Achieving micron‐level precision MR‐Histology correlation for validating Alzheimer’s disease biomarkers in the human hippocampus

Yixin Wang, William Hai Dang Ho, István N. Huszár, Hossein Moein Taghavi, Jeffrey Nirschl, Samantha Leventis, Philip Schlömer, Markus Axer, Wei Shao, Mirabela Rusu, Phillip DiGiacomo, Marios Georgiadis, Michael ZeinehAlzheimer 2024 · DOI

TL;DR

This paper presents a novel pipeline for precisely aligning high-resolution ex-vivo MRI data with corresponding histology images of the human hippocampus. This micron-level precision correlation is crucial for validating Alzheimer's disease (AD) biomarkers detected non-invasively by MRI, paving the way for more accurate AD diagnosis and research.

What Problem Does This Paper Address?

Alzheimer's disease (AD) is characterized by microscopic changes in the brain, such as tau tangles and amyloid plaques, which can be detected by MRI. However, validating these MRI-based observations with the "gold standard" of histology has been a significant challenge. The primary hurdle is the non-linear 3D deformation that occurs when tissue is processed for histology, making it difficult to accurately map MRI signals to specific cellular structures. This paper tackles this challenge by developing a method to achieve precise, micron-level alignment between post-mortem MRI and histology of the human hippocampus.

Methods

The authors developed a multi-step pipeline to achieve this precise correlation. First, they acquired ultra-high-resolution ex-vivo MRI of human hippocampal specimens. Simultaneously, they captured optical images of the tissue block's surface at each sectioning step (blockface images) using a polarized filter. These 2D blockface images were then reconstructed into a 3D blockface volume.

The core of their method involves two key registration steps:

  1. 3D Registration: The 3D MRI volume was registered to the 3D blockface volume using either Tensor Image Registration Library (TIRL) or ANTs SyN. This step aligns the overall structure of the MRI with the physical tissue block.
  2. 2D Registration: Once the 3D volumes were aligned, the 2D MRI slices corresponding to the stained histology slides were identified. These 2D MRI slices were then registered to their respective histology images using 2D deformable registration with TIRL or ANTs.

To quantify the accuracy of their alignment, they segmented key hippocampal regions (white matter, grey matter, and dentate gyrus) in both the MRI and histology data and calculated the Dice similarity coefficient.

Datasets

The study utilized post-mortem human hippocampal specimens. The paper does not specify the number of specimens or their origin beyond being formalin-fixed human brains. No new benchmark datasets were introduced; rather, the focus was on developing a methodology for correlating existing MRI and histology data.

Key Insights and Ideas

  • Bridging the Gap: The pipeline successfully bridges the gap between non-invasive MRI and detailed histology, enabling direct validation of MRI-detected AD pathology at a microscopic level.
  • Handling Deformation: The method's ability to account for non-linear 3D deformations is critical for accurate correlation, as these deformations are a major obstacle in traditional approaches.
  • TIRL vs. ANTs: The study highlights that while both TIRL and ANTs are powerful registration tools, TIRL demonstrated superior performance for internal hippocampal features, suggesting its suitability for detailed biomarker validation.
  • Micron-Level Precision: The achievement of micron-level precision in coregistration is a significant step forward, allowing for the examination of very fine pathological details.

Key Results

  • 3D Registration Accuracy: When comparing 3D MRI and blockface segmentations, TIRL achieved average Dice similarity coefficients of 86.2% (hippocampal head) and 86.76% (hippocampal tail). ANTs SyN achieved slightly lower scores of 83.91% and 78.26%, respectively.
  • 2D Registration Accuracy: For 2D MR registration to histology, TIRL achieved impressive average Dice coefficients of 94.12% and 96.95%. ANTs performed marginally better on the whole tissue mask, but TIRL showed superior performance within internal features.
  • Qualitative Validation: The high-fidelity alignment was visually confirmed by the accurate mapping of small features like focal hypointensities and the dentate granule cell layer between MRI and histology.

Limitations and Future Work

Limitations: * The study was conducted on post-mortem tissue, which may not perfectly replicate the in-vivo conditions for AD biomarker detection. * The paper does not detail the specific characteristics or number of the human brain specimens used.

Future Work: The authors suggest that their advanced correlative pipeline "propels the development of MRI-based AD biomarkers." This implies future work will focus on applying this methodology to larger cohorts and a wider range of AD-related pathologies to refine and validate MRI biomarkers for clinical use.

Summary updated

Abstract

Abstract Background MRI offers potential noninvasive detection of Alzheimer's micropathology. The AD hippocampus exhibits microscopic pathological changes such as tau tangles, iron accumulation and late‐stage amyloid. Validating these changes from ultra‐high‐resolution ex‐vivo MRI through histology is challenging due to nonlinear 3D deformations between MRI and histological samples. We aim to address the challenge by a pipeline (Figure 1) for precise alignment of post‐mortem MRI data with histological images. Method A human hippocampal specimen was dissected from a formalin‐fixed human brain and divided into hippocampal head and tail. We obtained high‐resolution multi‐echo gradient echo ex vivo MRI of each using a 7T Bruker scanner. The specimens were paraffin‐embedded and sectioned using a Leica Histocore Nanocut R microtome. An optical image of the tissue block's surface was captured before each section was cut (Blockface image) using a polarized filter at Brewster’s angle. The obtained 2D blockface images were aligned and filtered into 3D blockface volumes. These were registered with MRI using either Tensor Image Registration Library (TIRL) with Modality Independent Neighbourhood Descriptor, or ANTs SyN registration. Subsequently, the 2D MRI slices corresponding to the stained histology slides were identified in the 3D volume and underwent 2D deformable registration with TIRL or ANTs. To quantify the registration accuracy, we segmented hippocampal volumes into white matter, grey matter, and dentate gyrus for MRI, blockface, and histology data separately. Result When comparing 3D MR and blockface 3D image segmentations (Figure 2, top), the average Dice similarity coefficient of hippocampal head and tail were 86.2% and 86.76% with TIRL, and 83.91% and 78.26% with ANTs SyN. Evaluating 2D MR registration to histological images, TIRL achieves the average Dice coefficient of 94.12% and 96.95%. ANTs performs marginally better across the whole tissue mask, but TIRL shows superior performance in the internal features (Table 1). High‐fidelity alignment shows very small features such as focal hypointensities (Figure 2, bottom, blue circles) and the dentate granule cell layer (orange arrows). Conclusion Our advanced correlative MRI‐histology pipeline achieves micron‐level precision coregistration methods at micron‐scale precision, propelling the development of MRI‐based AD biomarkers.
BibTeX
@inproceedings{wang2024achieving,
  author = {Yixin Wang and William Hai Dang Ho and István N. Huszár and Hossein Moein Taghavi and Jeffrey Nirschl and Samantha Leventis and Philip Schlömer and Markus Axer and Wei Shao and Mirabela Rusu and Phillip DiGiacomo and Marios Georgiadis and Michael Zeineh},
  title = {{Achieving micron‐level precision MR‐Histology correlation for validating Alzheimer’s disease biomarkers in the human hippocampus}},
  booktitle = {Alzheimer},
  year = {2024},
  doi = {10.1002/alz.092110},
}
Figure