Slice distortion is absorbed rather than modelled, and stereotactic error goes unmeasured in deep MRI–histology registration

Question: How do deep-learning-based nonlinear registration methods for MRI-to-histology alignment handle slice-level tissue deformation and evaluate accuracy against stereotactic targets?

Answer

Across the retrieved literature, deep-learning MRI-to-histology pipelines do not model slice-level deformation as a physical, section-specific process; they absorb it into a generic dense 2D/3D deformation field, or they avoid it by acquiring an intermediary that carries the section through undistorted. The strongest histological resources in this corpus achieve metric accuracy by controlling deformation at acquisition — thick sections, blockface volumes, tape-transfer serial sectioning — and then handing a residual nonlinear refinement to registration, whether learned or classical [6] [2] [14]. Where deep networks do contribute, their role is simulated distortion priors and appearance disentanglement, and their headline gains are reported on histology-to-histology benchmarks rather than MRI-to-histology [10] [13]. Crucially, no retrieved paper evaluates a learned MRI–histology transform directly against stereotactic target coordinates: stereotactic address is assigned downstream by MRI-to-MRI normalisation, and accuracy is reported as Dice, Hausdorff distance, or landmark rTRE — quantities that the corpus itself shows can improve while target-scale geometric error does not [6] [4]. Taken together, the papers support a position that the field has solved the alignment problem at the resolution of anatomy and left the targeting problem at the resolution of surgery unaddressed.

Slice-level deformation enters the pipeline as an unmodelled nuisance absorbed by flexible warps

Coregistration is described as difficult because histological processing and sectioning impose microscopic and macroscopic nonlinear 3D deformation on the tissue, and pipelines respond by fitting deformations rather than by attributing them to sectioning [2]. Methodological reviews of whole-slide registration name structural and morphological change between non-consecutive sections, artefacts, tears and deformations as the defining obstacles, and catalogue deep architectures without a dedicated deformation model for sectioning [1]. The one framework that trains a CNN for cross-modality histological brain-section registration does so entirely on simulated data, using simulation to cover the appearance and geometry gap rather than to represent a measured sectioning process [10]. Deformation is handled elsewhere by region-specific regularisation; segmentation-driven registration decomposes images into anatomically coherent subregions and computes partial fields per region, with accuracy tracking segmentation quality near-linearly [17]. In a low-field MRI study, smoothness and Dice remained high while real-data accuracy dropped due to geometric distortion, indicating that flexible fields absorb distortions the metrics do not see [9].

The highest-accuracy histological pipelines control deformation physically, not statistically

The SWAN atlas reduced slice-level deformation by sectioning an entire human brain at high thickness, explicitly because thick sections are less prone to it, and then normalised to stereotactic space through a multistep transform chain rather than a single learned warp [6]. Blockface imaging is used as an undistorted intermediary volume, with the reported result of approximately 90% Dice for whole-tissue MRI-to-blockface alignment and above 95% for 2D MRI–histology registration [2]. Tape-transfer assisted serial sectioning yields contiguous series that are co-registered with MRI into reference volumes at 60 µm isotropic with a stereotactic coordinate system based on skull landmarks [14]. Rodent atlases report the same pattern, reassembling Nissl and myelin series into 20 µm isotropic volumes in stereotactic coordinates derived from co-registered in vivo MRI and CT [15]. Blockface-reconstructed 3D histology similarly underpins a nigrosome atlas co-aligned to MNI152 and validated against BigBrain rather than against a stereotactic target [16].

Deep networks contribute simulated distortion priors and appearance disentanglement, but the benchmarks are histology-to-histology

Learned approaches are motivated by the absence of large annotated datasets, which simulated training data is used to replace [10]. Macaque whole-brain registration uses deep models together with each specimen's own MRI to handle cross-modal contrast differences and the more pronounced non-uniform deformation of larger ex vivo specimens [5]. Stain-induced appearance shift is treated as the primary obstacle, with scene-appearance disentanglement achieving a median rTRE of 0.25% on ANHIR against 0.27% for the state-of-the-art comparator [13]. Detector-based and detector-free deep features with outlier rejection reach average-average rTRE of 0.0034 on ANHIR, 17% better than traditional approaches [12]. These numbers establish sub-percent in-plane accuracy for histology-to-histology alignment and say nothing about out-of-plane section correspondence or MRI.

Stereotactic address is inherited from a downstream MRI-to-MRI step, never validated end-to-end

Atlas resources state the workflow plainly: users transform the atlas into patient space by MRI-to-MRI registration, with nonlinear refinement left to SwanWarp or WarpDrive [6]. Rodent and primate tools follow the same architecture, with an atlas-alignment step followed by nonlinear refinement in WebWarp or WarpDrive [21]. Atlas-alignment software for microdissection exports region contours from atlas-registered sections without reporting target-space error at all [20]. Histology-derived segmentation tools are validated by Dice against the original Bayesian framework across in vivo MRI, ex vivo MRI, HiP-CT and photography, not by target displacement [23]. Deep-learning atlas registration for pathological MRI reports DSC 0.89–0.92, HD 6.79–7.60 mm and ASSD 0.63–0.77 mm with Jacobian plausibility checks, a template the histology field has not matched with target-level metrics [3].

Overlap and appearance metrics can improve while target-scale error does not

The registration accuracy available to stereotactic practice is millimetric and spatially structured: anatomical-fiducial registration error to MNI ranged 1.49–6.85 mm across brain regions, and misregistration accounted for 28% of variance in DBS electrode position along the axis of maximum variance, a median of 0.64 mm with a range of 0.05–2.05 mm [18]. Clinical thalamic lead placement achieves median proximity of 0.30 mm for stereo-EEG and 0.23 mm for DBS to atlas nucleus volumes [22], and even an MRI-free scalp-measurement method reaches mean localisation error of 3.82 mm across ten deep targets [19]. Against this budget, flexible transformation models are known to fit intensity features and noise so that normalised cross-correlation rises without genuine geometric gain, motivating landmark-derived target registration error as a complement [4]. Fitting error is characterised as necessary but not sufficient for deformable correction because the fitting parameters themselves can absorb poor registration [8]. Registration-induced residual misalignment further acts as structured label noise; networks partly learn the geometric convention of the pipeline that produced their targets, and voxel-wise metrics reward it, with perceptual and voxel-wise metrics disagreeing precisely when alignment is unreliable [11].

Conflicts and gaps

The corpus disagrees on whether deep learning is the mechanism of improvement at all: the highest-overlap MRI–histology results come from blockface-intermediated multistep chains that are not presented as learned end-to-end registration [2], while the competitive rTRE figures come from ANHIR multi-stain histology pairs with no MRI in the loop [13] [12]. Dice above 95% for 2D MRI–histology and registration errors of 1.49–6.85 mm in stereotactic space are not contradictory but are measured in different spaces and are not comparable, and no retrieved paper bridges them [2] [18]. Probabilistic registration uncertainty has been formalised for radiotherapy dose propagation, where the certainty-map design dominates the resulting dose bounds [7], but no retrieved histology pipeline propagates warp uncertainty to a target coordinate. Whether deformation should be physically staged or learned is unresolved, and blockface and thick-section strategies are unavailable for archival specimens, which are precisely the material stereotactic atlases are built from [6] [2].

What would change the answer

An end-to-end stereotactic benchmark would settle this: same-specimen MRI plus serial histology with expert-delimited nuclei, scored by nucleus-centroid target registration error in stereotactic coordinates rather than Dice, with classical blockface, ANTs, SwanWarp and learned pipelines run on identical pairs [6] [10]. Adapting the anatomical-fiducial framework — which already uncouples registration-related variance from other sources of electrode-position variance — to histology-to-MRI warps would supply exactly the missing millimetric readout and permit direct comparison with the 1.49–6.85 mm AFRE range [18]. An ablation isolating the contribution of simulated slice distortion during training, tested on real sectioned material with known deformation, would establish whether learned priors recover geometry or merely smooth it [10]. Finally, a held-out prediction task — predicting post-operative DBS or stereo-EEG contact locations from a histological atlas warped by a learned transform, scored against imaging-localised contacts at the 0.23–0.30 mm proximity scale — would convert the field's overlap percentages into a claim a surgeon could act on [22].

References

[1] Behnaz Elhaminia, Abdullah Alsalemi, Esha Nasir, Mostafa Jahanifar, Ruqayya Awan, Lawrence S. Young, Nasir M. Rajpoot, Fayyaz Minhas, Shan E Ahmed Raza. "From Traditional to Deep Learning Approaches in Whole Slide Image Registration: A Methodological Review". arxiv. 2025-02-27T00:00:00.000Z. URL: https://arxiv.org/abs/2502.19123

[2] Wang, Y., Ho, W., Huszar, I. N., DiGiacomo, P., Taghavi, H. M., Tao, L., Choi, M., Nguyen, N., Leventis, S., Camarillo, D. B., Schlomer, P., Axer, M., Shao, W., Rusu, M., Cobos, I., Nirschl, J., Georgiadis, M., Zeineh, M.. "Precise MRI-Histology Coregistration of Paraffin-Embedded Tissue with Blockface Imaging". bioRxiv. 2025-06-05T00:00:00.000Z. URL: https://doi.org/10.1101/2025.06.02.657335

[3] Nanna E. Wielenberg, Ilinca Popp, Oliver Blanck, Lucas Zander, Jan C. Peeken, Stephanie E. Combs, Anca-Ligia Grosu, Dimos Baltas, Tobias Fechter. "Deep-Learning Atlas Registration for Melanoma Brain Metastases: Preserving Pathology While Enabling Cohort-Level Analyses". arxiv. 2026-02-16T00:00:00.000Z. URL: https://arxiv.org/abs/2602.12933

[4] Ahmed Aboelela, Johannes Barcsay, Jana Friedhof, Miguel Gonçalves, Alexander Hann, Katharina Breininger. "Evaluating Transformation Models for pCLE Mosaic Registration". arxiv. 2026-09-22T00:00:00.000Z. URL: https://arxiv.org/abs/2609.2456

[5] Zhuo, J., Zhang, Z., Xiao, C., Wang, M., Bi, S., Shi, L., Luo, Q., Gong, H., Shen, Z., Yang, X., Li, P.. "Fully Automated and Scalable Pipeline for Macaque Brain Registration". bioRxiv. 2025-07-16T00:00:00.000Z. URL: https://doi.org/10.1101/2025.07.10.663842

[6] Lopes Alho, E. J., Friedrich, H., Alho, A. T. D. L., Alaminos Bouza, A., Amaro, E., Grinberg, L. T., Oxenford, S., Talamoni Fonoff, E., Horn, A., Heinsen, H.. "The SWAN atlas: an open multimodal resource combining neuroimaging and high-resolution histology of the human brain in stereotactic space". bioRxiv. 2026-10-19T00:00:00.000Z. URL: https://doi.org/10.64898/2026.10.12.758634

[7] Stefan Heldmann, Sven Kuckertz, Nasim Givehchi, Thomas Coradi, Mikel Byrne, Ben Archibald-Heeren, Nils Papenberg. "A practical probabilistic framework for deformable image registration uncertainty in radiotherapy dose propagation". arxiv. 2026-06-09T00:00:00.000Z. URL: https://arxiv.org/abs/2606.09253

[8] Fanwen Wang, Ke Wen, Yaqing Luo, Yinzhe Wu, Jiahao Huang, Dudley J. Pennell, Pedro F. Ferreira, Andrew D. Scott, Sonia Nielles-Vallespin, Guang Yang. "Is fitting error a reliable metric for assessing deformable motion correction in quantitative MRI?". arxiv. 2025-03-11T00:00:00.000Z. URL: https://arxiv.org/abs/2503.07293

[9] Iglesias, J. E., Johnson, I. P., Williams-Ramirez, J., Zemlyanker, D., Tian, L., Gopinath, K., Olchanyi, M., Farnan, A. D., Demopoulos, A., Rosen, M. S., Sheth, K. N., de Havenon, A., Kimberly, W. T., Sorby-Adams, A.. "On the accuracy of image registration in portable low-field 3D brain MRI". bioRxiv. 2026-02-14T00:00:00.000Z. URL: https://doi.org/10.64898/2026.02.11.705413

[10] Fang, S., Tward, D. J.. "Deep Learning Based Cross-Modality Histological Brain Section Registration in Multiple Species Using Synthetic Images". bioRxiv. 2026-07-22T00:00:00.000Z. URL: https://doi.org/10.64898/2026.07.17.738987

[11] Valentin Boussot, Cedric Hemon, Caroline Lafond, Jean-Claude Nunes, Jean-Louis Dillenseger. "When Misalignment Becomes Supervision: Structured Label Noise in Supervised Synthetic CT Generation". arxiv. 2026-09-25T00:00:00.000Z. URL: https://arxiv.org/abs/2609.29387

[12] Chulong Zhang, Yuming Jiang, Na Li, Zhicheng Zhang, Md Tauhidul Islam, Jingjing Dai, Lin Liu, Wenfeng He, Wenjian Qin, Jing Xiong, Yaoqin Xie, Xiaokun Liang. "A Hybrid Deep Feature-Based Deformable Image Registration Method for Pathology Images". arxiv. 2024-10-28T00:00:00.000Z. URL: https://arxiv.org/abs/2208.07655

[13] Jiahao Qin, Yiwen Wang. "Learning Domain-Invariant Representations for Cross-Domain Image Registration via Scene-Appearance Disentanglement". arxiv. 2026-01-23T00:00:00.000Z. URL: https://arxiv.org/abs/2601.08875

[14] Jayakumar, J., Sivaprakasam, M., Verma, R., Bota, M., Joseph, J., Mulay, S., Kumutha, J., Srinivasan, C., S, S., S, L., Kumar, H. E., Bhaduri, A., Nowakowski, T. J., Roy, P. K., Savoia, S., Banerjee, S., Tward, D., Mitra, P. P.. "A three-dimensional histological cell atlas of the developing human brain". bioRxiv. 2024-12-18T00:00:00.000Z. URL: https://doi.org/10.1101/2024.12.17.628811

[15] Tward, D. J., Flannery, P. J., Savoia, S., Mezias, C., Banerjee, S., Richman, M., Lodato, B., O'Rourke, J. R., Balani, S., Arima, K., Washington, S., Coronado-Leija, R., Zhang, J., Mitra, P. P.. "3D multimodal histological atlas and coordinate framework for the mouse brain and head". bioRxiv. 2025-05-01T00:00:00.000Z. URL: https://doi.org/10.1101/2025.04.30.650246

[16] Brammerloh, M., Alkemade, A., Bazin, P.-L., Jantzen, C., Jäger, C., Herrler, A., Pine, K. J., Morawski, M., Balesar, R., Amunts, K., Forstmann, B. U., Weiskopf, N., Kirilina, E.. "Giving you five: A neuroimaging atlas of the nigrosomes in the substantia nigra based on 3D histology". bioRxiv. 2026-04-28T00:00:00.000Z. URL: https://doi.org/10.64898/2026.04.24.720555

[17] Xiang Chen, Fengting Zhang, Qinghao Liu, Min Liu, Kun Wu, Yaonan Wang, Hang Zhang. "Ideal Registration? Segmentation is All You Need". arxiv. 2025-09-22T00:00:00.000Z. URL: https://arxiv.org/abs/2509.15784

[18] Abbass, M., Taha, A., Gilmore, G., Santyr, B., Chalil, A., Jog, M., MacDougall, K., Parrent, A. G., Peters, T. M., Lau, J. C.. "The impact of localization and registration accuracy on estimates of deep brain stimulation electrode position in stereotactic space". medRxiv. 2025-03-12T00:00:00.000Z. URL: https://doi.org/10.1101/2024.03.01.24303286

[19] Valter, Y., Huang, Y., Khadka, N., Datta, A., Bikson, M.. "Neuronavigation-free and MRI-free Localization of Deep Brain Targets". bioRxiv. 2026-02-06T00:00:00.000Z. URL: https://doi.org/10.64898/2026.02.04.703768

[20] Koneru, R., Anant, M. M.-G., Kim, H., Cheron, J., Kebschull, J. M.. "DART: A GUI Pipeline for Aligning Histological Brain Sections to 3D Atlases and Automating Laser Microdissection". bioRxiv. 2025-09-08T00:00:00.000Z. URL: https://doi.org/10.1101/2025.09.03.673836

[21] Puchades, M. A., Yates, S. C., Csucs, G., Carey, H., Balkir, A., Leergaard, T. B., Bjaalie, J. G.. "Software and pipelines for registration and analyses of rodent brain image data in reference atlas space". bioRxiv. 2025-04-28T00:00:00.000Z. URL: https://doi.org/10.1101/2025.04.25.650563

[22] Park, S., Ortiz-Guerrero, G., Permezel, F., Dabaja, H., Osman, G., Burkett, B., Messina, S., Starnes, K., Lundstrom, B., Wong-Kisiel, L., Burkholder, D., Hermes, D., Brinkmann, B., Worrell, G., Marsh, R., Miller, K. J., Van Gompel, J., Gregg, N. M.. "Thalamic stereoelectroencephalography: safety, accuracy, and thalamocortical connectivity". medRxiv. 2025-12-02T00:00:00.000Z. URL: https://doi.org/10.64898/2025.12.01.25341375

[23] Puonti, O., Nolan, J., Dicamillo, R., Balbastre, Y., Casamitjana, A., Mancini, M., Robinson, E., Peter, L., Annunziata, R., Althonayan, J., Crampsie, S., Blackburn, E., Billot, B., Atzeni, A., Schmidt, P., Hughes, J., Augustinack, J., Edlow, B. L., Zöllei, L., Thomas, D., Kliemann, D., Bocchetta, M., Strand, C., Holton, J., Jaunmuktane, Z., Iglesias, J. E.. "Fast segmentation with the NextBrain histological atlas". bioRxiv. 2025-09-25T00:00:00.000Z. URL: https://doi.org/10.1101/2025.09.22.673638

How this survey was made

This survey answers one question: “How do deep-learning-based nonlinear registration methods for MRI-to-histology alignment handle slice-level tissue deformation and evaluate accuracy against stereotactic targets?”

We keep our own collection of recent preprints and refresh it every day from bioRxiv, medRxiv, arXiv, PsyArXiv, SocArXiv and ChemRxiv. For each survey, we search that collection for recent work and keep the strongest matches: papers that use the same terms as the question, and papers on the same topic in different words.

The text was written from the titles and abstracts of those papers; the full texts were not read. Statements that draw on a particular paper are marked with a number in brackets. The reference list just above is the 23 papers whose work this text draws on, chosen from the 120 papers those searches returned; the rest of the shortlist is not printed. It is not a complete review of everything published on the question.

Treat this as a starting point and a reading list, not a peer-reviewed review. The searches behind this survey ran on October 22, 2026 and covered work from the previous two years.