Uncertainty-Calibrated Source-Free Continual Test-Time Adaptation for Longitudinal Multiparametric MRI Differentiation of Glioblastoma Recurrence from Pseudoprogression under Missing Sequences and Scanner Drift
International Journal of Electrical Engineering, Mathematics and Computer Science · 2026
DOI: 10.62951/ijeemcs.v3i3.415 ↗Kurz gefasst
Fachartikel in der Zeitschrift International Journal of Electrical Engineering, Mathematics and Computer Science, 2026. Verfasst von Roaa Ghanim. Der Volltext ist frei zugänglich (Open Access, CC BY-SA 4.0).
Abstract (Originalsprache)
One of the major and unmet problems in the management of glioblastoma is the ability to differentiate true recurrence from treatment-induced pseudoprogression in post-treatment MRI. Despite their promising performance, their application in a wide variety of hospitals is hindered by scanner and protocol drift, by a non-uniform distribution of MRI sequences and by the time-varying nature of the target distribution. The current literature focuses on addressing different problems of recurrence/pseudoprogression classification, missing-modality learning, domain adaptation, continual test-time adaptation, and uncertainty estimation alone. We introduce UC-SFCTTA, an uncertainty-calibrated source-free continual test-time adaptation framework, which combines all of these requirements. A missing-sequence-robust multimodal encoder with modality dropout and masked attention consumes both prior and current examinations, and models the evolution of lesions; when deployed, the model adapts to unlabelled target sequences without access to source sequences, gating the input examinations with calibrated predictive uncertainty, while an exponential-moving-average teacher with stochastic weight restoration prevents forgetting. Under sequential A→B→C→A scanner drift across the multicentre domains, UC-SFCTTA achieved a target domain AUROC of 0.902, an ECE of 0.041 and prevented catastrophic forgetting of 0.003, while showing good performance even without perfusion and diffusion sequences. The framework provides a viable pathway to safer, self-sustaining recurrence evaluation in actual multicentre practice.
Zitieren
APA
Ghanim, R. (2026). Uncertainty-Calibrated Source-Free Continual Test-Time Adaptation for Longitudinal Multiparametric MRI Differentiation of Glioblastoma Recurrence from Pseudoprogression under Missing Sequences and Scanner Drift. International Journal of Electrical Engineering, Mathematics and Computer Science. https://doi.org/10.62951/ijeemcs.v3i3.415
Vancouver
Ghanim R. Uncertainty-Calibrated Source-Free Continual Test-Time Adaptation for Longitudinal Multiparametric MRI Differentiation of Glioblastoma Recurrence from Pseudoprogression under Missing Sequences and Scanner Drift. International Journal of Electrical Engineering, Mathematics and Computer Science. 2026. doi:10.62951/ijeemcs.v3i3.415
BibTeX
@article{Ghanim2026,
author = {Ghanim, Roaa},
title = {Uncertainty-Calibrated Source-Free Continual Test-Time Adaptation for Longitudinal Multiparametric MRI Differentiation of Glioblastoma Recurrence from Pseudoprogression under Missing Sequences and Scanner Drift},
journal = {International Journal of Electrical Engineering, Mathematics and Computer Science},
year = {2026},
doi = {10.62951/ijeemcs.v3i3.415}
}
EUVIMED