An End-to-end Interactive Software for Pediatric Pneumonia Severity Assessment from Lung Ultrasound Images.

Purpose: To develop and validate an end-to-end interactive software that integrates lung ultrasound radiomics and machine learning for the objective assessment of pediatric pneumonia severity at the point of care. Materials and methods: This retrospective study included a dataset of lung ultrasound images from 293 pediatric patients (157 with mild pneumonia and 136 with severe pneumonia). A total of 104 radiomics features were extracted from clinician-delineated regions of interest. Feature selection was performed using Least Absolute Shrinkage and Selection Operator regression, and 10 machine learning algorithms were constructed and evaluated. The optimal model was interpreted using SHapley Additive exPlanations. The finalized classifier was integrated into an interactive software platform that guides users from image upload to severity prediction. Results: The Light Gradient Boosting Machine classifier, which forms the core of the software, demonstrated superior performance on the independent test set, achieving an accuracy of 89.8%, a sensitivity of 91.7%, a specificity of 88.6%, and an area under the curve of 90.1%. SHapley Additive exPlanation analysis identified and ranked the contribution of key predictive features, with morphological characteristics being the most influential. Conclusions: We present an end-to-end interactive software that successfully leverages lung ultrasound radiomics and machine learning to provide an objective, accurate, and rapid assessment of pediatric pneumonia severity. This tool has significant potential to standardize diagnosis and support clinical decision-making in real-world settings.