Abstract
BACKGROUND: Extrinsic adenomyosis exhibits heterogeneous clinical symptoms, with pain being more commonly reported. The relationship between magnetic resonance imaging (MRI) feature and symptom remains unclear. OBJECTIVE: To evaluate the performance of MRI radiomics model for differentiating symptom heterogeneity of extrinsic adenomyosis, pain, abnormal uterine bleeding (AUB), infertility, and no symptom. MATERIALS AND METHODS: This retrospective analysis included 405 patients with MRI-diagnosed extrinsic adenomyosis (January 2020-July 2022), randomly split 7:3 into training and test cohorts. Radiomic features were extracted from MRI-T2 image. Random forest algorithm was used to select the key radiomics features of different symptoms and develop the radiomic model by support vector machine algorithm. Multivariable logistic regression assessed clinical characteristics. A combined radiomics-clinical nomogram was created for symptom stratification. RESULTS: In total 405 patients presented with 496 clinical symptoms. In the training and test cohorts, radiomics models achieved areas under the curve (AUCs) of 0.73/0.72 (pain), 0.82/0.76 (AUB), 0.84/0.80 (infertility), and 0.80/0.71 (no symptom). The multi-signature model (radiomic + clinical features) showed improved performance, with the nomogram demonstrating good stratification ability: AUCs of 0.78/0.78 (pain), 0.87/0.85 (AUB), 0.89/0.88 (infertility), and 0.84/0.81 (no symptom) in the training/test cohort. CONCLUSION: We identified the correlation between key radiomic features and clinical symptom of extrinsic adenomyosis. The machine learning-based MRI radiomics models have potential for symptom stratification of extrinsic adenomyosis and may potentially reduce unnecessary treatment.