This project develops explainable deep learning methods for neuritic plaque segmentation in Alzheimer’s disease histopathology, with the objective of supporting traceable, interpretable and reproducible computational histopathology for patient stratification.
Research Themes
Project Overview
Alzheimer’s disease is associated with the misfolding and accumulation of amyloid-beta peptides and tau proteins. However, clinical presentation is highly heterogeneous, and in vivo imaging approaches do not provide the microscopic resolution available in post-mortem histopathology.
The project focuses on the automatic segmentation of tau protein aggregates, especially neuritic plaques, from whole-slide histopathological images. It also investigates how visual explainability and attention mechanisms can support the refinement of manual annotations and improve the interpretability of deep learning models.
Scientific Motivation
- Clinical heterogeneity: Alzheimer’s disease patients show highly variable clinical trajectories and pathological presentations.
- Microscopy-level information: Histopathology provides spatial and morphological detail that cannot be reached by conventional in vivo imaging.
- Tau pathology relevance: Tau-related lesions are strongly linked to clinical manifestations and disease progression.
- Annotation challenge: Manual annotation of neuritic plaques is difficult, time-consuming and subject to variability.
Dataset
The study uses two whole-slide image datasets with different antibodies, scanner conditions and patch sizes. These variations are important because acquisition protocols, staining conditions and contextual information can strongly influence segmentation performance.
| Dataset 1 | Dataset 2 |
|---|---|
| 6 WSI | 8 WSI |
| ALZ50 antibodies | AT8 antibodies |
| NanoZoomer 2.0-RS | NanoZoomer 2.0-RS & S60 |
| 128 × 128 patch size | 128 × 128 & 256 × 256 patch size |
| 20× magnification | 20× magnification |
Resources
Results and Visual Explanations
Context and Patch Size
The study shows that contextual information and patch size influence segmentation performance, underlining the importance of carefully designing spatial context in histopathological deep learning.
Attention-Based Explainability
Successive activation layers of attention U-Net highlight a progressive focusing mechanism. This visual explainability was used to support manual annotation refinement of neuritic plaques.
Application to Patient Stratification
The final application integrates the developed deep learning pipelines into a computational histopathology framework aimed at refining Alzheimer’s disease patient stratification.
Key Ideas from the Study
- Antibody choice can impact the detection and segmentation of tau aggregates in whole-slide images.
- AT8 staining may produce less compact structures, making plaque segmentation more challenging.
- Scanner variability and acquisition differences affect segmentation and annotation consistency.
- Context size influences segmentation performance and must be explicitly controlled.
- Visual explainability can help refine manual annotations and improve morphological interpretation.
- Explainable deep learning can outperform black-box commercial software while providing better traceability.
How to Cite
This project was published at MICCAI 2022. Please cite the work below if you use the associated code, dataset or methodology.
@InProceedings{Jimenez2022StratifIAD,
author = {Jimenez, Gabriel and Kar, Anuradha and Ounissi, Mehdi and Ingrassia, Lea
and Boluda, Susana and Delatour, Benoit and Stimmer, Lev and Racoceanu, Daniel},
title = {Visual Deep Learning-Based Explanation for Neuritic Plaques Segmentation
in Alzheimer's Disease Using Weakly Annotated Whole Slide Histopathological Images},
booktitle = {Medical Image Computing and Computer-Assisted Intervention -- MICCAI 2022},
year = {2022},
publisher = {Springer International Publishing}
}
