Abstract BACKGROUND Artificial intelligence (AI) and deep learning have had increased uses in healthcare. Initial studies have shown it is possible to predict multiple tumor types in adults and children using convolutional neural networks (CNN), a form of deep learning, using MRIs. In pediatric CNS tumors, molecular alterations are used in both diagnosis and treatment decisions. This project aims to verify that we can use deep learning to help diagnose and molecularly classify primary pediatric CNS tumors. Previously, we used Recurrent Health’s platform to predict high vs. low grade CNS lesions to >95% recall and 85% precision in pediatric patients. The objective of this study is to validate the AI’s ability to predict common genetic alterations with 90% accuracy. METHODS This retrospective study evaluated patients (ages 0-21) with histologically confirmed primary CNS tumors that underwent BRAF molecular analysis. We collected demographic data: age, tumor location, and presence of BRAF mutations/fusions. The data was divided into a training and validation set (80%/20% split). To determine the viability of models, we will measure specificity, sensitivity, and F1 Score (Sorenson-Dice coefficient) for predicting genetic alterations. Model score confidence was achieved through K-fold cross-validation and statistical power analysis with a finite population coefficient to determine the appropriate sample size. RESULTS Our retrospective test included 136 patients with multiple scans for training, model validation and testing. The CNN was able to predict presence of BRAF fusions with >95% recall/sensitivity, >76% precision/specificity and F1/accuracy >87% while predicting BRAFV600 mutations with >75% recall/sensitivity, >95% precision/specificity and F1/accuracy >85%. These metrics indicate strong discrimination in classification. CONCLUSIONS Preliminary data suggests the accuracy of the model in identifying the presence of BRAF mutations, providing a potential opportunity for a non-invasive tool to help aid in diagnosis and help guide initiation of targeted therapy.
Read full abstract