. 2020 Jul 13;12(7):1884. doi: 10.3390/cancers12071884

Table 2.

Characteristics of the deep-learning models according to tumor classification, tumor microenvironment analysis, and prognosis prediction of colorectal cancers using pathologic image analysis.

Model No.	Feature	Task	Dataset	External Cross-Validation	Base Model	Performance	Author/Team	Year	Country	Ref.
1	Tumor Classification	6 classes (cancer subtypes): NL/ADC/MC/SC/PC/CCTA	717 patches	Not done	AlexNet	Accuracy—97.5%	Xu	2017	China	[38]
2		5 classes (polyp subtypes):	2074 patches 936 WSI	Not done	ResNet	Accuracy—93.0%	Korbar	2017	USA	[39]
2		NL/HP/SSP/TSA/TA/TVA-VA,	2074 patches 936 WSI	Not done	ResNet	Accuracy—93.0%	Korbar	2017	USA	[39]
3		3 classes: NL/AD/ADC	30 multispectral image patches	Not done	CNN	Accuracy—99.2%	Haj-Hassan	2017	France	[40]
4		2 classes: NL/Tumor	57 WSI (10,280 patches)	Not done	VGG	Accuracy—93.5%,	Yoon	2018	South Korea	[41]
						Sensitivity—95.1%
						Specificity—92.8%
5		3 classes: NL/AD/ADC	27 WSI (13,500 patches)	Not done	VGG16	Accuracy—96 %	Ponzio	2018	Italy	[42]
6		4 classes: NL/HP/AD/ADC	393 WSI	Not done	CNN	Accuracy—80%	Sena	2019	Italy	[43]
6		4 classes: NL/HP/AD/ADC	(12,565 patches)	Not done	CNN	Accuracy—80%	Sena	2019	Italy	[43]
7		3 classes: NL/AD/ADC	4036 WSI	Not done	CNN/RNN	AUCs—0.96 (ADC)	Iizuka	2020	Japan	[44]
7		3 classes: NL/AD/ADC	4036 WSI	Not done	CNN/RNN	0.99 (AD)	Iizuka	2020	Japan	[44]
8.		2 classes: NL/Tumor	94 WSI,	Done using 378 DACHS data	ResNet18	AUC > 0.99	Kather	2019	Germany	[45]
			370 TCGA-KR,
			(60,894 patches)
			378 TCGA-DX,
			(93,408 patches)
9	Tumor Microenvironment Analysis	Classification, Segmentation and Detection: EC/IC/FC/MC	21,135 patches	Not done	DCRN/R2U-Net	Classification	Alom	2018	USA	[46]
						F1-score—0.81
						AUC—0.96
						Accuracy—91.1%
						Segmentation
						Accuracy—92.1%
						Detection
						F1 score—0.831
10		Detection of immune cell CD3⁺, CD8⁺	28 WSI IHC	Not done	FCN/LSM/U-Net	FI score—0.80 Sensitivity—74.0% Precision—86	Swiderska-Chadaj	2019	Netherland	[47]
11		Detection and classification EC/IC/FC/MC	853 patches & 142 TCGA images	Not done	CNN	Detection Accuracy—65% Classification Accuracy—76 %	Shapcott	2019	UK	[48]
12		Classification of 9 cell types ADI, BAC, DEB, LYM, MUC, SM, NL, SC and EC	86 WSI (100,000) NCT&UMM	Not done	VGG19	Accuracy—94–99%	Kather	2019	Germany	[20]
13	Prognosis Prediction	5-year disease-specific survival	420 TMA	Not done	LSTM	AUC—0.69	Bychkov	2018	Finland	[49]
14		Survival predictions	25 DACHS WSI	Not done	VGG19	Accuracy—94–99%	Kather	2019	Germany	[20]
			862 TCGA WSI
			409 DACHS WSI
15		MSI predictions	360 TCGA- DX (93,408 patches) 378 TCGA- KR (60,894 patches)	Done using 378 DACHS data	ResNet18	AUC TCGA-DX—0.77 TCGA-KR—0.84	Kather	2019	Germany	[45]

NL, normal mucosa; ADC, adenocarcinoma; MC, mucinous carcinoma; SC, serrated carcinoma; PC, papillary carcinoma; CCTA, cribriform comedo-type adenocarcinoma; HP, hyperplastic polyp; SSP, sessile serrated polyp; TSA, traditional serrated adenoma; TA, tubular adenoma; TVA, tubulovillous adenoma; VA, villous adenoma; WSI, whole slide images; TCGA, The Cancer Genome Atlas; DACHs, Darmkrebs Chancen der Verhütung durch Screening; ResNet, residual network architecture; VGG, visual geometry group; AD, adenoma; CNN, convolution neural network; RNN, recurrent neural network; AUC, area under the curve; IHC, immunohistochemistry; TMA, tissue microarray; EC, epithelial cell; SC, stromal cells; DCNN, deep convolution neural network; MCC, Matthew correlation coefficient; MC, miscellaneous; DCRN, densely connected recurrent convolutional network; R2U-Net, recurrent residual U-Net; FCN, fully convolutional networks; IC, inflammatory cell; FC, fibroblast cell; ADI, adipose; BAC, background; DEB, debris; LYM, lymphocytes; MUC, mucus; SM, smooth muscle; LSM, locality-sensitive method; LSTM, long short-term memory; MSI, microsatellite instability.