Research map / Label-efficient and robust learning / Recognition and long-tail

MosaicMRI: A Diverse Dataset and Benchmark for Raw Musculoskeletal MRI

NeurIPS 2026 · Label-efficient and robust learning › Recognition and long-tail · approach: Correspondence and causal representation learning · 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.

Introduces MosaicMRI, a 2,671-volume open raw musculoskeletal MRI dataset, and benchmarks E2E-VarNet, ESPIRiT and latent bridge matching on accelerated reconstruction, studying scaling, cross-anatomy and protocol transfer.

Problem

Deep learning MRI progress relies on brain- and knee-centric public raw datasets, so reliability and generalization of reconstruction models across diverse musculoskeletal anatomies are poorly studied.

Approach

A large, anatomically and protocol-diverse open collection of fully sampled raw MSK k-space data, used to study scaling, cross-anatomy transfer, protocol shift, and similarity-based data filtering for reconstruction.

Key results

• E2E-VarNet reaches 42.63 dB PSNR at 4x and 40.49 dB at 8x on MosaicMRI test, versus 33.59 dB and 31.37 dB for ESPIRiT.
• Scaling model cascades from 4 to 12 on the full training set moves 8x PSNR from 40.01 to 40.62 dB.
• A DreamSim-filtered subset of about 8.8k slices (roughly 15% of training) matches full-data knee PSNR (39.30 vs 39.32 dB).

Surprising findings

• LBM generative reconstruction gets only about 34 dB PSNR despite SSIM around 0.86, attributed to perception-distortion tradeoff.
• Only 30% of the DreamSim-filtered knee-query subset is knee data; spine and shoulder are most frequent.

Limitations

• Single scanner vendor and single 1.5T site (Siemens Avantofit).
• Evaluation limited to one reconstruction task and mainly one architecture (E2E-VarNet).
• Uneven anatomy distribution with long tail of small anatomy groups.

Experimental setup

Datasets: MosaicMRI fastMRI

Benchmarks: MosaicMRI fastMRI

Metrics: PSNR SSIM

Baselines: ESPIRiT LBM fastMRI-trained VarNet

Builds on: E2E-VarNet fastMRI DreamSim Latent Bridge Matching

Closest papers by meaning

• HIT: Estimating Internal Human Implicit Tissues from the Body Surface (CVPR 2024 · similarity 0.84)
• A Unified Model for Compressed Sensing MRI Across Undersampling Patterns (CVPR 2025 · similarity 0.84)
• VidSole: A Multimodal Dataset for Joint Kinetics Quantification and Disease Detection with Deep Learning (AAAI 2025 · similarity 0.84)
• AeSPa : Attention-guided Self-supervised Parallel Imaging for MRI Reconstruction (CVPR 2025 · similarity 0.84)
• MultiMorph: On-demand Atlas Construction (CVPR 2025 · similarity 0.84)
• Efficient Noise Calculation in Deep Learning-based MRI Reconstructions (ICML 2025 · similarity 0.84)