Table 1.

Artificial intelligence in diagnosis of colorectal cancer

Type of study	Ref.	No. of participants	Method	Control and interventions	Conclusion
Case control study	Yang et al[19], 2019	241	Depth-learning intelligent assistant diagnosis system	By comparing the accuracy of different algorithms on MRI images of patients with CRC, the algorithms that were conducive to the diagnosis of CRC were defined	T2-weighted imaging method had obvious advantages over other methods in differentiating CRC
Analytical research	Liu et al[20], 2011	429	SVM	Compared the performance of new and old classification methods in colorectal polyps CAD system	SVM could help CAD system get excellent classification performance
Review	Regge et al[21], 2013	NA	CAD system	NA	CAD system helped radiologists diagnose CRC with visual markers
Case control study	Summers et al[22], 2008	104	CAD system	The sensitivity of adenoma was measured by CAD system and compared with previous studies	CAD system had high accuracy in detecting and distinguishing adenoma
Descriptive research	Chowdhury et al[23], 2008	53	CAD-CTC system	The sensitivity of CAD-CTC system and manual CTC was compared through the image data of 53 patients	CAD-CTC system could effectively identify polyps and cancers with clinical significance in CT images
Case control study	Nappi et al[24], 2018	196	ResNets	Based on the clinical data of 196 patients, the classification performance of different models in distinguishing masses from normal colonic anatomy was compared	ResNets solved the practical problem of how to optimize the performance of DL
Case control study	Taylor et al[25], 2008	24	CAD system	The effectiveness of CAD system in detecting tumors was tested using the clinical data of 24 patients	CAD could effectively detect flat carcinoma by tumor morphology
Case control study	Summers et al[26], 2010	394	CAD-CTC system	The CTC data sets of 394 patients were trained in CAD system. It was confirmed that the experimental group could reduce the missed diagnosis rate of cancer	CAD-CTC system used advanced image processing and ML to reduce the occurrence of FP results
Case control study	Lee et al[27], 2011	65	CAD system	The CTC data sets of patient polyps were divided into a training data set and a test data set to compare the detection performance of CAD system	CAD system included colon wall segmentation, polyp specific volume filter, cluster size counting and thresholding, which had high detection performance of polyps and cancer tissue
Case control study	Nappi et al[28], 2015	154	DCNN	The clinical data were divided into a training data set and a test data set to compare the polyp detection performance of multiple classifiers	DCNN could greatly improve the accuracy of automatic detection of polyps in CTC
Case control study	Näppi et al[29], 2005	14	CAD system	The clinical data of 14 patients were used to test the effect of different staining methods on the effectiveness of polyp detection	CAD system helped to improve the ability to detect polyps in CTC
Case control study	van Wijk et al[30], 2010	84	CAD-CTC system	The polyp detection performance of different classification methods was tested through the clinical data of 84 patients	The sensitivity of the CAD-CTC system to distinguish polyps over 6 mm was very high
Case control study	Kim et al[31], 2007	35	CAD system	The sensitivity of CAD polyp detection was tested using colonoscopy data of 35 patients	CAD system helped to distinguish polyps and cancer tissue larger than or equal to 6 mm
Case control study	Nappi et al[32], 2017	101	CADe system	The polyp detection accuracy of novel and old CADe systems was compared by colonoscopy data of 101 patients	CADe system could improve the accuracy of detecting serrated polyps or cancer tissues
Case control study	Ma et al[33], 2020	681	Portal venous phase timing algorithm	Training through 479 CT scan data sets; 202 CT scans were used for retrospective analysis and algorithm development and verification	It was helpful to quantitatively describe the characteristics of tumor enhancement
Case control study	Soomro et al[34], 2018	12	3D fully convolutional neural networks	The effects of polyp segmentation and recognition of different models were compared using MRI data of 12 patients	3D fully convolutional neural networks provided a more accurate segmentation result of colon MRI
Case control study	Soomro et al[35], 2019	43	DL	43 patients with CRC were evaluated by MRI. The data set was divided into 30 volumes for training and 13 volumes for testing	DL achieved better performance in colorectal tumor segmentation in volumetric MRI
Retrospective study	Wang et al[36], 2020	240	Faster R-CNN	The Faster R-CNN was trained using pelvic MRI images to establish an AI platform. The diagnosis results of AI platform were compared with those of senior radiologists	It was highly feasible to segment the circumcision positive margin with Faster R-CNN in MRI image of rectal cancer
Retrospective study	Wu et al[37], 2021	183	Faster R-CNN	The MRI data of 183 patients were collected as training objects. The platform was constructed using Faster R-CNN. The diagnostic accuracy was compared with that of radiologists	AI could effectively predict the T stage of rectal cancer
Case control study	Joshi et al[38], 2010	10	Non-parametric mixture model	Compared the accuracy of the algorithm and expert conclusions through the patient's MRI images	The algorithm could be used to distinguish T3 and T4 tumors accurately
Case control study	Shiraishi et al[40], 2020	314	CNN	The prognostic significance was evaluated by CNN based on the expression of tumor markers in 314 patients	CNN could help to evaluate the diagnosis and prognosis of tumor markers
Case control study	Pham[41], 2017	NA	DL	NA	DL could reduce training time and improve classification rate
Case control study	Tiwari[42], 2018	10	CNN	CNN was used to compare the accuracy of image classification methods for seven different tissue types	CNN determined the most suitable color for cancer tissue classification (HSV color space) by classifying tissues in different color spaces
Case control study	Sirinukunwattana et al[43], 2016	100	SC-CNN	Through the comparative evaluation on the image data set of 100 cases of CRC, SC-CNN was helpful to the quantitative analysis of tissue components	SC-CNN can help to predict the nuclear class tags more accurately
Case control study	Koohababni et al[44], 2018	NA	DL	NA	DL could combine the probability maps of a single nucleus to generate the final image, so as to improve the diagnostic performance of complex colorectal adenocarcinoma datasets
Case control study	Zhang et al[45], 2018	NA	Faster R-CNN	NA	Faster R-CNN provided quantitative analysis of tissue composition in pathological practice
Case control study	Xu et al[46], 2016	1376	DCNN	Compared the classification effects of AI and manual methods on the same pathological image dataset	DCNN can help to improve the accuracy of differentiation between epithelial and mesenchymal regions in digital tumor tissue microarray
Retrospective study	Chen et al[47], 2017	85	Deep contour-aware network	The classification performance of different segmentation methods on the same pathological image dataset was compared	Output accurate probability map of gland cells, draw clear outline to separate the originally gathered cells, and further improve the segmentation performance
Case control study	Yoshida et al[48], 2017	1328	An automated image analysis system	The classification results of the same dataset by human pathologists and electronic pathologists were compared	Compared with manual classification, the system had higher classification accuracy
Retrospective study	Saito et al[49], 2013	NA	CAD system	NA	CAD system could be used for quality control, double check diagnosis, and prevention of missed diagnosis of cancer
Descriptive research	Jin et al[50], 2019	NA	AI	NA	AI accelerated the transformation of pathology to quantitative direction, and provided annotation storage, sharing, and visualization services
Case control study	Qaiser et al[51], 2019	75	CNN	The segmentation and recognition effects of different methods on the same pathological dataset were compared	CNN and PHPs can more accurately and quickly distinguish tumor regions from normal regions by simulating the atypical characteristics of tumor nuclei
Retrospective study	Zhou et al[53], 2020	120	DCNN	In the man-machine competition of 120 images, the accuracy of AI and endoscopists was compared	DCNN helped to establish an objective and stable bowel preparation system
Case control study	de Almeida et al[54], 2019	NA	CNN	NA	CNN improved the accuracy of polyp segmentation. It can help to automatically increase the sample number of medical image analysis dataset
Case control study	Taha et al[56], 2017	15	DL	The effectiveness of the DL method for identifying polyps in colonoscopy images was verified on the public database	In the early screening of CRC, it was better than other single models
Case control study	Yao et al[57], 2019	NA	DL	NA	A DL algorithm in HSV color space was designed to effectively improve the accuracy of diagnosis and reduce the cost
Case control study	Bravo et al[59], 2018	NA	Supervised learning model	NA	Supervised learning model could help to detect polyps more than 5 mm automatically with high accuracy
Review	de Lange et al[60], 2018	NA	CAD system	NA	CAD system could eliminate the leakage rate of polyps, thus avoiding polyps from developing into CRC
Case control study	Mahmood et al[61], 2018	NA	CAD system	NA	CAD system combined with depth map could more accurately identify polyps or early cancer tissue
Retrospective study	Mo et al[62], 2018	16	DL	Compared the performance of multiple algorithms in the same dataset	DL was in the leading position in many aspects such as the performance of evolutionary algorithm, and was an effective clinical method
Case control study	Zhu et al[63], 2010	50	CAD system	Through the database of 50 patients, the performance differences of different segmentation strategies were compared	Initial polyp candidates could greatly facilitate the FP reduction process of CAD system
Case control study	Komeda et al[64], 2017	1200	CNN-CAD system	The efficiency of CNN-CAD system was evaluated by maintaining cross validation for 10 times	CNN-CAD system can quickly diagnose colorectal polyp classification
Retrospective study	Zhang et al[65], 2018	18	CNN-CAD system	Through the video of 18 cases of colonoscopy, the efficiency of polyp detection between CNN-CAD system and existing methods was compared	CNN-CAD system can reduce the chance of missed diagnosis of polyps
Case control study	Zhu et al[66], 2019	357	CNN	The diagnostic performance of CNN was trained, fine-tuned, and evaluated using endoscopic data of 357 patients, and compared with that of manual diagnosis	The sensitivity of CNN optical diagnosis is higher than that of endoscopy, but the specificity is lower than that of endoscopy
Retrospective study	Akbari et al[67], 2018	300	FCN	The polyp segmentation method based on CNN was evaluated using CVC ColonDB database	FCN proposed a new method of image block selection and the probability map was processed effectively
Retrospective study	Yu et al[68], 2017	NA	3D-FCN	NA	3D-FCN could learn representative spatiotemporal features, and it had strong recognition ability
Case control study	Yamada et al[69], 2019	4395	AI	The AI system was trained through a large amount of data to make it sufficient to detect missed non polypoid lesions with high accuracy	AI could automatically detect the early features of CRC and improve the early detection rate of CRC
Retrospective study	Lund et al[71], 2019	20	DL	Polyp video dataset was used as training data. At the same time, a 5-fold cross validation method was used to evaluate the accuracy of the system	DL could improve the network training efficiency of polyp detection accuracy
Meta-analysis	Takamaru et al[73], 2020	NA	Endocytoscopy	NA	AI combined with endocytoscopy could greatly improve the efficiency of optical biopsy of CRC
Review	Djinbachian et al[76], 2019	NA	AI	NA	The sensitivity of optical diagnosis based on AI could be comparable to that of experienced endoscopists
Retrospective study	Kudo et al[77], 2019	69142	EndoBRAIN	A retrospective comparative analysis was performed between EndoBRAIN and 30 endoscopists on the diagnostic performance of endoscopic images in the same dataset	In the image of color cell endoscopy, EndoBRAIN could distinguish between tumor and non-tumor lesions accurately
Retrospective study	Mahmood et al[78], 2018	NA	CRF	NA	CRF estimated the depth of the colonoscopy image and reconstructed the surface structure of the colon
Case control study	Jian et al[81], 2018	2772	FCN	Quantitative comparison of manual and AI segmentation results of 2772 cases of CRC in MRI images	FCN was helpful for accurate segmentation of colorectal tumors
Case control study	Sivaganesan[82], 2016	20	RNN-ALGA	In the same database, milestone algorithms such as graph cut and level set were compared with RNN-ALGA algorithm	RNN-ALGA is suitable for abdominal slice of CT image, which can improve the accuracy and time efficiency of structure segmentation
Case control study	Gayathri et al[83], 2015	NA	NN	NA	NN can help to remove the colonic effusion and obtain the ideal colon segmentation effect
Retrospective study	Therrien et al[84], 2018	NA	SVM, CNN	NA	Using multiple datasets to train SVM and CNN could more accurately distinguish CRC staining tissue than single dataset
Case control study	Sun et al[85], 2019	NA	ML	NA	ML increased the chance of recognizing tumor bud by narrowing the region, thus providing effective tissue classification
Case control study	Shi et al[86], 2010	NA	DS-STM	NA	DS-STM could reduce the cost of diagnosis
Case control study	Su et al[87], 2012	212	MVMTM	The training set included 124 cases. The validation set included 88 cases. Comparedthe diagnostic efficiency of different methods for CRC	Compared with the traditional ML method, MVMTM has the advantages of low cost
Case control study	Kunhoth et al[88], 2017	80	Multispectral image acquisition system	A group of 20 samples were selected from 4 different types of colorectal cells. Compared the accuracy of different feature extraction methods	The database developed by this system had high classification accuracy
Case control study	Wang et al[89], 2018	1290	DL	Through the data of 1290 patients, an AI algorithm for real-time polyp detection was developed and verified	Compared with ML, DL could detect polyps in real time and reduce the cost
Meta-analysis	Barua et al[90], 2021	NA	AI	NA	AI based polyp detection system could increase the detection of small non-progressive adenomas and polyps
Randomized controlled study	Gong et al[91], 2020	704	ENDOANGEL system	704 patients were randomly assigned to use the ENDOANGEL system for colonoscopy or unaided (control) colonoscopy to compare the efficiency of ENDOANGEL system with conventional colonoscopy	The system significantly improved the detection rate of adenoma in colonoscopy
Meta-analysis	Lui et al[92], 2020	NA	AI	NA	AI system could improve the detection rate of adenoma and reduce the missed lesions in real-time colonoscopy
Case control study	Rodriguez-Diaz et al[93], 2011	134	A diagnostic algorithm with ESS	80 patients were randomly assigned to the training set, and the remaining 54 patients were assigned to the test set for prospective verification by the new algorithm	The algorithm with ESS reduced the risk and cost of biopsy, avoided the removal of non-neoplastic polyps, and reduced the operation time
Case control study	Kondepati et al[94], 2007	37	ANN	The tumor recognition accuracy of different algorithms was compared by collecting the spectra of cancer tissue and normal tissue	The spectrum was divided into cancer tissue group and normal tissue group by ANN, and the accuracy was 89%
Case control study	Angermann et al[95], 2016	NA	AL	NA	AL helped to realize real-time detection and distinguish between polyps and cancer tissues
Case control study	Ayling et al[96], 2019	619	ColonFlagTM	Through the clinical data of 619 patients, the performance of different systems in detecting CRC and high adenoma was compared	ColonFlagTM could help special patients establish an appropriate safety net
Meta-analysis	Tian et al[97], 2020	4560	EPE	Ten randomized controlled trials were included and 4560 participants were included for meta-analysis	EPE could guide the intestinal preparation of patients undergoing colonoscopy, and improve the detection rate of polyps, adenomas, and sessile serrated adenomas
Retrospective study	Javed et al[98], 2018	NA	QSL	NA	The prevalent communities found by QSL represented different tissue phenotypes with biological significance
Case control study	Wang et al[99], 2019	328	ANN	Different diagnostic models were established by back propagation and other methods, and the performance of each model was evaluated by cross validation test	ANN combined with gene expression profile data could improve the diagnosis mode of CRC
Case control study	Battista et al[100], 2019	345	ANN	The diagnostic performance and FP of the new model were measured in the experimental group (patients with CRC) and the control group (patients with good health)	ANN could help to establish an easily available, low-cost mathematical tool for CRC screening
Review	Zhang et al[101], 2021	NA	ML	NA	ML based on cell-free DNA and microbiome data helped diagnose CRC
Case control study	Wang et al[102], 2021	9631	DCNN	The diagnostic accuracy of AI tools and experienced expert pathologists was compared through the same database	A novel strategy for clinic CRC diagnosis using weakly labeled pathological whole-slide image patches based on DCNN
Review	Jones et al[103], 2021	NA	AI	NA	Electronic health record type data combined with AI could help diagnose early cancer
Case control study	Lorenzovici et al[104], 2021	33	A computer aided diagnosis system	The accuracy of the system in diagnosing CRC was tested through a dataset of 33 patients	The system used ML to improve the accuracy of CRC diagnosis
Review and Meta-analysis	Xu et al[105], 2021	NA	CNN	NA	Through the comparative study of online database, CNN system had good diagnostic performance for CRC
Case control study	Öztürk et al[106], 2021	NA	CNN	NA	CNN was the most successful method that could effectively classify gastrointestinal image datasets with a small amount of labeled data
Review	Echle et al[107], 2021	NA	DL	NA	DL could directly extract the hidden information from the conventional histological images of cancer, so as to provide potential clinical information

NA: Not available; DL: Deep learning; ML: Machine learning; AL: Active learning; QSL: Quasi-supervised learning; CNN: Convolutional neural network; CRC: Colorectal cancer; SVM: Support vector machine; CAD: Computer-aided diagnosis; CTC: Computed tomography colonography; CT: Computed tomography; FP: False-positive rate; DCNN: Deep convolutional neural network; CADe: Computer-aided detection; 3D: Three-dimensional; MRI: Magnetic resonance imaging; AI: Artificial intelligence; R-CNN: Region-based convolutional neural network; SC-CNN: Space-constrained convolutional neural network; PHPs: Persistent homology maps; HSV: Hue, saturation, value; FCN: Fully convolutional network; CRF: Conditional random field; DS-STM: Diagnosis strategy of serum tumor maker; MVMTM: Multiple tumor markers with multiple cut-off values; ANN: Artificial neural network.