging; NR, not reported; PZ, peripheral zone; T2WI, T₂-weighted imaging; TZ, transition zone; WP, whole prostate
^aThe study included 200 patients and 299 lesions, of which 70% were used to train train, 20% to test, 10% to fine-tune the models
^bDescribes the calibration cohort