Research map / Label-efficient and robust learning / Domain adaptation and generalization

Knowledge Transfer Scaling Laws for 3D Medical Imaging

NeurIPS 2026 · Label-efficient and robust learning › Domain adaptation and generalization · approach: Cross-modality medical segmentation adaptation · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes a transfer-aware scaling law with domain-specific power-law exponents and an asymmetric cross-domain transfer matrix to optimize data allocation for MAE pretraining across six 3D CT/MRI/PET domains.

Problem

Choosing how much of each 3D medical imaging domain (CT, MRI, PET) to use under a fixed pretraining token budget is typically heuristic (uniform or data-proportional), ignoring heterogeneous scaling and asymmetric cross-domain transfer.

Approach

Model per-domain MAE loss as a power law in effective budget, where effective budget is a sum of source allocations weighted by a directed transfer matrix tau_ij; fit parameters from K^2 small pure-domain proxy runs and optimize the mixture in closed form with SLSQP, yielding compute-dependent allocations with hub-and-island structure.

Key results

• Transfer-Aware allocation reduces mean MAE at B=50K to 0.0451 vs 0.0712 for Data-Proportional (up to 58% relative improvement).
• Scaling law fit on 10K-100K extrapolates to 200K with predicted 0.0227 vs observed 0.0224 mean loss; per-domain r=0.989.
• Transfer-Aware improves downstream segmentation Dice (e.g., HECKTOR 67.8 to 71.1, BraTS T2 61.4 to 65.4) and classification (SynBrain-PET AUC 94.5 to 95.9) over Data-Proportional.

Surprising findings

• Transfer-Aware allocation gives HN-PET (island) 56.3% of budget yet also improves floor domains via incoming transfer.
• Relative advantage over data-proportional narrows at B=200K, suggesting proxy model nears loss floor or capacity limit.

Limitations

• Assumes linear composition of transfer contributions, missing synergistic or interfering interactions between sources.
• Transfer matrix estimated once from small proxy runs and held constant across budgets; fixed model size (12M ViT).
• Optimizes MAE reconstruction proxy rather than clinical outcomes; advantage narrows at largest budget.

Experimental setup

Datasets: AMOS 2022 BraTS 2023 HECKTOR 2025 BHSD 2024 OASIS-1 SynBrain-PET OrganMNIST3D CPAISD

Benchmarks: AMOS 2022 BraTS 2023 HECKTOR 2025 BHSD OASIS-1 SynBrain-PET OrganMNIST3D

Metrics: MAE reconstruction loss Dice accuracy AUC Pearson correlation

Baselines: Uniform Data-Proportional Transfer-Naive From scratch Single-Modality only

Builds on: neural scaling laws data mixing laws RegMix UNETR masked autoencoder

Closest papers by meaning

• How Well Do Supervised 3D Models Transfer to Medical Imaging Tasks? (ICLR 2024 · similarity 0.90)
• Perturbating, Tuning, and Collaborating: Harnessing Vision Foundation Models for Single Domain Generalization on Medical Imaging (AAAI 2025 · similarity 0.87)
• Training Like a Medical Resident: Context-Prior Learning Toward Universal Medical Image Segmentation (CVPR 2024 · similarity 0.87)
• Autoregressive Sequence Modeling for 3D Medical Image Representation (AAAI 2025 · similarity 0.86)
• Better Tokens for Better 3D: Advancing Vision-Language Modeling in 3D Medical Imaging (NeurIPS 2025 · similarity 0.86)
• Learning General-purpose Biomedical Volume Representations using Randomized Synthesis (ICLR 2025 · similarity 0.86)