https://atlas.rsna.org/schemas/2025-11/model.json

MIMIR (Medical Inference on MRI with Image-based Regression)

https://github.com/tarolangner/ukb_mimir

1.0

Taro Langner, email: es.uu.icsgrus@rengnal.orat

Supported by the Swedish Heart-Lung Foundation and the Swedish Research Council (2016-01040, 2019-04756, 2020-0500, 2021-70492) and conducted under UK Biobank application no. 14237.

Work covered by a UK Biobank Research Tissue Bank approval and separate approval from the responsible Swedish ethics committee.

Convolutional Neural Network (ResNet50) performing mean-variance deep regression on 2D projections of neck-to-knee MRI; trained in PyTorch with Adam, batch size 32, random translations; learning rate 5e-5 for 8000 iterations then 5e-6 for 2000 iterations; four multi-output modules.

Open-source implementation on GitHub: https://github.com/tarolangner/ukb_mimir

Fast, automated estimation of body composition metrics and organ volumes to support large-scale medical research using UK Biobank MRI data; provides calibrated prediction intervals.

Fully automated estimation pipeline for target measurements from MRI scans.

Automated inference of 72 measurements (e.g., body composition metrics, organ volumes, anthropometrics, selected health-related properties) from UK Biobank neck-to-knee 1.5-T body MRI scans and from other studies reproducing a similar protocol and demographics; intended for research use.

Neck-to-knee 1.5-T two-point Dixon water/fat body MRI; images compressed to 2D projected representations.

Process UK Biobank neck-to-knee two-point Dixon MRI with the provided pipeline to output point estimates and prediction intervals for 72 targets. Uncertainty estimates calibrated via post hoc scaling factors. On RTX 2080 Ti (11 GB), ~1000 participants processed in ~10 minutes.

Not intended for medical diagnostics; performance for diagnosing health states (e.g., type 2 diabetes) is limited. Generalization likely restricted to UK Biobank imaging protocol, device type, and demographics; out-of-domain tasks may require retraining with hundreds of participants. Some measurements may be lost due to preprocessing or absent from MRI. 2D projection limits performance on small structures (e.g., higher error for kidney volume vs axial segmentation). Alternative dedicated protocols may be superior for specific targets (e.g., liver fat).

Use for large-scale research to automatically derive phenotypes from UK Biobank neck-to-knee MRI; not for clinical diagnosis or decision-making.

10-fold stratified cross-validation; post hoc uncertainty calibration; results replicated on subsequently released UKB test data for 12 body composition targets; open-source code provided.

Approximately 10 minutes to process 1000 participants on an NVIDIA RTX 2080 Ti (11 GB).