Title:The Efficiency of the CT Radiomics Model in Assessing the Microsatellite
Instability of Colorectal Cancer Liver Metastasis
Volume: 20
Author(s): Yun Wang, Luyao Ma, Haifeng Guo, Xuehua Wang, Zhaoxiang Ye*, Shuxuan Fan, Bulang Gao*Xiao-Ping Yin
Affiliation:
- Tianjin Medical University Cancer Hospital, Tianjin, 300181, China
- CT-MRI, Affiliated Hospital of Hebei University, Hebei, 071000, China
Keywords:
Radiomics, Colorectal liver metastasis, Random forest, 5-fold cross-validation, Microsatellite instability, Colorectal cancer.
Abstract:
Objective:
This study aims to investigate the efficiency of a radiomics model in identifying high-frequency microsatellite instability (MSI-H) and
microsatellite stability (MSS) of colorectal liver metastasis (CRLM) according to machine learning radiomics features of enhanced CT liver
images.
Materials and Methods:
A total of 12 patients with MSI-H CRLM and 96 patients with MSS CRLM were randomly divided into the training group and internal validation
group according to the ratio of 7: 3 (training: 75 cases, validation: 33 cases). From the enhanced CT (portal phase) image data of patients, 788
radiomics features were extracted, and a random forest model was established with the optimal features selected. The receiver operating
characteristics (ROC) curve analysis was performed to assess the model’s diagnostic efficacy.
Results:
The training group comprised 8 patients with MSI-H CRLM and 67 patients with MSS CRLM, and the internal validation group included 4
patients with MSI-H CRLM and 29 patients with MSS CRLM. After feature selection, 7 radiomics features good for distinguishing MSI-H CRLM
and MSS CRLM were screened out. The ROC curve analysis demonstrated that the random forest model had the AUC (area under the ROC curve)
value 0.88, accuracy 0.85, sensitivity 0.85, specificity 0.92, and F1 score 0.88 in the training group. The model had an AUC value of 0.75,
accuracy of 0.74, sensitivity of 0.81, specificity of 0.85, and F1_score of 0.78 in the internal validation group in identifying the MSI-H from the
MSS CRLM. In order to evaluate the robustness of the overall model, the 788 features obtained were all applied to the 5-fold cross-validation, with
the model being built on the random forest and analyzed with the ROC curve analysis. The AUC value of the model was 0.86 (P<0.05), accuracy
value 0.91, sensitivity 0.60, and specificity 0.95.
Conclusion:
The random forest prediction model built on the radiometric features extracted from enhanced CT images can be used to identify the MSI-H from
the MSS CRLM and may provide effective guidance for clinical immunotherapy of CRLM patients with unknown MSI status.
