Journal: NPJ precision oncology
This study reports the development and validation of an artificial intelligence model, Clinical-Ovarian Multi-Task Attention (Clinical-OMTA), to classify adnexal masses as benign, borderline, or malignant using ultrasound images plus age and CA125.
Key points:
- Model design:
- Dual-backbone architecture: one branch for benign vs non-benign, another for borderline vs malignant.
- Inputs: ultrasound images, patient age, and CA125, aiming for three-class classification.
- Data and setting:
- Retrospective multicenter cohort from 23 hospitals.
- 1882 patients used for training/validation/internal testing from 21 centers.
- External testing on two independent cohorts (340 and 159 patients).
- Diagnostic performance:
- On external datasets, Clinical-OMTA achieved AUCs comparable to ADNEX and accuracy comparable to expert subjective assessment.
- However, adding age and CA125 did not improve performance over an image-only version (OMTA), indicating minimal incremental value of these clinical variables for the model.
- Generalisability:
- Performance remained stable across different ultrasound acquisition modes, scanners, scanning methods, and institutions, with accuracy in the ~80–88% range.
- Clinical impact:
- When used as a decision support tool, Clinical-OMTA substantially improved radiologists’ inter-reader agreement and diagnostic accuracy.
- The authors suggest particular value in settings with limited access to expert ultrasound examiners, such as low-resource or remote environments.
Overall, this work supports AI-based adnexal mass classification as a robust and generalisable tool that can match expert performance and enhance diagnostic consistency, though multimodal (image + clinical) input did not surpass image-only modeling in this setting.