Objective:

To establish a clinical-radiomics prediction model for the differentiation degree of pancreatic cancer

and verify its performance in an independent cohort.

Methods:

Fifty-six patients diagnosed with pancreatic cancer by pathological examination were collected. Based on MaZda

whole-lesion region of interest (ROI) was placed to extract radiomics features

and a total of 30 radiomics features were extracted using three extraction methods

and dupli

cate radiomics features were removed. Then the high-collinearity features were removed from the radiomics feature detection. According to the normality test

independent sample t test and Mann-Whitney U test were performed

respectively. Then

combined with the clinical indicator carbohydrate antigen (CA)19-9

a clinical-image prediction model was constructed.

Results:

Teta2 and S(1

0) Entropy differed significantly between highly and moderately-poorly differentiated groups

and area under curve (AUC) was 0.68 and 0.70

respectively. The AUC obtained by the combination of Teta2 and S(1

0) Entropy was 0.74. The AUC of the clinical-radiomics mode which integrated CA19-9 was 0.815. In the validation group

the clinical-radiomics model also achieved good diagnostic performance (AUC=0.78).

Conclusion:

The clinical-radiomics prediction model based on texture features and clinical classic indicators might be helpful for predicting the differentiation degree of pancreatic cancer.