Machine learning-based radiomics models for the prediction of metachronous liver metastases in patients with colorectal cancer: A multimodal study

Abstract

The aim of the present study was to investigate whether a multimodal radiomics model powered by machine learning could accurately predict the occurrence of metachronous liver metastasis (MLM) in patients with colorectal cancer (CRC). A total of 157 patients diagnosed with CRC between 2010 and 2020 were retrospectively included in the present study; of these patients, 67 patients developed liver metastases within 2 years of treatment, while the remaining patients (n=90) did not. Radiomics features were extracted from annotated MR images of the tumor and portal venous phase CT images of the liver in each patient. Subsequently, machine learning-based radiomics models were developed and integrated with the clinical features for MLM prediction, employing Least Absolute Shrinkage and Selection Operator and Random Forest algorithms. The performance of the models were evaluated using the receiver operating characteristic curve analysis, while the clinical utility was measured using the decision curve analysis. A total of 922 and 1,082 radiomics features were extracted from the MR and CT images of each patient, respectively, which quantified the intensity, shape, orientation and texture of the tumor and liver. The mean area under the curve (AUC) values for the prediction of MLM were 0.80, 0.68 and 0.82 for the CT, MRI and merged models, respectively. For the clinical and clinical-merged models, the AUC values were 0.62 and 0.75, respectively. There was no significant difference between the CT model and the merged model (P>0.05). In conclusion, the preliminary results of the present study demonstrated the utility of machine learning-based radiomics models in the prediction of MLM in patients with CRC. However, further research is warranted to explore the potential of multimodal fusion models, due to the minimal improvement observed in diagnostic performance.

Keywords: colorectal cancer; machine learning; metachronous liver metastasis; multimodal; radiomics.

Figure 1.

Flow chart of the present study cohort.

Figure 2.

Image acquisition and segmentation. The ROI of the tumor was segmented and reconstructed in (A) and. (C) The ROI of the liver was segmented and reconstructed in (B) and. (D) ROI, region of interest.

Figure 3.

Radiomics feature extraction. GLSZM, gray level size zone matrix; GLCM, gray level co-occurrence matrix; NGTDM, neighboring gray tone difference matrix.

Figure 4.

Feature selection and model development. (A) Distribution and ranking of optimal radiomic features to predict whether the patients had MLM. (B) Coefficient profiles of the radiomic features in lasso model. Each colored line represents the coefficient path of a radiomic feature. MLM, metachronous liver metastases; lasso, least absolute shrinkage and selection operator; glszm, gray level size zone matrix; LLL, low low low; HHL, high high low.

Figure 5.

ROC-AUC of different models for the prediction of MLM. The models are represented as follows: Clinical, green line; merged, red line; CT, blue line; MRI, orange line; and clinical-merged, purple line. MLM, metachronous liver metastases; ROC, receiver operating characteristic; AUC, area under the curve; merged, the fusion radiomics signature based on features from CT and MRI; clinical-merged, the combined model incorporating clinical and radiological variables together.

Figure 6.

Decision curve analysis for the nomogram for the prediction of MLM in all patients. The solid black line indicated the net benefit of assuming that all patients have MLM; the dotted black line indicates the net benefit of the assumption that no patients have MLM and the red line indicates expected net benefit of each patient. At a threshold probability of 0.38, the net benefit of the individualized prediction (red line) equals that of the ‘Treat all’ strategy (black solid line), while at 0.93, it matches the net benefit of the ‘Treat none’ strategy (black dotted line). MLM, metachronous liver metastases.

Figure 7.

Nomogram for the prediction of MLM probabilities. A radiomics nomogram, which incorporated imaging histological features and clinical factors. Clinical factors included N stage and EMVI. MLM, metachronous liver metastases; MR_N, N stage of magnetic resonance; EMVI, extramural vascular invasion; glszm, gray level size zone matrix.