Tasks for artificial intelligence in prostate MRI

European Radiology Experimental

Table 1 Summary of the artificial intelligence (AI) development papers discussed in detail

First author [reference number]	Overall sample size	AI family	AI method	Public/external datasets used	Images used	Loss functions	AUC	Dice similarity coefficient
Wang [11]	90	Whole gland segmentation	3D CNN + skip connections	PROMISE12	T2-weighted	Cross-entropy + cosine loss		0.86−0.88
Ushinsky [12]	299	Whole gland segmentation	Hybrid 2D-3D CNN + skip connections		T2-weighted	Adam loss		0.88
Sanford [13]	648	Whole gland segmentation	Hybrid 2D-3D CNN	Five separate unaffiliated institutional independent datasets	T2-weighted	Dice similarity coefficient loss		0.931
Cao et al. [14]	417	Lesion detection	3D CNN FocalNet		T2-weighted, ADC maps, echo-planar	Mutual finding loss	0.81
Ishioka [15]	335	Lesion detection	U-net + ResNet50 (skip connections)		T2-weighted	Adam loss	0.64–0.65
Le [16]	364	Lesion classification	Two parallel 2D CNNs	The Cancer Imaging Database (TCIA)	T2, ADC maps	Similarity loss	0.91
Liu et al. [17]	341	Lesion classification	3D CNN XmasNet	PROSTATE-x	T2, ADC, diffusion-weighted, K_trans	Adam loss	0.84

2D Two-dimensional, 3D Three-dimensional, ADC Apparent diffusion coefficient, AUC Area under the curve, CNN Convolutional neural network