CT-Based Radiomics Models with External Validation for Prediction of Recurrence and Disease-Specific Mortality After Radical Surgery of Colorectal Liver Metastases.
Authors
Affiliations (11)
Affiliations (11)
- Medical Physics Laboratory, IRCCS Regina Elena National Cancer Institute, Rome, Italy.
- Radiology and Diagnostic Imaging Department, IRCCS Regina Elena National Cancer Institute, Rome, Italy. [email protected].
- Hepatopancreatobiliary Surgery, IRCCS, Regina Elena National Cancer Institute, Rome, Italy.
- Radiology and Diagnostic Imaging Department, IRCCS Regina Elena National Cancer Institute, Rome, Italy.
- Radiology Unit, Department of Medical-Surgical Sciences and Translational Medicine, Sant'Andrea University Hospital, "Sapienza" University of Rome, Rome, Italy.
- Biostatistics and Bioinformatic Unit, San Gallicano Dermatological Institute IRCCS, Rome, Italy.
- Department of Medical Surgical Science and Translational Medicine, Sant'Andrea University Hospital, "Sapienza" University of Rome, Rome, Italy.
- General Surgery and Liver Transplantation Unit, Fondazione IRCCS Istituto Nazionale Tumori, Milan, Italy.
- General Surgery 1 Unit, Azienda ULSS 2 Marca Trevigiana, Treviso, Italy.
- Department of Radiology, Ca' Foncello General Hospital, Treviso, Italy.
- HepatoBiliaryPancreatic Surgery, AOU Careggi, Department of Experimental and Clinical Medicine (DMSC), University of Florence, Florence, Italy.
Abstract
To build computed tomography (CT)-based radiomics models, with independent external validation, to predict recurrence and disease-specific mortality in patients with colorectal liver metastases (CRLM) who underwent liver resection. 113 patients were included in this retrospective study: the internal training cohort comprised 66 patients, while the external validation cohort comprised 47. All patients underwent a CT study before surgery. Up to five visible metastases, the whole liver volume, and the surrounding free-disease parenchymal liver were separately delineated on the portal venous phase of CT. Both radiomic features and baseline clinical parameters were considered in the models' building, using different families of machine learning (ML) algorithms. The Support Vector Machine and Naive Bayes ML classifiers provided the best predictive performance. A relevant role of second-order and higher-order texture features emerged from the largest lesion and the liver residual parenchyma. The prediction models for recurrence showed good accuracy, ranging from 70% to 78% and from 66% to 70% in the training and validation sets, respectively. Models for predicting disease-related mortality performed worse, with accuracies ranging from 67% to 73% and from 60% to 64% in the training and validation sets, respectively. CT-based radiomics, alone or in combination with baseline clinical data, allowed the prediction of recurrence and disease-specific mortality of patients with CRLM, with fair to good accuracy after validation in an external cohort. Further investigations with a larger patient population for training and validation are needed to corroborate our analyses.