A Review of Radiomics in Predicting Therapeutic Response in Colorectal Liver Metastases: From Traditional to Artificial Intelligence Techniques

Multiparameter magnetic resonance imaging-based radiomics model for the prediction of rectal cancer metachronous liver metastasis

BACKGROUND

The liver, as the main target organ for hematogenous metastasis of colorectal cancer, early and accurate prediction of liver metastasis is crucial for the diagnosis and treatment of patients. Herein, this study aims to investigate the application value of a combined machine learning (ML) based model based on the multiparameter magnetic resonance imaging for prediction of rectal metachronous liver metastasis (MLM).

AIM

To investigate the efficacy of radiomics based on multiparametric magnetic resonance imaging images of preoperative first diagnosed rectal cancer in predicting MLM from rectal cancer.

METHODS

We retrospectively analyzed 301 patients with rectal cancer confirmed by surgical pathology at Jingzhou Central Hospital from January 2017 to December 2023. All participants were randomly assigned to the training or validation queue in a 7:3 ratio. We first apply generalized linear regression model (GLRM) and random forest model (RFM) algorithm to construct an MLM prediction model in the training queue, and evaluate the discriminative power of the MLM prediction model using area under curve (AUC) and decision curve analysis (DCA). Then, the robustness and generalizability of the MLM prediction model were evaluated based on the internal validation set between the validation queue groups.

RESULTS

Among the 301 patients included in the study, 16.28% were ultimately diagnosed with MLM through pathological examination. Multivariate analysis showed that carcinoembryonic antigen, and magnetic resonance imaging radiomics were independent predictors of MLM. Then, the GLRM prediction model was developed with a comprehensive nomogram to achieve satisfactory differentiation. The prediction performance of GLRM in the training and validation queue was 0.765 [95% confidence interval (CI): 0.710-0.820] and 0.767 (95%CI: 0.712-0.822), respectively. Compared with GLRM, RFM achieved superior performance with AUC of 0.919 (95%CI: 0.868-0.970) and 0.901 (95%CI: 0.850-0.952) in the training and validation queue, respectively. The DCA indicated that the predictive ability and net profit of clinical RFM were improved.

CONCLUSION

By combining multiparameter magnetic resonance imaging with the effectiveness and robustness of ML-based predictive models, the proposed clinical RFM can serve as an insight tool for preoperative assessment of MLM risk stratification and provide important information for individual diagnosis and treatment of rectal cancer patients.

Computed Tomography-Based Radiomics with Machine Learning Outperforms Radiologist Assessment in Estimating Colorectal Liver Metastases Pathologic Response After Chemotherapy

OBJECTIVES

This study was designed to assess computed tomography (CT)-based radiomics of colorectal liver metastases (CRLM), extracted from posttreatment scans in estimating pathologic treatment response to neoadjuvant therapy, and to compare treatment response estimates between CT-based radiomics and radiological response assessment by using RECIST 1.1 and CT morphologic criteria.

METHODS

Patients who underwent resection for CRLM from January 2003-December 2012 at a single institution were included. Patients who did not receive preoperative systemic chemotherapy, or without adequate imaging, were excluded. Imaging characteristics were evaluated based on RECIST 1.1 and CT morphologic criteria. A machine-learning model was designed with radiomic features extracted from manually segmented posttreatment CT tumoral and peritumoral regions to identify pathologic responders (≥ 50% response) versus nonresponders. Statistical analysis was performed at the tumor level.

RESULTS

Eighty-five patients (median age, 62 years; 55 women) with 95 tumors were included. None of the subjectively evaluated imaging characteristics were associated with pathologic response (p > 0.05). Inter-reader agreement was substantial for RECIST categorical response assessment (K = 0.70) and moderate for CT morphological group response (K = 0.50). In the validation cohort, the machine learning model built with radiomic features obtained an area under the curve (AUC) of 0.87 and outperformed subjective RECIST assessment (AUC = 0.53, p = 0.01) and morphologic assessment (AUC = 0.56, p = 0.02).

CONCLUSIONS

Radiologist assessment of oligometastatic CRLM after neoadjuvant therapy using RECIST 1.1 and CT morphologic criteria was not associated with pathologic response. In contrast, a machine-learning model based on radiomic features extracted from tumoral and peritumoral regions had high diagnostic performance in assessing responders versus nonresponders.