Artificial intelligence for the prediction of synchronous and metachronous liver metastasis in colorectal cancer patients: a systematic review and meta-analysis.

backgroundColorectal cancer (CRC) is the third most common cancer and the second leading cause of cancer mortality in the world. Liver is the most common metastasis site (synchronous and metachronous) in this cancer, affecting 12.8% of these patients. Artificial intelligence (AI) models have shown significant potential in aiding the diagnosis of CRLM. Radiomic models extract quantitative features from medical images to characterize tumor heterogeneity, such as tumor shape, texture, and intensity, while deep learning models with their automated pipeline have emerged as the new research trend in this field. In this systematic review and meta-analysis, we aim to evaluate whether radiomics and deep learning models can predict the occurrence of synchronous and metachronous liver metastasis in patients with colorectal cancer (CRC).

methodsA comprehensive literature search was conducted on PubMed, Scopus, Embase, and Web of Science. Our criteria included studies that trained and validated AI models for predicting SLM and MLM in colorectal cancer patients, based on input imaging data. The Quality Assessment of Diagnostic Accuracy Studies (QUADAS) and the Radiomics Quality Score (RQS) tools were used for our quality assessment. Two different sets of analyses were conducted to calculate the pooled estimates for sensitivity and specificity, and a complementary bivariate analysis to evaluate both. Different moderators were assessed in our subgroup analyses and meta-regression to find the possible source of heterogeneity.

resultsOur systematic search yielded 6759 studies, of which, after title/abstract and full text screening, there were 21 final included studies. The pooled sensitivity was 0.80 (95% CI, 0.74-0.84) and the pooled specificity was 0.81 (95% CI, 0.72-0.88). The SROC curve (AUC) was 0.84 (p-AUC = 0.81), indicating a reasonably high discriminative ability of these models. Our clinical utility analyses revealed a PLR of 4.29 and an NLR of 0.25. Based on the pre-test probability of 38%, the final pooled PPV and NPV of the models were 0.72, 0.87 respectively. Our quality assessment resulted in a generally low risk of bias in the QUADAS tool, except for the reference standard and the index test domain, with frequent high and unclear risks of bias. Our RQS tool yielded a mean score of 16.4 out of 36, which indicates poor methodological quality among studies.

conclusionAI models show promising results for the accurate prediction of synchronous and metachronous colorectal cancer liver metastases. Although there are significant shortcomings in the methodological and reporting quality of included studies, highlighting a need for more robust standard research in the future.