EP22.03: Integrating ultrasound omics data and genomic information for early diagnosis of cervical cancer

The accurate diagnosis of early cervical cancer and cervical precancerous lesions continues to present challenges. This study aimed to create a machine learning model that could correlate ultrasound omics data with CC and Cervical intraepithelial neoplasia. By integrating CIN susceptibility-associated SNPs genomic data with clinicopathological features, the goal was to improve the differentiation of HSIL and stage IA CC in diagnosis. From January 2014 to February 2020, 96 patients with pathologically confirmed HSIL and 28 patients with stage IA CC were retrospectively included. NUK software was used to delineate the ROI and extract features from the collected cervical ultrasound images. Based on GWAS reports and previously published literature, 19 SNPs loci of 10 genes associated with tumour susceptibility were screened. Genotyping was performed using multiple PCR combined with next generation sequencing. The model was established using ultrasound image features, SNP data and clinicopathological data of patients. All subjects were randomly divided into a training set and a verification set in a 7:3 ratio. MRMR algorithm, LASSO and support vector machine method are used for feature selection and prediction modelling. The performance of the differential HSIL and early CC models was evaluated for calibration and clinical application. For the differentiation of HSIL and early CC models, the method combining ultrasonography, clinicopathological data and SNP data showed higher classification efficiency, with an AUC value of 0.945 [95%CI: 0.881-1.000] in the training cohort and 0.850 [95%CI: 0.713-0.987]in validation cohort, than the performance with only the ultrasound omics model (AUC: training, 0.902 [95%CI: 0.836-0.967]; validation, 0.885 [95%CI: 0.751-1.000]). The model developed and constructed in this study, which combines ultrasonic imaging features with clinicopathological features and SNP genomic data, has the potential to distinguish HSIL from Stage IA CC in the early stage, and has good clinical application value.