Interpretable classification of Alzheimer’s disease pathologies with a convolutional neural network pipeline

Ziqi Tang, Kangway V. Chuang, Charles DeCarli, Lee‐Way Jin, Laurel Beckett, Michael J. Keiser, Brittany N. Dugger

TL;DR

This paper introduces a deep learning pipeline that can accurately classify and localize Alzheimer's disease pathologies (amyloid plaques and cerebral amyloid angiopathy) in digitized brain tissue slides. By developing an interpretable CNN and a robust annotation platform, the researchers aim to provide a scalable and precise tool to aid neuropathologists in disease diagnosis and research.

What Problem Does This Paper Address?

Diagnosing Alzheimer's disease (AD) neuropathologically relies on expert neuropathologists examining vast brain tissue samples to identify and quantify amyloid-beta (Aβ) plaques and cerebral amyloid angiopathy (CAA). Current methods, like semi-quantitative scoring (e.g., CERAD), are often coarse-grained, lack precise neuroanatomic localization, and can be time-consuming and subject to inter-rater variability. This limits the ability to perform large-scale, detailed analyses of AD pathologies and their correlation with clinical outcomes. The paper addresses the need for more precise, scalable, and interpretable methods to characterize these pathologies.

Methods

The core of the approach is a convolutional neural network (CNN) pipeline designed for analyzing whole slide images (WSIs) of immunohistochemically-stained brain tissue. Key methodological components include:

  • Automated Preprocessing: A pipeline to normalize slide color and segment stained objects (plaques) from WSIs.
  • Cloud-Based Annotation Platform: A custom web interface built on AWS Elastic Beanstalk for rapid, expert annotation of candidate plaque image tiles. This platform incorporated gamification elements to maintain high annotation rates.
  • Deep Learning Model: A six-layer convolutional architecture with two dense layers, trained as a multi-task CNN for classifying cored plaques, diffuse plaques, and CAA. The model was optimized using data augmentation and minority class oversampling.
  • Interpretability Techniques:
    • Prediction Confidence Maps: Visualizations that show the model's confidence in predicting specific pathologies across an entire WSI, aiding in understanding their distribution and neuroanatomic localization.
  • Guided Grad-CAM: Used to identify salient visual features that the CNN focuses on for its predictions, highlighting important morphological characteristics.
  • Feature Occlusion Studies: A method to systematically mask parts of an image to observe how it affects the CNN's prediction confidence, further revealing feature importance.

Datasets

  • Training and Validation Data: A dataset of over 70,000 annotated plaque candidates extracted from 43 whole slide images (WSIs) of temporal gyri. This dataset was built in phases, with an initial large set and subsequent enrichment of minority classes (cored plaques and CAA).
  • Hold-out Test Sets:
    • A 10-WSI hold-out set for evaluating the CNN's classification performance.
  • A 30-WSI blinded hold-out set (combining 10 WSIs from the initial hold-out and 20 new WSIs) used to compare CNN-derived WSI scores with established semi-quantitative scores.

Key Insights and Ideas

  • Deep Learning for Neuropathology is Feasible: The study demonstrates that CNNs can achieve expert-level performance in classifying subtle Aβ pathologies, moving beyond coarse semi-quantitative measures.
  • Interpretability is Crucial: The success of the pipeline hinges not just on accuracy but also on making the CNN's decisions understandable through confidence maps and feature visualization, enabling expert validation and trust.
  • Annotation at Scale is Achievable: The custom cloud-based platform and gamification elements highlight a practical pathway for generating large, high-quality annotated datasets required for deep learning in pathology.
  • Morphology Matters: The CNN learns to distinguish between different Aß plaque morphologies (cored, diffuse, CAA) based on their visual characteristics, aligning with established neuropathological understanding.
  • Neuroanatomic Localization is Possible: Confidence maps reveal that different plaque types have distinct spatial distributions within the brain, offering a more granular view than traditional methods.

Key Results

  • High Classification Accuracy: The CNN achieved strong performance on a 10-WSI hold-out set, with an Area Under the Receiver Operating Characteristic (AUROC) of 0.993 for cored plaques and 0.743 for Precision-Recall Curve (AUPRC).
  • Correlation with Semi-Quantitative Scores: CNN-derived WSI scores for Aβ pathologies showed strong correlation with established semi-quantitative CERAD-like scores on a 30-WSI blinded hold-out set, demonstrating clinical relevance.
  • Efficient Training: The study found that significant performance gains could be achieved with relatively modest amounts of expert labeling time (e.g., ~15 hours for initial gains).
  • Interpretable Feature Learning: Guided Grad-CAM and feature occlusion studies confirmed that the CNN learns features consistent with accepted pathological characteristics of cored plaques, diffuse plaques, and CAA.

Limitations and Future Work

  • Single Expert Annotation: The current dataset and model were based on annotations from a single neuropathologist. Future work could explore inter-rater reliability and the impact of diverse annotation criteria.
  • Single Stain and Region: The study focused on Aβ IHC staining in temporal gyri. Extending the pipeline to other stains, anatomic regions, and pathology types (e.g., tau) is a natural next step.
  • Dataset Size for CAA: The relatively smaller dataset for CAA might limit its classification performance compared to cored and diffuse plaques. Further annotation of CAA morphologies could improve this.
  • Generalizability: While the model showed good generalization to unseen cases, further validation across diverse cohorts, brain banks, and staining protocols is necessary.
  • Clinical Correlation: The paper lays the groundwork for deeper clinicopathological correlations, which is a key future direction.
  • Refinement of Workflow: The authors suggest that future workflows could incorporate reinforcement learning from expert feedback on misclassified examples.

Summary updated

Abstract

Neuropathologists assess vast brain areas to identify diverse and subtly-differentiated morphologies. Standard semi-quantitative scoring approaches, however, are coarse-grained and lack precise neuroanatomic localization. We report a proof-of-concept deep learning pipeline that identifies specific neuropathologies-amyloid plaques and cerebral amyloid angiopathy-in immunohistochemically-stained archival slides. Using automated segmentation of stained objects and a cloud-based interface, we annotate > 70,000 plaque candidates from 43 whole slide images (WSIs) to train and evaluate convolutional neural networks. Networks achieve strong plaque classification on a 10-WSI hold-out set (0.993 and 0.743 areas under the receiver operating characteristic and precision recall curve, respectively). Prediction confidence maps visualize morphology distributions at high resolution. Resulting network-derived amyloid beta (Aβ)-burden scores correlate well with established semi-quantitative scores on a 30-WSI blinded hold-out. Finally, saliency mapping demonstrates that networks learn patterns agreeing with accepted pathologic features. This scalable means to augment a neuropathologist's ability suggests a route to neuropathologic deep phenotyping.
BibTeX
@inproceedings{tang2019interpretable,
  author = {Ziqi Tang and Kangway V. Chuang and Charles DeCarli and Lee‐Way Jin and Laurel Beckett and Michael J. Keiser and Brittany N. Dugger},
  title = {{Interpretable classification of Alzheimer’s disease pathologies with a convolutional neural network pipeline}},
  booktitle = {Nature Communications},
  year = {2019},
  doi = {10.1038/s41467-019-10212-1},
}
Figure