Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Improving Neuropathological Reconstruction Fidelity via AI Slice Imputation

Marina Crespo Aguirre, Jonathan Williams-Ramirez, Dina Zemlyanker, Xiaoling Hu, Lucas J. Deden-Binder, Rogeny Herisse, Mark Montine, Theresa R. Connors, Christopher Mount, Christine L. MacDonald, C. Dirk Keene, Caitlin S. Latimer, Derek H. Oakley, Bradley T. Hyman, Ana Lawry Aguila, Juan Eugenio Iglesias

TL;DR

The paper tackles the challenge of constructing isotropic brain volumes from thick coronal slabs of postmortem tissue, which typically yield coarse reconstructions. It introduces a lightweight super-resolution step—slice imputation—via a residual 2D U‑Net trained on domain-randomized synthetic data, predicting missing slices between adjacent slabs to produce $1$ mm isotropic volumes. Key contributions include a robust synthetic-data generator, a four-channel input with residual learning, and demonstrated improvements in cortical surface fidelity, region-wise Dice segmentation, and atlas registration across datasets and slab thicknesses ($2$–$12$ mm). The approach enhances morphometric analyses and strengthens the bridge between neuropathology and neuroimaging, evidenced by significant accuracy gains and public availability of the pipeline, while acknowledging limitations and future directions such as improved edge handling and multi-channel simulations.

Abstract

Neuropathological analyses benefit from spatially precise volumetric reconstructions that enhance anatomical delineation and improve morphometric accuracy. Our prior work has shown the feasibility of reconstructing 3D brain volumes from 2D dissection photographs. However these outputs sometimes exhibit coarse, overly smooth reconstructions of structures, especially under high anisotropy (i.e., reconstructions from thick slabs). Here, we introduce a computationally efficient super-resolution step that imputes slices to generate anatomically consistent isotropic volumes from anisotropic 3D reconstructions of dissection photographs. By training on domain-randomized synthetic data, we ensure that our method generalizes across dissection protocols and remains robust to large slab thicknesses. The imputed volumes yield improved automated segmentations, achieving higher Dice scores, particularly in cortical and white matter regions. Validation on surface reconstruction and atlas registration tasks demonstrates more accurate cortical surfaces and MRI registration. By enhancing the resolution and anatomical fidelity of photograph-based reconstructions, our approach strengthens the bridge between neuropathology and neuroimaging. Our method is publicly available at https://surfer.nmr.mgh.harvard.edu/fswiki/mri_3d_photo_recon

Improving Neuropathological Reconstruction Fidelity via AI Slice Imputation

TL;DR

The paper tackles the challenge of constructing isotropic brain volumes from thick coronal slabs of postmortem tissue, which typically yield coarse reconstructions. It introduces a lightweight super-resolution step—slice imputation—via a residual 2D U‑Net trained on domain-randomized synthetic data, predicting missing slices between adjacent slabs to produce mm isotropic volumes. Key contributions include a robust synthetic-data generator, a four-channel input with residual learning, and demonstrated improvements in cortical surface fidelity, region-wise Dice segmentation, and atlas registration across datasets and slab thicknesses (– mm). The approach enhances morphometric analyses and strengthens the bridge between neuropathology and neuroimaging, evidenced by significant accuracy gains and public availability of the pipeline, while acknowledging limitations and future directions such as improved edge handling and multi-channel simulations.

Abstract

Neuropathological analyses benefit from spatially precise volumetric reconstructions that enhance anatomical delineation and improve morphometric accuracy. Our prior work has shown the feasibility of reconstructing 3D brain volumes from 2D dissection photographs. However these outputs sometimes exhibit coarse, overly smooth reconstructions of structures, especially under high anisotropy (i.e., reconstructions from thick slabs). Here, we introduce a computationally efficient super-resolution step that imputes slices to generate anatomically consistent isotropic volumes from anisotropic 3D reconstructions of dissection photographs. By training on domain-randomized synthetic data, we ensure that our method generalizes across dissection protocols and remains robust to large slab thicknesses. The imputed volumes yield improved automated segmentations, achieving higher Dice scores, particularly in cortical and white matter regions. Validation on surface reconstruction and atlas registration tasks demonstrates more accurate cortical surfaces and MRI registration. By enhancing the resolution and anatomical fidelity of photograph-based reconstructions, our approach strengthens the bridge between neuropathology and neuroimaging. Our method is publicly available at https://surfer.nmr.mgh.harvard.edu/fswiki/mri_3d_photo_recon
Paper Structure (19 sections, 2 equations, 11 figures, 3 tables)

This paper contains 19 sections, 2 equations, 11 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Results
  3. Qualitative Results
  4. Surface Reconstruction
  5. Volume Segmentation
  6. Atlas Registration
  7. Discussion and Conclusion
  8. Materials and Methods
  9. Datasets
  10. Training data
  11. Test data
  12. Methods
  13. Data generator
  14. AI slice imputation
  15. Acknowledgment
  16. ...and 4 more sections

Figures (11)

  • Figure 1: Input and outputs for a sample case. (a) 3D Surface scan of the left human hemisphere acquired prior to dissection. (b) Photographs of dissected coronal slabs (thickness$\approx$8 mm), post pixel size calibration, with digital ruler overlaid. (c) 3D reconstruction of the photographs into an imaging volume. (d) Sagittal cross-section of the volume prior to imputation. (e) Sagittal cross-section of the volume after imputation with our approach, which recovers high-resolution detail. (f) Corresponding slice of automated segmentation obtained from (e) with Photo-SynthSeg gazula2024machine. (g) Pial surface with overlaid parcellation, obtained from (e) using Recon-Any gopinath2025recon. (h) White matter surface obtained with Recon-Any.
  • Figure 1: Pial surfaces with overlaid parcellations, computed with Recon-Any on one case from the MADRC dataset. (a) Reference surface from MRI (gold standard). (b) Surface obtained from the 3D reconstruction of slab photographs. (c) Surface obtained with the proposed imputation.
  • Figure 2: Axial and sagittal views of 3D reconstructions before and after imputation. (a) Sample case from the UW dataset, showing the original 4-mm slabs and the 8-mm and 12-mm variants. (b) Sample case from the MADRC dataset, comprising 8-mm thick slices. On the left, 3D reconstructions from our previous work gazula2024machine; on the right, results of the proposed imputation.
  • Figure 2: Color coded illustration of cortical thickness error distributions of the original (left) and our proposed method of imputation (right), overlaid on inflated surface hemispheres computed with Recon-Any. (a) Distribution of errors from reconstructions of UW dataset at three slab thicknesses (4, 8, 12 mm) (b) Distribution of errors from MADRC reconstructions (8 mm).
  • Figure 3: Pial surface meshes with overlaid parcellations of Recon-Any on one case from the UW dataset at three slab thicknesses, with and without imputation. (a) Reference pial surface from the MRI (gold standard); (b) surface results from the 3D reconstruction of slab photographs, and (c) surface results from the proposed imputations.
  • ...and 6 more figures