Accurate classification of ovarian adnexal masses is important for clinical decision-making, but ultrasound interpretation can be difficult when lesions have complex internal structures. A deep learning framework was developed to improve this process using routinely acquired ultrasound images. Built on data from 230 adults with adnexal masses, the approach combines automated segmentation, radiomic analysis and multimodal classification. It separately analyses fluid and solid components of each mass and also includes an explainability feature that retrieves similar prior cases.
Segmenting Masses and Their Internal Components
The framework is built around two core stages: segmentation and classification. Segmentation is handled by an nnU-Net-based model that automatically delineates adnexal masses and separates their fluid and solid components from B-mode ultrasound images. This design reflects the view that important diagnostic information is not confined to the lesion as a whole but also lies in its internal composition. By isolating these components, the system can analyse structures that may carry distinct diagnostic value.
After segmentation, the model defines the adnexal mass region and prepares the image data for downstream classification. The classification framework then combines deep learning-derived image features with radiomic features extracted from the ultrasound data. The image branch uses an ensemble architecture incorporating ResNet18 and EfficientNetB2S. Instead of treating the full image content as a single input stream, the framework uses separate ensemble networks for B-mode images, fluid regions and solid regions. This arrangement reduces feature space complexity and allows each network to learn from a more specific representation of lesion morphology.
The radiomic branch complements the image models by using multiple spatial filters and a feature-wise gated multilayer perceptron. These radiomic features are integrated with the deep learning features within a multimodal network-based classifier. The resulting structure combines visual representation learning with tabular feature analysis in a single classification pipeline.
Multimodal Design Improves Classification Results
The framework achieved strong performance in distinguishing benign from malignant adnexal masses. At image level, it reached 90% accuracy and an AUC of 0.94. Reported performance also included specificity of 0.92, precision of 0.90, sensitivity of 0.90 and an F1-score of 0.90. These results exceeded those of the radiomic models and standard deep learning models evaluated in the same analysis. The framework was also described as being on par with ADNEX, O-RADS 2019, O-RADS 2022 and the two-step approach.
Must Read:Gastric Contrast Ultrasound in Tumour Assessment
Part of this performance is attributed to gain to the use of separate networks for B-mode, fluid and solid components. This structure improved performance by reducing feature space complexity and optimising learning for each component. Rather than relying on a single representation of the lesion, the multimodal design captures different characteristics from the whole mass and from its internal substructures. The classifier therefore uses complementary forms of information that are processed in a coordinated way.
At patient level, the framework achieved 91% accuracy and 92% AUC. The abstract compares these figures with 77% accuracy and 92% AUC for ADNEX, 84% accuracy and 89% AUC for O-RADS 2019, and 80% accuracy and 88% AUC for O-RADS 2022. These comparisons place the framework in a favourable position against established tools while maintaining a workflow based on routinely acquired ultrasound images. The reported results support the value of combining segmentation, substructure analysis, radiomics and deep learning within the same model.
Explainability Adds a Decision-Guiding Layer
Beyond classification accuracy, the framework adds an explainability pipeline intended to improve interpretability. One part of this pipeline visualises differentiation between benign and malignant samples using principal component analysis and t-distributed stochastic neighbour embedding. These methods provide a visual representation of how samples cluster within the learned feature space. The framework also includes a decision-guiding tool that identifies the top-k training samples most similar to a test case on the basis of cosine similarity. This retrieval approach is designed to add an extra layer of interpretability and to increase confidence in decision-making.
The explainability component is closely tied to the stated clinical purpose of the framework. Rather than generating a malignancy prediction alone, the system is intended to provide predictions together with visualisations of similar historical cases. This combination supports a more transparent use of AI assistance in adnexal mass assessment. The model therefore addresses classification and interpretability in parallel.
The overall contribution rests on several linked elements: a curated cohort of adults with ultrasound-detected adnexal lesions, automated segmentation of masses and their fluid and solid components, a multimodal classifier combining deep learning and radiomic features, and an explainability framework based on visualisation and similarity retrieval. Together, these elements form a decision-support approach for ultrasound-based classification of adnexal masses.
This framework combines automated segmentation, multimodal classification and similarity-based explainability to support ultrasound assessment of ovarian adnexal masses. It analyses B-mode images alongside fluid and solid components and integrates deep learning-derived features with radiomic features from multiple spatial filters. Reported performance was strong at both image and patient level, and the model compared favourably with established tools including ADNEX and O-RADS. The framework also adds a decision-guiding layer by retrieving similar prior cases and visualising sample differentiation in feature space. Taken together, these elements position the model as a structured approach to more accurate and more interpretable classification using routinely acquired ultrasound images.
Source: Journal of Imaging Informatics in Medicine
Image Credit: iStock
References:
Dhar MK, De Vitis L, Gregory AV et al. (2026) A Deep Learning Framework for Enhanced Ovarian Adnexal Mass Classification Using Routinely Acquired Ultrasound Images. J Digit Imaging Inform Med: In Press.
