Machine learning-based radiomics approach assessing preoperative non-contrast CT for microsatellite instability prediction in colon cancer.

Objectives: To assess the feasibility of non-contrast CT-based radiomics model for predicting microsatellite instability (MSI) status in colon cancer.

Methods: Leveraging non-contrast abdominal CT imaging data from 57 retrospectively enrolled patients with balanced class distribution (training cohort: n = 38, 19 non-MSI-H and 19 MSI-H; test cohort: n = 19, 9 non-MSI-H and 10 MSI-H), we implemented a voxel volume-based tumor feature selection method. Feature selection integrated four feature selection filters-correlation analysis, univariate logistic regression, least absolute shrinkage and selection operator (LASSO), and recursive feature elimination (RFE). We comparatively evaluated multiple classifiers using cross-validation combined with accuracy for choosing the best classifier.

Results: A multilayer perceptron-based classification model was developed, achieving average multifold accuracy of 0.871 in cross-validation on the training cohort. In the test cohort, the model achieved an AUC of 0.944 (95% CI 0.841-1.000) with accuracy of 0.842, while maintaining sensitivity of 0.889 and specificity of 0.800, demonstrating excellent and comparable performance to previous contrast-enhanced CT-based radiomics models.

Conclusion: We validated the feasibility of non-contrast CT for MSI prediction in colon cancer with radiomics analysis, highlighting its potential as a flexible and cost-effective preliminary screening tool. This approach, which does not require supplementary medical examination, may enhance clinical decision-making by providing a valuable tool for identifying MSI-H molecular subtypes in colon cancer patients.