Predicting Neurocognitive Decline in Multiple Brain Metastases Patients Undergoing Distributed Stereotactic Radiosurgery.

PURPOSE/OBJECTIVE(S): Stereotactic radiosurgery (SRS) is the standard of care for treating a limited number (<3) of brain metastasis (BMs), which offers reduced neurotoxicity compared to whole brain radiotherapy (WBRT). Contemporary advancements in SRS made it possible to also commonly treat multiple (>4) BMs (mBMs). Emphasizing the value of preserving quality of life (QoL) after SRS, there is an urgent need for a systematic study of potential neurocognitive decline in patients receiving SRS treatment for mBMs. The purpose of this study is to use routine MRIs to predict neurocognitive decline for patients treated with distributed SRS, allowing for timely and effective treatment strategy design.

MATERIALS/METHODS: This study uses data from an institutional phase I/II clinical trial to determine the neurocognitive decline in patients with (>6) mBMs treated with distributed SRS. In the first 12 months post-SRS, participants are followed and evaluated with routine MRIs and the Hopkins Verbal Learning Test-Revised (HVLT-R) at 2 to 3-month intervals. Changes in HVLT-Delayed Recall scores between two visits are used to define neurocognitive decline. For each visit, an in-house deep learning model segments 66 cortical and 55 subcortical brain regions of interest (ROIs) from the T1 structural MRI and extracts 253 ROI features, including the surface area and thickness of cortical ROIs, and the volume of all ROIS. The difference in ROI features between two visits, together with other clinical factors (e.g., prescription, number of BMs, etc.), is considered as one sample. The study included 22 subjects with 91 visits, resulting in 171 samples with neurocognitive decline labels. The entire sample set is split into 10 folds on patient level for cross validation. In each fold, feature engineering is conducted to remove redundancy and to select the most-important features. The top 20% most frequently selected features are applied with Support Vector Machine to predict the neurocognitive decline label of each sample.

RESULTS: As a preliminary result, the proposed method achieves an accuracy of 76%, with an area under the curve (AUC) of 0.75, sensitivity of 0.65 and specificity of 0.83 for predicting neurocognitive decline in mBMs SRS patients using only routine T1 MRIs. The volume of lateral occipital complex, the thickness of inferior parietal lobe and postcentral gyrus, and the surface area of lateral orbitofrontal cortex and pars triangularis are identified as the 5 most important features for this task.

CONCLUSION: Our method shows promising findings for post-SRS neurocognitive decline prediction solely based on routine baseline and follow-up MRIs. In addition, it can identify critical brain ROIs associated with the post-SRS cognitive function. This method has the potential to assist treatment planning strategy to help preserve patients' QoL.