Table 1.

Studiesincluded in the systematic review, divided into the four previously defined categories, along with their main parameters. Mag.: magnification level. n/a: not available. #: Number. xAI: studies that provide elements for explainable AI, e.g., GradCAMs or attention mechanism.

	Study	Year	Studied Structures	Mag.	# WSIs	Patch Size	Pre-Processing	DL Method	GPU Used	# Sources	Metadata	xAI
Comparison vs. pathologists	Ba et al. [42]	2021	Tumor	40×	781	256 × 256	Image quality review	CNN and random forest	n/a	2	no	yes
	Bao et al. [36]	2022	Tumor	40×	981	224 × 224	Random patch selection, structure-preserving color normalization	ResNet-152	NVIDIA GTX 2080Ti	3	no	no
	Brinker et al. [43]	2022	Tumor	n/a	100	n/a	n/a	ResNeXt50	n/a	n/a	no	yes
	Hekler et al. [14,15]	2019	Tumor	10×	695	n/a	n/a	ResNet50	n/a	1	no	no
	Phillips et al. [27]	2019	Tumor, dermis, and epidermis	40×	50	512 × 512	Subtraction	Modified FCN	NVIDIA GTX 1080 Ti	10 †	no	yes
	Sturm et al. [16]	2022	Mitosis	20×	102	n/a	n/a	n/a	n/a	1	yes	no
	Wang et al. [37]	2020	Tumor	20×	155	256 × 256	Random cropping to 224 × 224, data enhancement, and augmentation	VGG16	n/a	2	no	yes
	Xie et al. [28]	2021	Tumor, dermis, and epidermis	20×	701	224 × 224	Discard blank patches (Otsu)	ResNet50	n/a	3 †	yes	no
	Xie et al. [17]	2021	Tumor	n/a	841	256 × 256	Discard blank patches (Otsu)	ResNet50	NVIDIA TITAN RTX	1	no	yes
Diagnosis	Del Amor et al. [19]	2021	Tumor	10×	51	512 × 512	Discard blank patches (Otsu)	VGG16 with attention	NVIDIA DGX A100	1	no	yes
	Del Amor et al. [18]	2022	Tumor	5×, 10×, 20×	43	512 × 512	Discard blank patches and with less than 20% of tissue (Otsu)	ResNet18 with late fusion of multiresolution feature maps	NVIDIA GP102 TITAN Xp	1	no	yes
	Hart el al. [34]	2019	Tumor	40×	300	299 × 299	n/a	InceptionV3	4 NVIDIA GeForce GTX 1080	n/a	no	yes
	Höhn et al. [38]	2021	Tumor	n/a	431	512 × 512	Remove patches with more than 50% of background, random selection of 100 tiles per slide	ResNeXt50 with fusion model to combine patient data and image features	NVIDIA GeForce GTX 745	2	yes	yes
	Li et al. [29]	2021	Tumor, dermis, and epidermis	20×	701	224 × 224	Discard blank patches (Otsu)	ResNet50	n/a	2 †	yes	yes
	Van Zon et al. [20]	2020	Tumor	40×	563	256 × 256	Data augmentation	U-Net	NVIDIA 2080	1	no	no
	Xie et al. [21]	2021	Tumor	40×	312	500 × 500	Filter out background tiles	Transfer learning vs fully trained: InceptionV3, ResNet50, MobileNet	n/a	1	no	no
Prognosis	Brinker et al. [13]	2021	Tumor	n/a	415	256 × 256	n/a	ResNeXt50	n/a	3	yes	no
	Kim et al. [30]	2022	Tumor, inflammatory cells, and other	20×	305	299 × 299	n/a	Inception v3 with fivefold cross-validation	n/a	2 †	yes	no
	Kulkarni et al. [40]	2020	Tumor, inflammatory cells, and other	40×	n/a	500 × 500	Downsample to 100 × 100, nuclear segmentation with watershed cell detection	n/a	n/a	2	yes	no
	Moore et al. [41]	2021	Tumor, inflammatory cells, and other	40×, 20×	n/a	100 × 100	n/a	QuIP TIL CNN [44]	NVIDIA GP102GL [Quadro P6000]	2	yes	no
	Zormpas-Petridis et al. [31]	2019	Tumor, inflammatory cells, and other	20×, 5×, 1.25×	105	2000 × 2000 (20× WSIs)	n/a	Spatially constrained CNN with spatial regression, neighboring ensemble with softmax	NVIDIA Tesla P100-PCIE-16GB	1 †	yes	no
ROI/histological features	Alheejawi et al. [22]	2021	Tumor, inflammatory cells, and epidermis	40×	4	960 × 960	Divide patches into 64 × 64 blocks	ResNet50	NVIDIA GeForce GTX 745	1	no	no
	De Logu et al. [39]	2020	Tumor and healthy tissues	20×	100	299 × 299	Data augmentation, discard patches with more than 50% background	Inception-ResNet-v2	n/a	3	no	yes
	Kucharski et al. [23]	2020	Tumor	10×	70	128 × 128	Data augmentation, overlapping only for minority class to balance data set	Autoencoders	n/a	1	no	yes
	Liu et al. [24]	2021	Tumor	10×	227 ROIs ‡	1000 × 1000	Downscale magnification to 5×	Mask R-CNN	4 NVIDIA GeForce GTX 1080	1	no	no
	Nofallah et al. [25]	2021	Mitosis	40×	22	101 × 101	Data augmentation	ESPNet, DenseNet, ResNet, and ShuffleNet	NVIDIA GeForce GTX 1080	1	no	no
	Zhang et al. [26]	2021	Tumor	n/a	30	1024 × 1024	Data augmentation, color analysis for tissue-contained patch selection, normalization of patches to a uniform size, resize patches to 512 × 512	CNN, feature fusion	NVIDIA RTX 2080-12G	1	no	no

† At least one of the source institutions is open source, i.e., TCGA or NCI. ‡ Images are ROIs extracted from initial WSIs.