Objective:

To explore the predictive value of radiomics models based on endorectal ultrasound (ERUS) images for KRAS gene mutations in rectal cancer patients.

Methods:

A total of 225 patients with rectal cancer admitted to Fujian Medical University Union Ho

spital were included in this study. Those patients were randomly separated into training cohort (191 cases) and testing cohort (34 cases) according to the ratio of 17∶3. The ERUS images of each patient’s clearest and deepest tumor infiltration section were selected for manual segmentation and feature extraction. After dimensionality reduction and selection

three classification algorithms

logistic regression (LR)

support vector machine (SVM) and random forest (RF) were used to construct models to predict the

KRAS

gene status of rectal cancer. Receiver operating characteristic (ROC) curve

calibration curve and decision curve analysis (DCA) were drawn to evaluated the predictive performance

goodness of fit and clinical value of the models

respectively. DeLong test was used to compare the efficiency differences of the three models.

Results:

After features selection

the best 8 features were used to construct models for predicting

KRAS

gene mutations in rectal cancer patients. The area under the curve (AUC) of SVM

LR and RF models were 0.94

0.93

0.91 and 0.82

0.88

0.85 in the training cohort and testing cohort

respectively. There was no significant difference in the AUC of the three models ( all

P

＞0.05 ) by DeLong test. The DCA showed that all three models had certain clinical benefits

and the LR model had the highest clinical benefit in the testing cohort. The calibration curve showed that the three models fitted well.

Conclusion:

The radiomics models of ERUS images have great predictive value for

KRAS

gene mutations in rectal cancer. It can be used as a supplementary method for non-invasive evaluation of

KRAS

gene mutations in rectal cancer patients and has great guiding significance for clinical selection of targeted therapy.