The question
MRI survival models trained in one cohort may not improve predictions in another. This study asks whether transfer between pediatric brain tumors and adult glioblastoma depends on its direction.
The approach
The study evaluated 115 BraTS-PEDs patients and 115 UPENN-GBM patients using FLAIR, contrast-enhanced T1, and combined T1c+FLAIR. It compared zero-shot inference, frozen-encoder transfer, and full fine-tuning with cohort-specific models trained from scratch across five patient-level folds and three random seeds.
What we know
Zero-shot transfer was limited in both directions. Pediatric-to-adult fine-tuning underperformed scratch across all inputs. For combined T1c+FLAIR, adult-to-pediatric fine-tuning showed an uncertain gain, while the estimated directional asymmetry interval included zero.
What remains open
The findings are exploratory. All evaluated asymmetry confidence intervals included zero, so directional asymmetry was not statistically established. Independent cohorts are needed to validate the observed trend.
- Does the direction of transfer change performance in independent cohorts?
- Which imaging inputs are most robust across age groups?
- When does source pretraining outperform a target-cohort model trained from scratch?
ASPNR abstract
Bidirectional Transfer of MRI-Based Survival Models Between Pediatric Brain Tumors and Adult Glioblastoma: Evidence of Direction-Dependent Performance
ASPNR 2027 (American Society for Pediatric Neuroradiology) · Accepted abstract
Purpose
To evaluate whether MRI-based survival representations transfer between pediatric brain tumors and adult glioblastoma, and whether the benefit of transfer differs by direction.
Materials and Methods
This retrospective study included 115 patients from BraTS-PEDs and 115 patients from UPENN-GBM. All pediatric survival durations represented observed events; the adult cohort included 114 events and one censored observation. FLAIR, contrast-enhanced T1-weighted imaging (T1c), and combined T1c+FLAIR were evaluated using a three-dimensional residual survival network. The encoder comprised a 32-channel stem and four residual blocks with 64, 128, 256, and 256 channels, incorporating squeeze-and-excitation attention, batch normalization, and dropout. A fully connected head generated a scalar survival-risk score. Training minimized the Breslow Cox partial-likelihood loss using learning-rate warmup followed by cosine decay. Both transfer directions were assessed using source-model zero-shot inference, a frozen source encoder with a new target head, and full-encoder fine-tuning with a new target head. Target-cohort models trained from scratch served as comparators. Five-fold patient-level cross-validation with three random seeds yielded 15 evaluations per configuration; inner validation selected learning rates and training duration. Transfer gain was defined as the target C-index difference relative to scratch. Asymmetry was defined as pediatric-to-adult gain minus adult-to-pediatric gain. Conditional 95% confidence intervals were estimated using 2,000 patient-bootstrap samples.
Results
Scratch C-indices ranged from 0.566–0.606 in pediatric patients and 0.505–0.525 in adults. Zero-shot performance was limited in both directions, with mean C-indices of 0.438–0.475. Pediatric-to-adult fine-tuning underperformed scratch across all input configurations. For T1c+FLAIR, adult performance decreased from 0.525 to 0.478, corresponding to a gain of −0.0472 (95% CI, −0.0946 to −0.0007). Conversely, adult-to-pediatric fine-tuning increased performance from 0.570 to 0.585, although this gain was uncertain (+0.0158; 95% CI, −0.0289 to 0.0565). The resulting asymmetry was −0.0630 (95% CI, −0.1235 to 0.0006). All evaluated asymmetry intervals included zero.
Conclusion
Cross-cohort pretraining did not consistently improve MRI-based survival prediction. Combined-modality fine-tuning showed a directional trend favoring adult-to-pediatric relative transfer performance, driven primarily by negative pediatric-to-adult transfer. However, bidirectional asymmetry was not statistically established. These exploratory findings highlight the need to evaluate transfer against cohort-specific scratch baselines and validate directional effects in independent cohorts.