IMG-130. Integrating AI-based models from physiologic MRI with MR spectra for differentiating treatment-effect from glioma tumor recurrence

INTRODUCTION An ongoing challenge faced in neuro-oncology is non-invasively distinguishing treatment-induced effects (TxE) following chemotherapy and/or radiation therapy from true tumor recurrence (rTumor). Previous research has explored the utility of AI-based models for this task but has overlooked within-lesion heterogeneity and the non-enhancing, T2-lesion. We investigated the value of integrating models trained using: 1) multi-parametric MRI (mpMRI) including anatomical, diffusion-weighted, and perfusion-weighted images, and 2) individual spectra from 1cc regions surrounding the tissue-sample locations, to improve the discrimination of treatment-effect from tumor recurrence. METHODS This retrospective study included 144 high-grade glioma patients who underwent MRI scans before surgical resection for suspected recurrence. Imaging included standard anatomical, diffusion-weighted, and dynamic-susceptibility-contrast perfusion-weighted MRI, and lactate-edited ¹H-MRSI. 324 spatially-localized tissue-samples were histopathologically classified as TxE or rTumor. The mpMRI model utilized 10 mm volumetric patches of each standardized image contrast (T2-FLAIR, T1-post-contrast, peak height and %-recovery from perfusion, ADC and FA from diffusion) centered on the tissue sample coordinates and generated 20 ensemble predictions. The AI-based spectra model predicted Ki-67, cellularity, and a composite tumor aggressiveness index from the entire 1D-spectrum reconstructed at the location of the tissue-sample. These predictions were concatenated into 4 machine-learning classifiers, which were trained on the combined feature set and on individual imaging and spectral features to assess model contributions. RESULTS Balancing the dataset enhanced the performance of all models, most notably that of gradient boosting (AU-ROC: 0.655 to 0.724). The radiopathomic spectra model increased performance of the weighted ensemble by 4% to 0.73+/-0.04 AU-ROC when integrated with our previously-developed mpMRI model. Three imaging features consistently ranked among the top five predictors across classifiers. CONCLUSION Integrating radiopathomic-derived features from an AI-based model using the entire spectrum with mpMRI features preserved diagnostic performance across all models, with a maximum improvement of 5.26% in AU-ROC. Current work is evaluating different strategies for combining models for contrast-enhancing and non-enhancing samples separately.