Noninvasive diagnostic models based on CT scans for differentiating solitary pulmonary metastasis in colorectal cancer patients by artificial intelligence: A multicenter study

Indeterminate solitary pulmonary nodules are often encountered on CT scans, and diagnosis of solitary pulmonary metastasis (PM) is important for patients with colorectal cancer (CRC). Indeterminate solitary pulmonary nodules are often encountered on CT scans, and diagnosis of solitary pulmonary metastasis (PM) is important for patients with colorectal cancer (CRC). We aim to build and validate proper noninvasive artificial intelligence diagnostic models based on routine chest CT for solitary PM in CRC patients.

All patients (n=212) with pretreated solitary CRC PM on chest CT were reviewed in the local database in Zhejiang Cancer Hospital between 2012 and 2022, and randomly divided into the training or internal validation groups on 7: 3. A total of 185 patients with pretreated T1 stage primary lung cancer and 256 patients with benign solitary pulmonary nodules were randomly selected. Artificial intelligence models based on machine learning (Decision Tree, Extra Trees, Light GBM, Radom Forest, Support Vector Machine, XGBoost) and deep learning were built to classify the solitary pulmonary nodules as PM, benign lesion or primary lung cancer. External validation group included 44 CRC patients with solitary PM from two independent hospitals.

For classification between PM and benign solitary pulmonary nodule, the machine learning model based on support vector machine showed the best diagnostic ability, with a 0.995 area under the curve (AUC) in the internal validation, a 0.977 AUC in external validation. The deep learning model showed a 0.966 AUC in the internal validation, and 0.893 AUC in external validation. For classification between PM and lung cancer, the best machine learning model based on support vector machine showed a 0.991 AUC in internal validation and the deep learning model showed a 0.949 AUC.

Non-contrast CT based radiomic analyses can be useful for noninvasively differentiating solitary PM and benign pulmonary nodule or primary T1 stage lung cancer which can aid clinical decision in CRC patients with indeterminate solitary pulmonary nodules detected by CT.

The authors.

Has not received any funding.

All authors have declared no conflicts of interest.