LGG-25. ADVANCING RESPONSE ASSESSMENT OF LOW-GRADE GLIOMAS: IDENTIFYING VOLUME CORRELATIONS OF BI-PERPENDICULAR DIAMETER CHANGES WITH MACHINE LEARNING

Abstract

Abstract BACKGROUND Assessing brain tumor response is key for effective management and clinical trial benchmarks. Traditionally, bi-perpendicular diameter (BPD) measurements have been used to estimate tumor size. However, recent studies suggest volumetric assessment yields a more accurate measure of response and estimate of tumor burden. Advances in technology enable rapid volume assessments. Integrating volumetric analysis into clinical settings is challenged by the absence of volume-based standards aligned with traditional RAPNO bi-dimensional criteria. Our goal was to correlate volume change percentages with bi-perpendicular diameter changes using machine learning. METHODS In this project, we compared 2 machine learning models to establish a correlation between percentage change of sum of products of BPD and percent change in volume using a generalized linear model (GLM) and an artificial neural network (ANN). We analyzed 49 patients with low grade glioma, treated with focal irradiation (NCT04065776) (n=27) or mirdametinib (NCT04923126) (n=22). We trained and validated the models at train-test split ratios of 80/20, 75/25, and 70/30, incorporating a proportionate split of tumors from each group. The models were compared on the validation data set using mean-squared error (MSE). RESULTS At percentage changes of +25%, -25%, -50%, and -75% in the sum of BPDs, the GLM predicts changes in volume to be +33%, -31%, -59%, and -83%, respectively. For the ANN, the predicted changes are +20%, -29%, -53%, and -78%. The ANN demonstrated a lower MSE compared to the GLM. The two models showed higher concordance in their volume predictions at negative percentage changes in the sum of BPDs: -25% (p=0.15), -50% (p=0.03), and 75% (p=.08) CONCLUSION We identified volume percentage changes matching BPD changes per RAPNO criteria using two machine learning methods, both agreeing on negative values. Further testing with a larger dataset is needed to reliably correlate percentage changes.