Artificial intelligence in the diagnosis and management of colorectal cancer liver metastases

Table 1 Summary of the studies considered in this review

Author	Study design	AI model type	Data source	Total sample size/training cohort/validation cohort	AUC training/AUC validation	Sensitivity/specificity	PPV/NPV	Accuracy
CRLM development
Li et al[23] (2020)	Retrospective; Single center	Radiomics/ML	CT images ± clinical data	100/NA/80	0.90/0.906	81%/84%	85%/79%	NA
Taghavi et al[24] (2021)	Retrospective; Multicenter	Radiomics/ML	CT images ± clinical data	91/70/21	0.952-0.683-0.954/0.862-0.713-0.864	NA/NA	NA/NA	NA
Lee et al[25] (2020)	Retrospective; Single center	Radiomics/CNN	CT images ± clinical data	2019/1413/606	NA/0.6062-0.7093-0.7474	NA/NA	NA/NA	NA
Diagnosis
Vorontsov et al[26] (2017)	Retrospective; Single center	Radiomics/CNN	CT images	40/32/8	NA/NA	84%/92%	NA/NA	88%
Vorontsov et al[28] (2019)	Retrospective; Single center	Radiomics/CNN	CT images	156/115/15	NA/NA	59%5/NA	80%5/NA	NA
Ma et al[30] (2020)	Retrospective; Multicenter	CNN	CT images	909/479/202 (2286)	NA/0.837-0.8446	82%6/74%5	75%6/81%6	NA
Kim et al[31] (2021)	Retrospective; Single center	DL	CT images	587/502/85	NA/0.631	81.82%/22.22%	NA/NA	NA
Khalili et al[34] (2020)	Retrospective; Single center	CNN	CT images ± liver metastatic status	199/150/49	NA/0.84-0.957	(81.5%-81.5%7)/(76.2%-96.4%7)	NA/NA	78.3%; 90.6%6
Jansen et al[38] (2019)	Retrospective; Single center	CNN	MRI images	121/3341/861	NA/NA	99.8%/NA	NA/NA	NA
Steenhuis et al[39] (2020)	Retrospective; Single center	ML	VOCs	62/NA/NA	NA/0.86	88%/75%	72%/90%	81%
Miller-Atkins et al[40] (2020)	Prospective; Single center	ML	VOCs	296/284/NA	NA/NA	51%/94%	NA/NA	86%
Kiritani et al[41] (2021)	Retrospective; Single center	ML	Histologic markers	183/NA/40	NA/0.999	100%/99%	NA/NA	99.5%
Han et al[47] (2020)	Retrospective; Single center	Radiomics/ML	MRI images ± clinical data	107/611/311	0.9742-0.6593-0.9714/0.9122-0.6763-0.9094	95.2%2-57.1%3-95.2%4/80.0%2-70.0%3-70.0%4	NA/NA	90.3%2; 61.3%3; 87.1%4
Chemotherapy response
Maaref et al[54] (2020)	Retrospective; Single center	DL CNN	CT images	202/70%/10%	0.97/0.88	98%/54%	NA/NA	91%8; 78%9
Wei et al[55] (2021)	Retrospective; Single center	Radiomics/DL	CT images ± CEA	192/144/48	0.90310-0.93511/0.82010-0.83011	90.9%/73.3%	88.2%/78.6%	85.4%
Giannini et al[57] (2020)	Retrospective; Multicenter	Radiomics/ML	CT images	38/28/10	NA/NA	92%/86%	96%/75%	NA
Nakanishi et al[58] (2021)	Retrospective; Single center	Radiomics	CT images	42/941/321	0.8512/0.7792	NA/NA	NA/NA	NA
Local ablative therapies efficacy
Taghavi et al[59] (2021)	Retrospective; Single center	Radiomics/ML	CT images	90/63/27	NA/0.782-0.563-0.794	NA/NA	NA/NA	NA
Survival prediction
Mühlberg et al[60] (2021)	Retrospective; Single center	Radiomics/ML	CT images ± WLTB ± TBS	103/NA/NA	NA/0.7012–0.7313-0.7614	NA/NA	NA/NA	NA
Hao et al[62] (2017)	Retrospective; Multicenter	ML	DNA methylation	17921/NA/8841 (7181,6)	NA/NA	NA/NA	NA/NA	98.4%
Dercle et al[64] (2020)	Retrospective; Multicenter	ML	CT images	667/438/229	0.83/0.80	80%/78%	NA/NA	NA
Spelt et al[65] (2013)	Retrospective; Single center	ANN	Clinical variables	241/NA/NA	NA/NA	NA/NA	NA/NA	72%
Paredes et al[66] (2020)	Retrospective; Multicenter	ML	Clinical variables	1406/703/703	0.52715-0.52516-0.69317/0.52415-0.50116-0.64217	NA/NA	NA/NA	NA

¹Number of lesions.

²Model based on radiomics data only.

³Model based on clinical data only.

⁴Model based on both radiomics and clinical data.

⁵Per patient values.

⁶Values calculated on the external validation set.

⁷Model based on both convolutional neural network and liver metastatic status.

⁸For differentiating treated and untreated lesions.

⁹For predicting the response to a FOLFOX + bevacizumab-based chemotherapy regimen.

¹⁰Model based on both deep learning and radiomics signature.

¹¹Model based on deep learning and radiomics signature considering carcinogenic embryonic antigen values.

¹²Model based on tumor burden score.

¹³Model based on geometric metastatic spread of whole liver tumor burden.

¹⁴Model based on the Aerts radiomics prior model.

¹⁵Model based on Fong/Blumgart clinical risk score for predicting 1-year recurrence.

¹⁶Model based on Brudvik–Vauthey clinical risk score for predicting 1-year recurrence.

¹⁷Model based on Paredes–Pawlik clinical risk score for predicting 1-year recurrence.

AI: Artificial intelligence; ANN: Artificial neural network; AUC: Area under the curve; CEA: Carcinogenic embryonic antigen; CNN: Convolutional neural network; CRLM: Colorectal cancer liver metastases; DL: Deep learning; ML: Machine learning; NPV: Negative predictive value; PPV: Positive predictive value; TBS: Tumor burden score; VOCs: Volatile organic compounds; WLTB: Whole liver tumor burden.