Addressing Domain Shift in Automated Brain Tumor Segmentation: Evaluating TumorSynth on a Latin American Clinical Data.

The translation of deep learning models for brain tumor segmentation to public healthcare systems is hindered by domain shift, particularly in heterogeneous clinical cohorts. This study presents a zero-shot analysis of a promising model TumorSynth using local data. We evaluated these methodologies on the local FLENI dataset for segmentation accuracy and the public LUMIERE dataset for automated Response Assessment in Neuro-Oncology (RANO) classification. Results on the FLENI dataset indicate that the model achieved a median Dice coefficient of 0.90 for the whole tumor. Conversely, the contrast-agnostic model successfully identified active tumor regions and edema but encountered significant difficulties in accurately resolving necrotic and non-enhancing with median Dice score of 0.60. For longitudinal monitoring, both automated volumetric approaches exhibited significant limitations in RANO classification, frequently misclassifying progressive disease as stable disease and yielding F1-scores near random chance levels. Furthermore, they underscore that relying solely on automated volumetrics remains insufficient for robust clinical-level disease progression assessment.