Timepoint-Specific Benchmarking of Deep Learning Models for Glioblastoma Follow-Up MRI.

Area of Science:

• Neuro-oncology
• Medical Imaging
• Artificial Intelligence

Background:

• Distinguishing true tumor progression (TP) from treatment-related pseudoprogression (PsP) in glioblastoma is a significant clinical challenge, particularly in early follow-up scans.
• Accurate differentiation is crucial for timely treatment adjustments and improved patient outcomes.

Purpose of the Study:

• To benchmark the performance of various deep learning (DL) architectures for differentiating TP from PsP using follow-up MRI scans.
• To assess the impact of imaging timepoint on the diagnostic accuracy of DL models.

Main Methods:

• Cross-sectional benchmarking of eleven DL model families (CNNs, LSTMs, hybrids, transformers, selective state-space models) on the Burdenko GBM Progression cohort (n=180).
• Models were trained using a unified, quality-controlled pipeline with patient-level cross-validation, analyzing different post-radiation therapy (RT) timepoints independently.
• Evaluation focused on accuracy, F1 score, and Area Under the Curve (AUC) to assess discrimination capabilities.

Main Results:

• Overall accuracies were comparable across timepoints (~0.70-0.74), but discrimination improved at the second follow-up for several models.
• A Mamba+CNN hybrid model demonstrated the best accuracy-efficiency trade-off.
• Transformer variants achieved competitive AUCs but with higher computational costs; lightweight CNNs were efficient but less reliable. Model performance was sensitive to batch size.

Conclusions:

• The study establishes a timepoint-aware benchmark for DL models in glioblastoma progression assessment.
• Findings suggest that incorporating longitudinal data, multi-sequence MRI, and larger multi-center cohorts may enhance diagnostic performance.
• Further research is motivated to improve the modest absolute discrimination observed, highlighting the inherent difficulty of TP vs. PsP differentiation.