Artificial intelligence in gastrointestinal cancer: Recent advances and future perspectives

Table 1 Previous studies on upper endoscopy of gastric cancer using artificial intelligence

Ref.	Targets	Sample sizes	Inputs	Tasks	Analysis method	Diagnostic performance
Yoon et al[28]	GC (ESD/surgery)	800 cases	GC/non-GC images in close-up and distant views	Detection and invasion depth prediction	CNN	AUC: detection, 0.981; depth, 0.851
Zhu et al[29]	GC	993 images	GC images	Diagnosis of invasion depth	CNN	Sensitivity: 76.4%, PPV: 89.6%
Li et al[30]	GC and healthy	386 GC and 1702 NC images	NBI images	Diagnosis of GC	CNN	Sensitivity: 91.1%, PPV: 90.6%
Hirasawa et al[31]	GC	13584 training and 2296 test images	GC images	Diagnosis of GC	CNN	Sensitivity: 92.2%, PPV: 30.6%
Ishioka et al[32]	EGC	62 cases	Real-time images	Detection	CNN	Detection rate: 94.1%
Luo et al[33]	GC	1036496 images	GC images	Detection	CNN	PPV: 0.814, NPV:0.978
Horiuchi et al[34]	GC and gastritis	1492 GC and 1078 gastritis images	NBI images	Detection	CNN	Sensitivity: 95.4%, PPV: 82.3%

GC: Gastric cancer; CNN: Convolutional neural network; AUC: Area under the curve; PPV: Positive predictive value; NC: Non-cancer; NBI: Narrow-band image; EGC: Early gastric cancer.

Table 2 Previous studies on colonoscopy using artificial intelligence

Ref.	Targets	Sample sizes	Inputs	Tasks	Analysis method	Diagnostic performance
Akbari et al[35]	Screening endoscopy	300 polyp images	Polyp images	Auto segmentation of polyps	CNN	Accuracy: 0.977, Sensitivity: 74.8%
Jin et al[36]	Screening endoscopy	Training: 2150 polyps, test: 300 polyps	NBI images	Differentiation of adenoma and hyperplastic polyps	CNN	The model reduced the time of endoscopy and increased accuracy by novice endoscopists
Urban et al[37]	Screening endoscopy	8641 polyp images and 20 colonoscopy videos	Polyp images	Detection of polyps	CNN	AUC: 0.991, Accuracy: 96.4%
Yamada et al[38]	Screening endoscopy	4840 images, 77 colonoscopy videos	Real-time images	Differentiation of the early signs of CRC	CNN	Sensitivity: 97.3%, Specificity: 99.0%

CNN: Convolutional neural network; NBI: Narrow-band image; AUC: Area under the curve.

Table 3 Previous studies on the pathology of gastric cancer using artificial intelligence

Ref.	Targets	Sample size	Input	Task	Analysis method	Diagnostic performance
Qu et al[39]	GC	15000 images	Pathological images	Evaluation of stepwise methods	CNN	AUC: 0828-0.920
Yoshida et al[40]	GC	3062 biopsy samples	Pathological images stained by H&E	Automatic segmentation, diagnosis of carcinoma	CNN	Sensitivity: 89.5%, specificity: 50.7%
Mori et al[41]	GC (surgery)	516 images from 10 GC cases	Pathological images stained by H&E	Diagnosis of invasion depth in signet cell carcinoma	CNN	Sensitivity: 90%, Specificity: 81%
Jiang et al[42]	GC (surgery)	786 cases	IHC (CD3, CD8, CD45RO, CD45RA, CD57, CD68, CD66b, and CD34)	Prediction of survival	SVM	The immunomarker SVM was useful for predicting survival

GC: Gastric cancer; AUC: Area under the curve; H&E: Hematoxylin eosin staining; CNN: Convolutional neural network; IHC: Immunohistochemistry; SVM: Support vector machine.

Table 4 Previous studies on the pathology of colorectal cancer using artificial intelligence

Ref.	Targets	Sample size	Input	Task	Analysis method	Diagnostic performance
Van Eycke et al[43]	CRC		H&E staining, IHC image	Segmentation of the glandular epithelium	TMA, CNN	F1 value: 0,912
Graham et al[44]	CRC		H&E staining	Differentiation of intratumor glands	CNN	F1 values: 0.90
Abdelsamea et al[45]	CRC	333 samples	H&E staining, IHC (CD3)	Differentiation of the tumor epithelium	TMA, CNN	Accuracy: 0.93-0.94
Yan et al[46]	CRC		H&E staining	Tumor classification,segmentation of tumors,	CNN	Accuracy: Classification, 97.8%; segmentation, 84%
Haj-Hassan et al[47]	CRC		Multispectral images	Segmentation of carcinoma	CNN	Accuracy: 99.1%
Rathore et al[48]	CRC	Biopsy samples	H&E staining	Detection and grading of tumors	Texture and morphology patterns, SVM	Recognition rate: Detection, 95.4%; grading; 93.4%
Yang et al[49]	CRC	180 samples	H&E staining	Diagnosis of benign tumors, neoplasms, and carcinoma	SVM, histogram, texture	AUC: 0.852
Chaddad et al[50]	CRC	30 cases	H&E staining	Diagnosis of carcinoma, adenoma, and benign tumors	Automatic segmentation, texture	Accuracy: 98.9%
Yoshida et al[51]	CRC	1328 samples	H&E staining	Diagnosis of benign tumors, neoplasms, and carcinoma	CNN, automatic analysis of structure	Undetected rate of carcinoma and adenoma: 0-9.3% and 0-9.9%, respectively
Takamatsu et al[52]	CRC surgery	397 samples	H&E staining	Prediction of lymph node metastasis	LR, shape analysis	AUC: 0.94
Weis et al[53]	CRC	596 cases	IHC (AE1/AE3)	Automatic evaluation of tumor budding	TMA, CNN	Correlation; R2 value: 0.86
Bychkov et al[54]	CRC surgery	420 cases	H&E staining	Prediction of survival	TMA, CNN	Good biomarker for predicting survival
Kather et al[55]	CRC	973 slides	H&E staining	Prediction of survival	Stromal pattern, CNN	Good biomarker for predicting survival
Reichling et al[56]	CRC surgery	1018 cases	HE, IHC (CD3, CD8)	Prediction of survival	RF, monogram	Good biomarker for predicting survival

CRC: Colorectal cancer; H&E: Hematoxylin eosin staining; IHC: Immunohistochemistry; TMA: Tissue microarray; CNN: Convolutional neural network; SVM: Support vector machine; AUC: Area under the curve; LR: Linear regression.

Table 5 Previous studies on the radiological diagnosis of gastric cancer using radiomics or artificial intelligence

Ref.	Targets	Sample size	Input	Task	Analysis method	Diagnostic performance
Li et al[57]	GC, radical surgery	181 cases	Primary tumor, preoperative CT	Prediction of survival	Manual segmentation, radiomics, Nomograms	The TNM stage and radiomics signature were good biomarkers
Zhang et al[58]	GC, radical surgery	669 cases	Primary tumor, preoperative CT	Predication of early recurrence	Manual segmentation, radiomics, Nomograms	AUC: 0.806-0.831
Li et al[59]	GC, radical surgery	204 cases	Primary tumor, pre-operative dual-energy CT	Pre-operative diagnosis of LNM	Manual segmentation, radiomics, Nomogram	AUC; 0.82--.84
Li et al[60]	GC, radical surgery	554 cases	Primary tumor, preoperative CT	Prediction of a pathological status, survival	Semi-automatic segmentation, radiomics	AUC for prediction of the pathological status: 0.77, the TNM stage and radiomics signature were good biomarkers
Wang et al[61]	GC, radical surgery	187 cases	Primary tumor, preoperative dynamic CT	Pre-operative prediction of intestinal-type GC	Manual segmentation, radiomics, Nomograms	AUC: 0.904
Jiang et al[62]	GC, surgery	214 cases	Primary tumor, preoperative PET-CT	Prediction of survival	Manual segmentation, radiomics, Nomograms	C-index: DFS, 0.800; OS, 0.786
Chen et al[63]	GC, surgery	146 cases	Primary tumor, preoperative MRI	Pre-operative diagnosis of lymph node metastasis	Manual segmentation, radiomics analysis	AUC: 0.878
Gao et al[64]	GC, surgery	627 cases, 17340 images	Lymph nodes, preoperative CT	Pre-operative diagnosis of lymph node metastasis	Manual segmentation, deep learning	AUC: 0.9541.
Huang et al[65]	GC, surgery		Primary tumor, preoperative CT	Pre-operative diagnosis of peritoneal metastasis	Manual segmentation, CNN	Ongoing, retrospective cross-sectional study

GC: Gastric cancer; CT: Computed tomography; AUC: Area under the curve; LNM: Lymph node metastasis; DFS: Disease-free survival; MRI: Magnetic resonance imaging; CNN: Convolutional neural network.

Table 6 Previous studies on the radiological diagnosis of colorectal cancer using radiomics or artificial intelligence

Ref.	Targets	Sample size	Input	Task	Analysis method	Diagnostic performance
Trebeschi et al[66]	LRC	140 cases	Primary tumor, MRI	Automatic detection, segmentation	CNN	DSC: 0.68-0.70, AUC: 0.99
Wang et al[67]	LRC	568 cases	Primary tumor, MRI	Automatic segmentation	CNN	DSC: 0.82
Wang et al[68]	LRC	93 cases	Primary tumor, MRI	Automatic segmentation	Deep learning	DSC: 0.74
Men et al[69]	LRC	278 cases	Primary tumor, CT	Automatic segmentation	CNN	DSC: 0.87
Shayesteh et al[70]	LRC, NCRT followed by surgery	98 cases	Primary tumor, pre-treatment MRI	Prediction of CRT responses	Manual segmentation, radiomics, machine learning	AUC: 0.90
Shi et al[71]	LRC, NCRT followed by surgery	45 cases	Primary tumor, pre-treatment MRI, mid-radiation MRI	Prediction of CRT responses	Manual segmentation, CNN	AUC: CR, 0.83; good response, 0.93
Ferrari et al[72]	LRC, NCRT followed by surgery	55 cases	Primary tumor, MRI before, during and after CRT	Prediction of CRT responses	Manual segmentation, radiomics, RF	AUC: CR: 0.86, non-response: 0.83
Bibault et al[73]	LRC, NCRT followed by surgery	95 cases	Primary tumor, pre-operative CT	Prediction of CRT responses	Manual segmentation, radiomics, CNN	80% accuracy
Dercle et al[74]	CRC, FOLFILI with/without cetuximab	667 cases	Metastatic tumor, CT	Prediction of tumor sensitivity to chemotherapy	Manual segmentation, radiomics, machine learning	AUC: 0.72-0.80
Ding et al[75]	LRC, radical surgery	414 cases	Lymph nodes, pre-operative MRI	Pre-operative diagnosis of lymph node metastasis	Manual segmentation, CNN	AI system > radiologist
Taguchi et al[76]	CRC	40 cases	Primary tumor, CT	Prediction of the KRAS status	Manual segmentation, radiomics	AUC: 0.82

LRC: Lower rectal cancer; MRI: Magnetic resonance imaging; CNN: Convolutional neural network; DSC: Dice similarity coefficient; AUC: Area under the curve; NCRT: Neoadjuvant chemoradiotherapy; CR: Complete response; RF: Random forest; CT: Computed tomography; CRC: Colorectal cancer.