Automating Microinfarct Screening in Hematoxylin and Eosin‐stained Human Brain Tissues: A Machine Learning Approach

Luca Cerny Oliveira, Joohi Chauhan, Zhengfeng Lai, Samson Cheung, Amparo C. Villablanca, Lee‐Way Jin, Charles DeCarli, Chen‐Nee Chuah, Brittany N. DuggerAlzheimer 2024 · DOI

TL;DR

This paper introduces a machine learning pipeline to automatically detect microinfarcts in human brain tissue images, addressing a critical bottleneck in vascular dementia research. By leveraging convolutional neural networks and a novel multi-FOV approach, the system achieves high accuracy in identifying affected slides, paving the way for more efficient and objective neuropathological analysis.

What Problem Does This Paper Address?

Vascular dementia (VaD) is often characterized by microinfarcts, which are small areas of tissue damage in the brain. Accurately identifying and quantifying these microinfarcts is crucial for diagnosing and grading VaD. However, microinfarcts are notoriously difficult to assess due to their varied appearance and imprecise boundaries. This heterogeneity leads to significant inter-rater variability among expert pathologists, making consistent and objective grading a major challenge. Current manual analysis of whole slide images (WSIs) is time-consuming and prone to subjective interpretation. This paper aims to bridge this gap by developing an automated machine learning approach for microinfarct screening and detection in these complex neuropathological images.

Methods

The researchers adapted a patch-based pipeline using convolutional neural networks (CNNs) to automate the screening and detection of microinfarcts within WSIs. The core of their method involves breaking down the large WSIs into smaller patches, which are then fed into a CNN for analysis. They explored different "fields-of-view" (FOVs) – essentially different zoom levels or patch sizes – to understand their impact on performance and computational cost. To further refine the results and combat false positives, they incorporated a post-processing step and experimented with leveraging information from multiple FOVs simultaneously. The chosen CNN architecture for their experiments was ResNet-18.

Datasets

The study was validated on a dataset of 66 annotated WSIs. This dataset comprised 40 infarcted WSIs and 26 non-infarcted WSIs. These images were sourced from the frontal, parietal, and occipital white matter regions of 22 different cases. The annotations, which included delineations of microinfarct regions and grading of white matter rarefaction, were performed by a single trained expert.

Key Insights and Ideas

  • The inherent variability of microinfarcts is a major hurdle: The paper highlights that the diverse appearances of microinfarcts make them challenging to identify consistently, even for human experts. This variability is a key reason why automated methods are needed.
  • Screening vs. Detection: Two sides of the same coin: The study distinguishes between "screening" (identifying if a WSI contains any microinfarcts) and "detection" (precisely localizing the microinfarct regions within a WSI). Achieving high accuracy in both is critical for a useful automated system.
  • Multi-FOV approach is key to robustness: Simply analyzing patches at a single resolution isn't enough. Combining information from different FOVs significantly improves both the ability to screen slides and to accurately detect the microinfarct regions, while also reducing false alarms.

Key Results

The developed machine learning models demonstrated notable efficacy in screening WSIs for the presence of microinfarcts. Specifically, the ResNet-18 model achieved an impressive 100% accuracy at the WSI level, meaning it could correctly classify all slides as either containing infarcts or not. However, the sensitivity for detecting the precise microinfarct regions within a WSI was lower, falling below 50% when using an Intersection over Union (IoU) threshold of 20% (a common metric for object detection accuracy). The paper also notes that leveraging multiple FOVs improved both detection and screening performances and helped reduce false positives.

Limitations and Future Work

A significant limitation identified is the relatively low sensitivity in precisely localizing microinfarct regions, particularly when requiring a high degree of overlap (IoU > 20%). This suggests that while the model is good at knowing if an infarct is present, it struggles with pinpointing its exact boundaries. The authors propose that leveraging multiple FOVs is a promising direction to mitigate false positives and improve overall performance. Future work could focus on further refining the detection capabilities, potentially by exploring more advanced CNN architectures, different annotation strategies, or incorporating additional imaging modalities if available. The current study is a proof-of-concept, and further validation on larger and more diverse datasets would be beneficial.

Summary updated

Abstract

Abstract Background Microinfarcts, characteristic lesions of vascular dementia (VaD), are heterogenous and vary in appearance, which pose a considerable challenge for VaD grading as there is great interrater variability in microinfarct assessment. We propose a novel application of machine learning (ML) in the automated screening of microinfarcts, addressing a gap in the post‐mortem analysis of VaD in whole slide images (WSIs) from human brain. Method Our study adapts a patch‐based pipeline with convolutional neural networks (CNNs) to automate microinfarct screening in WSIs. We compare performance and computational cost of different fields‐of‐view (FOV) for the patch‐based method. Additionally, we propose a postprocessing step and leverage multiple FOVs to mitigate false positives. This study is validated against 66 annotated WSIs (N = 40 infarcted, N = 26 un‐infarcted) from Frontal, Parietal, and Occipital white matter regions across 22 cases. Annotations are from a single trained expert, who delineated microinfarct regions and graded white matter rarefaction. Result We report screening performance, i.e, the ability to distinguish infarct‐positive from infarct‐negative WSIs, and detection performance, i.e, the ability to localize the microinfarct regions within a WSI. Despite the inherent challenges of the inexact boundaries of microinfarcts, our models demonstrate notable efficacy in screening WSIs for infarcts, with ResNet‐18 achieving 100% accuracy at WSI level. However, the sensitivity in detecting infarct regions is below 50%, based on Intersection over Union thresholds above 20% overlap (a performance metric used to evaluate the accuracy of object detection models). Leveraging multiple FOV improves both detection and screening performances, reducing false positives. Conclusion This work presents a proof‐of‐concept pipeline driven by machine learning to automate microinfarct screening & detection in WSIs. This workflow provides new avenues in the automated analysis of neuropathological images of microinfarcts, bringing potential advancement in ML‐based dementia and neuropathology research.
BibTeX
@inproceedings{oliveira2024automating,
  author = {Luca Cerny Oliveira and Joohi Chauhan and Zhengfeng Lai and Samson Cheung and Amparo C. Villablanca and Lee‐Way Jin and Charles DeCarli and Chen‐Nee Chuah and Brittany N. Dugger},
  title = {{Automating Microinfarct Screening in Hematoxylin and Eosin‐stained Human Brain Tissues: A Machine Learning Approach}},
  booktitle = {Alzheimer},
  year = {2024},
  doi = {10.1002/alz.095132},
}
Figure