Improving Neuropathological Reconstruction Fidelity via AI Slice Imputation

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

💡 Research Summary

This paper addresses a critical limitation of previous work that reconstructed three‑dimensional brain volumes from two‑dimensional dissection photographs: the anisotropic resolution caused by thick coronal slabs (often several millimeters) leads to overly smooth and anatomically inaccurate reconstructions, especially in cortical regions where precise surface geometry is essential. To overcome this, the authors introduce a computationally efficient super‑resolution step they call “AI slice imputation.” The method takes two adjacent photographic slices as input and predicts an intermediate “missing” slice using a 2‑D U‑Net architecture. By iteratively applying the network, an isotropic volume with a target through‑plane spacing (typically 1 mm) can be generated from arbitrarily spaced slabs, without requiring large memory or fixed inter‑slice distances.

Training such a model on real ultra‑thin (≈1 mm) slabs would be impractical, so the authors adopt a synthetic data paradigm. They start from existing 1 mm isotropic MRI datasets, randomly simulate a wide variety of image contrasts, illumination conditions, camera parameters, and slab thicknesses, and then down‑sample to create paired low‑ and high‑resolution examples. This domain‑randomized synthetic training enables the network to generalize robustly to real dissection photographs acquired under diverse protocols, as demonstrated in prior work on domain‑agnostic MRI analysis.

The evaluation uses two independent post‑mortem brain collections: the University of Washington Alzheimer’s Disease Center (UW) dataset, with homogeneous slabs of 4 mm thickness and artificially subsampled versions at 8 mm and 12 mm, and the Massachusetts Alzheimer’s Disease Research Center (MADRC) dataset, which contains heterogeneous slabs of roughly 8 mm. Qualitative inspection shows that the imputed volumes retain fine cortical folding and subcortical structures that are lost in the baseline reconstructions, especially at larger slab thicknesses.

Quantitative downstream analyses are performed on four fronts:

1. Surface reconstruction – Using the FreeSurfer‑based Recon‑Any pipeline, pial and white‑matter surfaces are extracted from both baseline and imputed volumes and compared to gold‑standard surfaces derived from ex‑vivo 1 mm FLAIR MRI. Mean closest‑point distance errors decrease by up to 0.4 mm after imputation, with statistically significant improvements (Wilcoxon p < 0.001) across all slab thicknesses.

2. Cortical thickness estimation – Thickness errors, computed as the distance between the two surfaces, are similarly reduced, confirming that the imputed volumes provide more accurate cortical geometry.

3. Automated segmentation – Gold‑standard segmentations are obtained by applying SynthSeg to the MRI scans. Photo‑SynthSeg (trained for large coronal spacings) is applied to baseline volumes, while SynthSeg (contrast‑agnostic) processes the imputed volumes. Region‑wise Dice scores improve markedly for almost every structure, with the most pronounced gains in cortex and white matter. Only two out of 36 region‑thickness combinations show a slight decline (ventricle and amygdala in MADRC).

4. Atlas registration – The MNI‑ICBM152 atlas is non‑rigidly registered to each reconstruction using NiftyReg. Dice overlap between the warped atlas labels and the MRI‑derived gold‑standard labels is higher after imputation for every dataset, thickness, and region, again with strong statistical significance.

These results collectively demonstrate that slice imputation restores anatomically plausible detail lost due to through‑plane undersampling, and that a synthetic training regime can bridge the domain gap between MRI‑derived data and real photographic slabs. The approach is computationally lightweight, requires only modest hardware for inference, and does not need any fine‑tuning on the target photographs.

The authors acknowledge limitations: the current 2‑D U‑Net only leverages two neighboring slices, potentially missing broader 3‑D context; extremely thick slabs (>12 mm) and specimens with severe pathological deformation were not extensively tested. Future work could explore 3‑D or recurrent architectures, incorporate uncertainty estimation, and integrate the imputation step into a fully automated neuropathology pipeline.

All code and pretrained models are publicly released at https://surfer.nmr.mgh.harvard.edu/fswiki/mri_3d_photo_recon, facilitating adoption by the neuropathology and neuroimaging communities.