EP23.23: Diagnosis of adnexal lesions using ultrasound‐based artificial intelligence

Adnexal lesions pose a diagnostic challenge as most are benign, and ovarian cancer is discovered only rarely but has high mortality rate. Computer-aided diagnosis models can support clinical decision-making by automating segmentation, feature extraction, and classification. The study aims to evaluate the performance of computer-extracted radiomic features merged with clinical assessment of solid elements in distinguishing benign and malignant lesions on sonographic images. Combined segmentation model of the lesion's border (see figure 1A) and unsupervised clustering of internal compartments (see figure 1B) was applied to a single-centre dataset of 133 adnexal lesions (70% training/validation; 30% test) with 26.3% malignancy prevalence (1 image/case). Eight component-based radiomic features were merged with solid elements to train a classifier to distinguish malignant and benign. AUC (proper binormal) was obtained by randomly sampling output 2000 times with replacement; sensitivity, specificity, PPV, and NPV were evaluated from sampling and calculated at threshold of the maximum Youden Index from training/validation. Merging computer-extracted component-based radiomic features with solid elements yielded AUC of 0.89 (95% CI 0.83, 0.95) in training/validation set and 0.91 (0.83, 0.98) in test set (see figure 1C). Sensitivity, specificity, PPV, and NPV were 0.95, 0.77, 0.63, and 0.43 in the test set. A combined approach of component-based radiomics with solid elements demonstrated potential in distinguishing benign and malignant adnexal lesions. Future studies will explore automating solid assessments and validating the model in different cohorts, including patients managed conservatively, to determine the added value in supporting clinical decision-making. Please note: The publisher is not responsible for the content or functionality of any supporting information supplied by the authors. Any queries (other than missing content) should be directed to the corresponding author for the article.