Table 4.

Summary of selected works on artificial intelligence from the literature on bladder cancer.

#	Application	Reference	Dataset	Methods	Conclusion
1	Segmentation	Dolz et al. (2018) [123]	60 patients (training 40, validation 5, test 15)	U-Net yields precise segmentation of bladder walls and tumors on T2w.	Higher accuracy than standard CNN, especially for tumors.
2		Li et al. (2020) [122]	1092 MR images	U-Net with priors is applied to segment bladder walls and tumors on T2w.	The method improved the accuracy of bladder wall segmentation.
3		Yu et al. (2022) [126]	245 patients (training 220, test 25)	Path augmentation U-Net segmentation for bladder walls and tumors on T2w.	It can precisely extract bladder structures, especially small tumors.
4		Coroamă et al. (2023) [127]	33 patients	A low-complexity 3D U-Net with less than five layers for segmentation of bladder walls and tumors on T2w.	System for automated diagnosis of bladder tumors that can lead to higher reporting accuracy.
5		Moribata et al. (2023) [128]	170 patients (training 140, test 30)	U-Net could segment bladder cancer, and robust high-order radiomics features were extracted from ADC maps.	The model performed accurate segmentation of bladder cancer, and the extracted radiomics exhibited high reproducibility.
6	Classification	Zou et al. (2022) [130]	468 patients	Inception V3, CNN on T2w, recognizes the position of bladder walls and tumors.	Reliable method that can be more focused on features from the surrounding area of the tumor.
7		Sevcenco et al. (2018) [131]	51 patients (training 36, test 15)	A multilayer perceptron with one hidden layer on ADC maps.	Classifier model combining the ADC values with clinical–pathological information can identify patients at high risk for survival.
8		Li et al. (2023) [133]	Multicenter cohort of 89 (121) patients (tumors), 61 (93) from center 1, and 28 (28) from center 2. Tumors for training 93, test 28	3D ResNet50 CNN on T2w as a multitask model exhibits good diagnostic performance in predicting MIBC.	The method was lesion-focused and more reliable for clinical decisions.
9	Denoising	Taguchi et al. (2021) [124]	68 patients	VI-RADS validation CNN, with denoising reconstruction on T2w, discriminates between NMIBC and MIBC.	Combining VI-RADS with denoising CNN might improve diagnostic accuracy.
10		Watanabe et al. (2022) [125]	163 patients	VI-RADS validation CNN with denoising reconstruction on T2w and DW- predicts accurate MIBC without using DCE-MRI.	It achieved a comparable predictive accuracy for MIBC to that of conventional VI-RADS.