henhaobio
Sign in

bioRxivNeurosciencePreprint

Automatic Quality Control and Error Correction in MRI linear registration via a Residual Parameter Prediction Network for T1w MRI

Zhaojin CheniD, Roqaie MoqadamiD, Amelie MetziD, Walter Adame‐GonzaleziD, The Alzheimer’s Disease Neuroimaging Initiative (ADNI), Yashar ZeighamiiD, Mahsa DadariD

PDFDOICode
In-depth full-text digest available · Jump to digest ↓

Abstract

Errors in linear registration can propagate to downstream nonlinear registration and bias volumetric estimations, deformation-based morphometry (DBM) and voxel-based morphometry (VBM) analyses. Subtle linear registration errors are particularly challenging as they are difficult to detect and may not result in obvious failures in nonlinear registration but still affect downstream results. Therefore, accurate identification and correction of these errors are critical. In this study, we present the Residual Affine COefficient Optimization Network (RACOON), a framework designed to identify and correct linear registration errors in T1w MRI scans registered to the MNI-ICBM152 space. RACOON's correction module achieved a residual misalignment RMSE of 0.778 mm on synthetic dataset, comparable to the variability observed among repeated QC-passed registrations using the same pipeline. For the classification module, RACOON achieved a balanced accuracy of 76.8% and a precision of 74.4%, outperforming existing state-of-the-art methods. RACOON is open source and publicly available at https://github.com/ZhaojinChen/RACOON.

In-depth summary

Claude Opus 5.5 · xhigh effort · 8 Nov 2026 · from the full text of version 1 (16 Sept 2026) · CC BY 4.0

A 3D ResNet18 trained on synthetically misaligned T1w scans predicts residual affine parameters, correcting synthetic misalignments to 0.778 mm residual RMSE and flagging failed registrations better than RegQCNET or DARQ.

Errors in linear registration, including subtle ones that nonlinear registration only partly compensates, bias volumetric, deformation-based and voxel-based morphometry results, and manual quality control (QC) takes about 30 hours for roughly 10K registrations. Prior automated QC, per the authors, mostly uses preprocessed images, ignores defacing, targets large errors and only flags failures; RACOON runs on unpreprocessed, possibly defaced scans and also corrects failed transforms.

Key results

  1. ShownOn held-out datasets with synthetic misalignments, the residual-parameter network leaves less than 1 mm of residual error. Iterative inference lowers it to about the level of disagreement between two QC-passed registrations of the same scans.

    Model trained on 13,000 QC-passed scans: RMSE over six brain-boundary points was 0.861 ± 0.299 mm on the five independent test datasets in a single pass and 0.778 ± 0.286 mm with iterative inference (t=-8.36, p<0.001), vs 0.786 mm between two QC-passed registration runs in ADNI (the most variable of three datasets). NIFD, with the largest parameter MAE: 0.92 ± 0.35 mm. · 3.1.2 Training sample size and final model…; 3.1 Variability of real QC-passed images; 2.6 Assessing the variability of QC-passed…; 4. Discussion

  2. ShownRACOON-C (fine-tuned from the residual-parameter network) had the highest accuracy and precision of all methods tested; RACOON-P (SVM on predicted parameters) ranked second in accuracy, roughly on par with direct classification and ahead of DARQ and RegQCNET. Authors suggest many real failures were subtle.

    Five-fold averages on the held-out 10% of Dataset C (the Table 4 caption calls it Dataset D): accuracy 0.767 for RACOON-C and 0.721 for RACOON-P, ahead of direct classification, DARQ and RegQCNET; RACOON-C precision 0.744. Abstract: balanced accuracy 76.8%. Gains over direct classification of about 4.8 points (balanced accuracy) and 6.8 points (precision). · 3.3 Classification Module; 4. Discussion; 2.5 Performance Evaluation; Abstract

  3. ShownDefacing augmentation removed the accuracy drop seen when the network was applied to a defaced independent dataset (CamCAN).

    Without defacing augmentation, RMSE rose from 1.39 ± 0.60 mm on validation to 1.69 ± 0.56 mm on defaced CamCAN (t=-11.68, p<0.001). With augmentation (nine facial masks, four manual and five from AFNI, FreeSurfer, FSL, pydeface, mridefacer), RMSEs were 1.04 ± 0.37 mm and 1.03 ± 0.31 mm, not significantly different. · 4. Discussion; 2.4.2 Augmentation

  4. SuggestedThe predicted residual parameters fixed real ANTs registration failures. In the example shown, a corrected scan passed visual QC performed blind to correction status.

    Qualitative examples only (Fig. 7). One scan had a scaling error in all three views and a rotation error clearest in axial slices; after correction its agreement with the template outline improved in every view. No count or rate of corrected real failures is given. · 3.4 Correction Module

  5. SuggestedPrediction errors show the same axis dependence as the variability between repeated registrations: more rotation error about x, less scaling and translation error along x. This hints that part of the residual error is intrinsic to affine registration rather than a model bias.

    Synthetic test sets: x-axis rotation MAE 0.305° to 0.336° vs about 0.19° to 0.23° for y and z; x-axis translation MAE 0.091 to 0.139 mm vs about 0.20 to 0.26 mm for y and z. Repeated QC-passed registrations in ADNI, UKBB and PPMI showed the same pattern. · 3.1.2 Training sample size and final model…; 3.1 Variability of real QC-passed images; 4. Discussion

Methods, data and reproducibility

Rotation, scaling and translation offsets were applied to QC-passed T1w scans registered to MNI-ICBM152-2009c; a 3D ResNet18 learned to predict these residual affine parameters from unpreprocessed volumes. Predicted parameters feed an SVM pass/fail classifier (RACOON-P) and an iterative correction of the original transform; a second classifier (RACOON-C) is fine-tuned from the network's backbone on the 3D volume.

  • T1w MRI from 13 public datasets (incl. ADNI, UK Biobank, PPMI, CamCAN); Siemens, GE, Philips; 1.5 T and 3.0 T; healthy aging, frontotemporal dementia, Parkinson's and Alzheimer's disease · human (clinical cohorts such as ADNI, UK Biobank) · n = 13 datasets
  • Dataset B: QC-passed scans with synthetic affine misalignment for the residual-parameter network; five datasets held out as independent tests, eight form the training pool · human · n = 13,000 training scans (final model)
  • Dataset A: subset for the template-input ablation and the scaling/translation loss-weight grid search · human
  • Dataset C: equal numbers of real QC-passed and QC-failed scans with manual labels; 10% held out for testing, 90% for five-fold cross-validation · human
  • Repeated linear registrations (two runs, both QC-passed) used to estimate acceptable affine variability · human · n = 3 datasets (ADNI, UKBB, PPMI)
  • CamCAN, a defaced dataset, used to test the effect of defacing augmentation · human
Methods & tools
3D ResNet18, SVM (RBF kernel), ANTs, PELICAN, MNI-ICBM152-2009c template, PyTorch, TorchIO, SimpleITK, AFNI, FreeSurfer, FSL, pydeface, mridefacer (defacing masks), RegQCNET, DARQ
Software & environment
PyTorch with automatic mixed precision (torch.cuda.amp.autocast) and TorchIO; no package versions, requirement files or containers stated.
Assay conditions & compute
Computing resources from the Digital Research Alliance of Canada; hardware, run time and memory not stated.
Reproducibility gaps
Trained models are released with the code. Dataset descriptions and the compositions and sizes of subsets A, B and C are deferred to Supplementary material. No data URLs or accessions are given in the text, and the release of the manual QC labels is not mentioned.
Limitations (6)
  • Real QC-failed registrations came mainly from ANTs, so the range of real failure types may be narrow. The authors argue that the synthetic parameter sampling is tool-agnostic, but suggest adding failures from other registration methods. · stated by the authors
  • Shearing was not among the training transformations, so registration failures caused by shearing are not corrected (Fig. S8). · stated by the authors
  • Only linear registration is checked; QC of nonlinear registration, skull stripping and tissue segmentation is left for future work. · stated by the authors
  • Developed and validated for MNI-ICBM152-2009c; templates whose stereotaxic alignment differs substantially would need retraining. · stated by the authors
  • Correction of real failures is shown only through example scans (Fig. 7), and no proportion of real QC failures that pass after correction is reported. The headline correction accuracy therefore rests on synthetic misalignments. · noted in this summary
  • The classification comparison uses one held-out split (10% of Dataset C) labelled with the authors' own QC criteria, and the test-set size is not given in the text. DARQ and RegQCNET were run as released models, with RegQCNET's threshold set by ROC analysis. · noted in this summary

Read this if you register large T1w cohorts linearly to MNI space and want automated QC that can also repair failed affine transforms instead of discarding the scans.

Citations
0

Reproducibility ledger

Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.

Sign in to file a reproducibility report.

Related papers

Similarity Map
SelectedSame fieldCross-field
THISTummala '21Fonov '18Fonov '21Lange '24Ardekani '22Hsu '21
66% semantic matchNeuroscience
Sudhakar Tummala, Niels K. Focke·IEEE International Symposium on Biomedical Imaging (ISBI)·Apr 13, 2021· 5 citations