没有合适的资源?快使用搜索试试~ 我知道了~
人工智能在神经病理学中的应用:基于深度学习的tau蛋白病变评估.docx
0 下载量 171 浏览量
2024-04-23
07:43:09
上传
评论
收藏 2.82MB DOCX 举报
温馨提示
试读
22页
人工智能在神经病理学中的应用:基于深度学习的tau蛋白病变评估.docx
资源推荐
资源详情
资源评论
Published in final edited form as:
Lab Invest. 2019 July ; 99(7): 1019–1029. doi:10.1038/s41374-019-0202-4.
Artificial intelligence in neuropathology: deep learning-based
assessment of tauopathy
Maxim Signaevsky
1,2,3
, Marcel Prastawa
1,4
, Kurt Farrell
1,2,3
, Nabil Tabish
1,2,3
, Elena
Baldwin
1,2,3
, Natalia Han
1,2,3
, Megan A. Iida
1,2,3
, John Koll
1,4
, Clare Bryce
1,2,3
, Dushyant
Purohit
1,2,5
, Vahram Haroutunian
5,6
, Ann C. McKee
7,8,9,10,11
, Thor D. Stein
8,9,10,11
, Charles
L. White III
12
, Jamie Walker
12
, Timothy E. Richardson
12
, Russell Hanson
1,2,3
, Michael J.
Donovan
1,4
, Carlos Cordon-Cardo
1,4
, Jack Zeineh
1,4
, Gerardo Fernandez
1,4
, John F.
Crary
1,2,3
1
Department of Pathology, Icahn School of Medicine at Mount Sinai, New York, NY 10029, USA
2
Department of Neuroscience, Icahn School of Medicine at Mount Sinai, New York, NY 10029,
USA
3
Ronald M. Loeb Center for Alzheimer’s Disease, Icahn School of Medicine at Mount Sinai, New
York, NY 10029, USA
4
Center for Computational and Systems Pathology, Icahn School of Medicine at Mount Sinai, New
York, NY 10025, USA
5
Departments of Psychiatry and Neuroscience, Icahn School of Medicine at Mount Sinai, New
York, NY 10029, USA
6
J. James Peters VA Medical Center, Bronx, NY, USA
7
Department of Neurology, Boston University School of Medicine, Boston, MA 02118, USA
8
Department of Pathology, Boston University School of Medicine, Boston, MA 02118, USA
9
Alzheimer’s Disease Center, CTE Program, Boston University School of Medicine, Boston, MA
02118, USA
10
Mental Illness Research, Education and Clinical Center, James J. Peters VA Boston Healthcare
System, Boston, MA 02130, USA
11
Department of Veteran Affairs Medical Center, Bedford, MA 01730, USA
12
Neuropathology Laboratory, Department of Pathology, UT Southwestern Medical Center, Dallas,
TX 75390, USA
Abstract
Accumulation of abnormal tau in neurofibrillary tangles (NFT) occurs in Alzheimer disease (AD)
and a spectrum of tauopathies. These tauopathies have diverse and overlapping morphological
John F. Crary, john.crary@mountsinai.org.
Conflict of interest The authors declare that they have no conflict of interest.
Publisher’s note: Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
HHS Public Access
Author manuscript
Lab Invest. Author manuscript; available in PMC 2020 November 24.
Author Manuscript
Author Manuscript
Author Manuscript
Author Manuscript
Lab Invest. Author manuscript; available in PMC 2020 November 24.
Signaevsky et al. Page 2
phenotypes that obscure classification and quantitative assessments. Recently, powerful machine
learning-based approaches have emerged, allowing the recognition and quantification of
pathological changes from digital images. Here, we applied deep learning to the neuropathological
assessment of NFT in postmortem human brain tissue to develop a classifier capable of
recognizing and quantifying tau burden. The histopathological material was derived from 22
autopsy brains from patients with tauopathies. We used a custom web-based informatics platform
integrated with an in-house information management system to manage whole slide images (WSI)
and human expert annotations as ground truth. We utilized fully annotated regions to train a deep
learning fully convolutional neural network (FCN) implemented in PyTorch against the human
expert annotations. We found that the deep learning framework is capable of identifying and
quantifying NFT with a range of staining intensities and diverse morphologies. With our FCN
model, we achieved high precision and recall in naive WSI semantic segmentation, correctly
identifying tangle objects using a SegNet model trained for 200 epochs. Our FCN is efficient and
well suited for the practical application of WSIs with average processing times of 45 min per WSI
per GPU, enabling reliable and reproducible large-scale detection of tangles. We measured
performance on test data of 50 pre-annotated regions on eight naive WSI across various
tauopathies, resulting in the recall, precision, and an F1 score of 0.92, 0.72, and 0.81, respectively.
Machine learning is a useful tool for complex pathological assessment of AD and other
tauopathies. Using deep learning classifiers, we have the potential to integrate cell- and region-
specific annotations with clinical, genetic, and molecular data, providing unbiased data for
clinicopathological correlations that will enhance our knowledge of the neurodegeneration.
Introduction
Tau-related neurodegenerative disorders, the tauopathies, comprise a heterogeneous group of
disorders with a clinical spectrum that includes primary motor symptoms, movement
disorder, psychiatric dysfunction, and cognitive impairment [1]. Histomorphologically,
tauopathies are characterized by intracellular deposition of hyperphosphorylated tau protein.
Various isoform compositions, morphology, and anatomical distributions of intracellular tau
represent distinct diagnostic features of tauopathies [1–3]. How pathological tau causes
neuronal dysfunction and degeneration is unclear. Several mechanisms have been
implicated, including both genetic and environmental risk factors, but most cases are
idiopathic [1, 3–5]. Sporadic tauopathies, such as the vast majority of Alzheimer disease
(AD) and progressive supranuclear palsy (PSP) cases, are associated with common genetic
risk alleles [1, 3]. Rare highly penetrant mutations in the microtubule-associated protein tau
gene are associated with some forms of frontotemporal lobar degeneration [6].
Environmental factors, such as traumatic brain injury in the case of chronic traumatic
encephalopathy (CTE) or putative neurotoxins, have also been implicated [7, 8].
Pathological changes in tau metabolism and post-translational modifications result in the
accumulation of toxic forms of misfolded tau aggregates in neurons and glial cells in various
brain regions. These misfolded aggregates are associated with loss of function and
ultimately cell death [1, 2].
Pathological tau forms inclusions in neurons and glia with histomorphologically
distinguishable features. In neurons, these take the form of the classical flame-shaped
Author Manuscript
Author Manuscript
Author Manuscript
Author Manuscript
Lab Invest. Author manuscript; available in PMC 2020 November 24.
Signaevsky et al. Page 3
intracellular neurofibrillary tangles (NFTs), granular pre-NFTs, extracellular “ghost”
tangles, ring tangles, and globose tangles, among others [9]. In glia, there is a spectrum of
characteristic histomorphological forms that are commonly associated with specific diseases,
including glial plaques of corticobasal degeneration, tufted astrocytes of PSP, globular
astroglial inclusions in globular glial tauopathy, ramified astrocytes of Pick disease, and
thorn-shaped astrocytes as well as granular fuzzy astrocytes of aging-related tau
astrogliopathy [9–11]. One recently proposed classification scheme codifies seven primary
tauopathies, and two secondary tauopathies under the umbrella of neurodegenerative
diseases, each with a unique constellation of regional vulnerability and histomorphology of
tau aggregates that define them [1, 2]. Pathological accumulation of hyperphosphorylated
tau is also described in various infectious/post-infectious, metabolic, genetic/chromosomal,
neoplastic/hamartomatous, and myopathic diseases [12]. Given the complexity and
morphological overlap, diagnosing these diseases is a challenge for neuropathologists, and
commands a high degree of expertise.
Microscopic analysis of stained postmortem sections by a trained expert remains the only
modality of confirmatory diagnosis of tauopathies. Despite the continuous effort and
improvements in the field, the analyses required for definitive diagnosis and subtyping of
neurodegenerative diseases remain highly time- and cost-consuming and are subject to a
substantial degree of inter- and intra-observer variability, thus lacking overall accuracy and
precision. The gold standard for histomorphological assessment of tau burden and
progression in Alzheimer’s disease is the Braak staging system, which focuses on the
hierarchical sequence of tau accumulation, but not a quantitative measurement of tau burden,
although distribution and qualitative NFT and thread density are correlated in this staging
system [13]. Despite this limitation, the Braak staging system has been widely accepted and
adopted for decades for its simplicity and robustness. Recent interest in differential semi-
quantitative assessment of tau burden in AD is exemplified in the work of Jellinger [14].
Further, various stages of intracellular pathologic tau accumulation are described (e.g., pre-
tangles, mature NFTs, and so-called “ghost” tangles—the remnants of the tau fibrillary
scaffold after neuronal cell death; Fig. 1). The Braak staging approach does not address
these features, and thus inherently lacks granularity and quantification. At the same time, the
field of diagnostic neuropathology is facing challenges related to the overall lack of
accuracy, demanded by the ever-evolving research and healthcare standards, and
discrepancies with clinicopathological correlations, with a recognized need to address these
issues [15].
Recently, there has been an increasing interest in developing computational methods to assist
the pathologist in histological analysis via digital microscopic whole slide images (WSI).
This is primarily intended to reduce the human error rate and bring about uniformity and
accuracy in pathological diagnosis [16]. One of the approaches that has been anticipated and
sought after for nearly half a century is artificial intelligence (AI) [17, 18]. The most
advanced AI, called deep learning (DL), is now used for complex tasks such as speech
recognition, language translation, and image recognition and interpretation [19–21]. Litjens
et al. provide a comprehensive survey of published studies on the use of AI/DL in medical
image analysis including WSI in pathology [17]. Although machine learning-based methods
have had limited application in diagnostic pathology to date, due to the variability of
Author Manuscript
Author Manuscript
Author Manuscript
Author Manuscript
Lab Invest. Author manuscript; available in PMC 2020 November 24.
Signaevsky et al. Page 4
laboratory standards and outcomes, and lack of reliable computer-backed platforms,
advances have been made recently. The relevance and potential of automated classification
algorithms in surgical pathology are exemplified by its application to the histologic grading
and progression of breast and prostate cancer [17, 22, 23]. These endeavors pave a way
toward increased use of machine learning for improving stratification, characterization, and
quantification for many other disease processes, including the neuropathological assessment
of tauopathies and AD cohorts. To date, no datasets derived from the application of
machine-based learning to neurodegenerative disease are available.
We aimed to develop and test a novel DL algorithm using convolutional neural networks
[20] that would be able to recognize, classify, and quantify diagnostic elements of
tauopathies on WSI of postmortem human brain tissue specimens from patients with tau-
associated neurodegenerative conditions in order to better stratify patients for clinical and
other correlative studies (Fig. 2). In this study, we focused on the development, validation,
and testing of the DL algorithms for recognition and quantification of NFT in an array of
tauopathies. This will allow us to apply these trained networks for larger disease-specific
cohorts and to generate quantitative data for clinicopathological correlations, as well as for
molecular and genetic studies, and enable further diagnostic and therapeutic strategies.
Materials and methods
Case material
De-identified autopsy brain tissues were obtained from 22 representative individuals with
AD, primary age-related tauopathy (PART), PSP, and CTE [24] (Table 1). This cohort was a
convenience sample selected by the investigators. We used the following selection criteria:
(i) clinical/pathological: well-characterized clinical case, representative of a variety of
pathognomonic diagnostic histomorphological features, and with minimal or absent
neuropathological comorbidities; (ii) technical: adequately stained tissue with minimal or no
artifacts.
Immunohistochemistry
We used standard histological coronal sections from formalin-fixed paraffin-embedded
(FFPE) postmortem brain tissue, representing hippocampal formation and dorsolateral
prefrontal cortex. For PART and AD cases, the immunohistochemistry (IHC) of all cases
was performed at the University of Texas Southwestern (UTSW) using anti-phosphorylated
tau antibodies (AT8, Invitrogen, Waltham, MA) at 1:200 dilution using a Leica Bond III
automated immunostainer (Leica Microsystems, Buffalo Grove, IL). PSP and CTE cases
were immunostained at the Neuropathology Research Core at Mount Sinai with anti-
phosphorylated tau antibodies (AT8, Invitrogen) at 1:2000 dilution using a Ventana
autostainer (Roche Diagnostics, Rotkreuz, Switzerland).
Slide digitization
All sections were digitized to obtain digital WSI. For PSP and CTE, WSI were acquired
using the Ultra Fast Scanner Digital Pathology Slide Scanner (Philips, Amsterdam,
Netherlands), which scans histological samples mounted on standard glass slides at x40
Author Manuscript
Author Manuscript
Author Manuscript
Author Manuscript
Lab Invest. Author manuscript; available in PMC 2020 November 24.
Signaevsky et al. Page 5
magnification (0.25 µm/pixel) and saves them in the proprietary iSyntax format. For PART
and AD cases, all slides were scanned using an Aperio CS image scanner (Leica
Microsystems) at x20 magnification (0.50 µm/pixel) and saved in .svs format. All images in
proprietary formats were then converted into a GeoTIFF and stored on the server behind the
hospital firewall for interactive display over the intranet.
Pathological annotations
WSI were uploaded to the Precise Informatics Platform (PIP) developed by the Center for
Computational and Systems Pathology at Mount Sinai (MP, JK, JZ, and GF), which allows
for the management of thousands of images with pathologist annotations. Authors
previously have applied machine learning to prostate cancer for Gleason grading [23, 25],
and it is currently being used in our CLIA-approved laboratory. In addition, PIP enables
graphics processing unit (GPU)-accelerated DL for rapid validation and visualization of how
DL classifiers perform in different scenarios (brain regions, cell types, and staining).
Annotations were generated using the PIP collaborative web-based user interface for
outlining (Fig. 3). An NFT was operationalized as an object, i.e. “foreground”, with
cytoplasmic fine granular, coarse granular, or fibrillary/condensed AT8 immunopositivity
morphologically consistent with a neuron based on the histological context. In addition,
extracellular AT8-positive structures morphologically consistent with the neuronal
somatodendritic compartment were counted as ghost tangles. Partial neurites lacking
connection to the soma or hillock were excluded. Other AT8-positive structures including
neuropil threads, neuropil granules/grains, and ambiguous non-neuronal phospho-tau
staining were categorized as “background”. The total number of 22 WSI was divided into 14
for training and validation (model selection), with 8 reserved as a test set for performance
evaluation.
We conducted a concordance study to assess the inter-rater reliability using a custom
interface within the PIP platform. A total of 471 unique patches of mixed human expert-
annotated ground truth NFTs and AI-detected false positives were independently assessed by
three neuropathologists (MS, JFC, or CB) and compared using a Fleiss’ kappa statistic.
Fully convolutional network (FCN) training and model selection
The training dataset consisted of WSI of sections from 14 subjects (Table 1). In total, 178
representative rectangular regions of interests (ROI) were selected by the investigators for
analysis. The criteria for ROI were as follows: (1) a representative cortical area with an
adequate IHC of diagnostic quality, (2) a representative variety of recognizable distinct
histological AT8-stained elements, and (3) intact tissue without detachment or large tissue
folds. All NFT forms were computed together. The total number of AT8-positive NFTs of
various morphologies ranging from pre-tangles to mature NFTs and ghost tangles used for
fully convolutional neural network training and model selection was 2221. We further
extracted image patches of size 512 × 512 pixels at x20 by partitioning the ROIs. The total
number of patches was 3177, comprising 2414 from Aperio scanned PART and AD WSIs,
as well as 763 from Philips scanned CTE and PSP WSIs (Fig. 4). We further assigned 200
patches from this dataset to the validation set (for model selection), with the remainder used
for training a neural network classifier.
Author Manuscript
Author Manuscript
Author Manuscript
Author Manuscript
剩余21页未读,继续阅读
资源评论
百态老人
- 粉丝: 1664
- 资源: 2万+
上传资源 快速赚钱
- 我的内容管理 展开
- 我的资源 快来上传第一个资源
- 我的收益 登录查看自己的收益
- 我的积分 登录查看自己的积分
- 我的C币 登录后查看C币余额
- 我的收藏
- 我的下载
- 下载帮助
安全验证
文档复制为VIP权益,开通VIP直接复制
信息提交成功