Sunday, November 30 | 11:20 a.m.-11:30 a.m. | S2-SSPD01-6 | Room E350
Session-goers will find out how a deep learning-based clinical support model that uses ultrasound imaging parameters can reduce unnecessary biopsies in pediatric head and neck patients.
This approach achieved high marks in the noninvasive diagnosis of malignant pediatric lymphadenopathy, according to findings to be presented by Gal Ben-Arie, MD, from Soroka University Medical Center in Beersheba, Israel.
The research team trained a feed-forward neural network on clinical data. It used a two-step approach that included predicting surgical intervention as a supportive task and fine-tuning for predicting malignancy. The team compared the performance of the multimodal model to that of a single-modality clinical model and an ultrasound imaging model.
The retrospective single-center study included 283 children with an age range of 1 to 17. The patients had subacute craniocervical lymphadenopathy. Of the total, 36 had malignancy confirmed on biopsy.
The multimodal model achieved an area under the curve (AUC) value of 0.91, a balanced accuracy of 85%, a sensitivity of 89%, and a specificity of 81%. It outperformed the clinical and imaging models.
On SHapley Additive exPlanations (SHAP) analysis, the researchers found that nodal size and duration of lymphadenopathy were the top predictors of malignancy. And saliency maps highlighted cortical thickening and hilar vascularity.
The researchers highlighted that this approach could make way for more timely tissue diagnosis and better outcomes in child patients.
Find out more by attending this session.