Automated Segmentation of Coronal Brain Tissue Slabs for 3D Neuropathology

TL;DR

A deep learning model is presented to automate segmentation of postmortem brain tissue from photographs, using a U-Net architecture that was trained with a combination of 1,414 manually segmented images of both fixed and fresh tissue from specimens with varying diagnoses.

Abstract

Advances in image registration and machine learning have recently enabled volumetric analysis of postmortem brain tissue from conventional photographs of coronal slabs, which are routinely collected in brain banks and neuropathology laboratories worldwide. One caveat of this methodology is the requirement of segmentation of the tissue from photographs, which currently requires costly manual intervention. In this article, we present a deep learning model to automate this process. The automatic segmentation tool relies on a U-Net architecture that was trained with a combination of 1,414 manually segmented images of both fixed and fresh tissue, from specimens with varying diagnoses, photographed at two different sites. Automated model predictions on a subset of photographs not seen in training were analyzed to estimate performance compared to manual labels, including both inter- and intra-rater variability. Our model achieved a median Dice score over 0.98, mean surface distance under 0.4mm, and 95\% Hausdorff distance under 1.60mm, which approaches inter-/intra-rater levels. Our tool is publicly available at surfer.nmr.mgh.harvard.edu/fswiki/PhotoTools.

Figures (10)

• Figure 1: Close-up of a dissection photograph (a) and its corresponding reference segmentation (b). The segmentation excludes the cortical surface (shown in gray), which lies outside the plane of the block face and is not relevant for downstream analysis. Segmenting this region is often time-consuming, as simple thresholding (effective for other parts of the image) fails to distinguish it accurately. Therefore, manual tracing is typically required, making the process labor-intensive. Consequently, fully automated segmentation methods are highly desirable to improve efficiency and scalability.
• Figure 2: Example dissection photography setups across the three photographic datasets used in this study. (a) Photograph from the MADRC dataset showing formalin-fixed slabs from a whole brain, with a ruler for scale calibration. (b) Image from the UW-fixed dataset, also depicting formalin-fixed whole-brain slabs with two orthogonal rulers included for spatial reference. (c,d) Images from the UW-fresh dataset showing coronal slabs from single hemispheres photographed on two slightly different tables. Four fiducial markers placed in a rectangular configuration are used for pixel size and perspective correction.
• Figure 3: (a) Sample photograph from the MADRC dataset. (b) Corresponding reference mask. (c) Automated segmentation. (d-f) Example from UW-fixed dataset; note the non-target slabs leading to irrelevant false positives that are factored out of our accuracy metrics. (g-h) Example from the UW-fresh dataset. A more comprehensive set of examples can be found in the supplement Figure \ref{['fig:s1']} to \ref{['fig:s6']}
• Figure 4: Box plots for Dice scores, mean surface distance, and HD95 for manual and automated methods. "Intra-1": intra-rater variability of Labeler 1, measured on 20 images. "Intra-2": intra-rater variability of Labeler 2 on the same 20 images. "Inter": inter-rater variability of the same 20 images. "Auto-in": accuracy of the proposed automated method on the in-distribution data (159 photographs). "Auto-out": accuracy of the proposed automated method on the out-of-distribution data (218 photographs).
• Figure S1: Additional automated segmentations for sample images from MADRC dataset (in distribution).