AI-Based Quantification of Hippocampal Volume Loss after Whole-Brain Radiotherapy and Stereotactic Radiosurgery
Abstract
Purpose
Cognitive decline following brain radiotherapy is strongly associated with hippocampal injury, yet scalable methods for assessing treatment-related hippocampal structural change in clinical practice remain limited. This study evaluates hippocampal volume loss following whole-brain radiotherapy (WBRT) versus stereotactic radiosurgery (SRS) using an automated deep-learning segmentation pipeline and examines its relationship with hippocampal radiation dose.
Methods
Patients treated with WBRT (n=15) or SRS (n=15) for brain metastases underwent longitudinal T1-weighted MRI at baseline, 6 months, and 12 months post-treatment. Hippocampal segmentation was performed using FastSurfer, a convolutional neural network–based neuroimaging tool enabling rapid, reproducible volumetric analysis from standard MRI. Hippocampal volumes were extracted at each timepoint, and rates of volume change (%/month) were calculated. Segmentations were co-registered with treatment plans to derive dosimetric parameters, including maximum hippocampal dose (Dmax). Group comparisons were performed using an unpaired t-test, and dose–response relationships were evaluated using Pearson correlation analysis.
Results
WBRT patients demonstrated significantly greater hippocampal volume loss compared with SRS patients at both 6 and 12 months (p=0.0001). The average rate of hippocampal atrophy following WBRT was approximately sevenfold higher than after SRS. Across all radiotherapy patients, hippocampal volume loss was strongly correlated with hippocampal Dmax (r=−0.66, p<0.0001), indicating a dose-dependent structural effect. In contrast, cerebellar volume loss showed no significant association with cerebellar dose, supporting regional specificity.
Conclusion
Deep learning–based hippocampal segmentation using FastSurfer enables robust, high-throughput quantification of radiation-associated hippocampal atrophy from routine MRI. WBRT is associated with substantially greater hippocampal volume loss than SRS, with magnitude tightly linked to hippocampal dose. These findings support hippocampal sparing and dose-aware treatment selection, and demonstrate the value of AI-enabled morphometric biomarkers for advancing treatment planning, toxicity assessment, and outcome modeling in contemporary medical physics.