Using Whole Slide Image Representations from Self-Supervised Contrastive Learning for Melanoma Concordance Regression

2025-05-06 0 0 3.5MB 15 页 10玖币

侵权投诉

Using Whole Slide Image Representations from

Self-Supervised Contrastive Learning for

Melanoma Concordance Regression ⋆

Sean Grullon1, Vaughn Spurrier1, Jiayi Zhao1, Corey Chivers1, Yang Jiang1,

Kiran Motaparthi2, Jason Lee3, Michael Bonham1, and Julianna Ianni1

1Proscia, Inc. Philadelphia, United States

2Department of Dermatology, University of Florida College of Medicine

3Department of Dermatology, Sidney Kimmel Medical College at Thomas Jeﬀerson

University

Abstract. Although melanoma occurs more rarely than several other

skin cancers, patients’ long term survival rate is extremely low if the

diagnosis is missed. Diagnosis is complicated by a high discordance rate

among pathologists when distinguishing between melanoma and benign

melanocytic lesions. A tool that provides potential concordance informa-

tion to healthcare providers could help inform diagnostic, prognostic, and

therapeutic decision-making for challenging melanoma cases. We present

a melanoma concordance regression deep learning model capable of pre-

dicting the concordance rate of invasive melanoma or melanoma in-situ

from digitized Whole Slide Images (WSIs). The salient features corre-

sponding to melanoma concordance were learned in a self-supervised

manner with the contrastive learning method, SimCLR. We trained a

SimCLR feature extractor with 83,356 WSI tiles randomly sampled from

10,895 specimens originating from four distinct pathology labs. We trained

a separate melanoma concordance regression model on 990 specimens

with available concordance ground truth annotations from three pathol-

ogy labs and tested the model on 211 specimens. We achieved a Root

Mean Squared Error (RMSE) of 0.28 ±0.01 on the test set. We also in-

vestigated the performance of using the predicted concordance rate as a

malignancy classiﬁer, and achieved a precision and recall of 0.85 ±0.05

and 0.61 ±0.06, respectively, on the test set. These results are an impor-

tant ﬁrst step for building an artiﬁcial intelligence (AI) system capable

of predicting the results of consulting a panel of experts and delivering a

score based on the degree to which the experts would agree on a partic-

ular diagnosis. Such a system could be used to suggest additional testing

or other action such as ordering additional stains or genetic tests.

⋆Acknowledgments: The authors thank the support of Jeﬀ Baatz, Ramachandra V.

Chamarthi, Nathan Langlois, and Liren Zhu at Proscia for their engineering sup-

port; Theresa Feeser, Pratik Patel, and Aysegul Ergin Sutcu at Proscia for their

data acquisition and Q&A support; and Dr. Curtis Thompson at CTA and Dr.

David Terrano at Bethesda Dermatology Laboratory for their consensus annotation

support.

arXiv:2210.04803v1 [cs.CV] 10 Oct 2022

2 S. Grullon et al.

Keywords: self supervised learning, contrastive learning, melanoma,

weak supervision, multiple instance learning, digital pathology

1 Introduction

More than 5 million diagnoses of skin cancer are made each year in the United

States, about 106,000 of which are melanoma of the skin [1]. Diagnosis requires

microscopic examination of hematoxylin and eosin (H&E) stained, paraﬃn wax

embedded biopsies of skin lesion specimens on glass slides. These slides can be

manually observed under a microscope, or digitally on a Whole Slide Image

(WSI) scanned on specialty hardware. The 5-year survival rate of patients with

metastatic malignant melanoma is less than 20% [15]. Melanoma occurs more

rarely than several other types of skin cancer, and its diagnosis is challeng-

ing, as evidenced by a high discordance rate among pathologists when distin-

guishing between melanoma and benign melanocytic lesions (∼40% discordance

rate; e.g. [10], [7]). The high discordance rate highlights that greater scrutiny is

likely needed to arrive at an accurate melanoma diagnosis, however patients re-

ceive diagnoses only from a single dermatopathologist in many instances. This

tends to increase the probability of misdiagnosis, where frequent over-diagnosis

of melanocytic lesions results in severe costs to a clinical practice and addi-

tional costs and distress to patients. [26]. In this scenario, the decision-making

of the single expert would be further informed by knowledge of a likely con-

cordance level among a group of multiple experts in a given case under con-

sideration. Additional methods of providing concordance information to health-

care providers could help further inform diagnostic, prognostic, and therapeutic

decision-making for challenging melanoma cases. A method capable of predicting

the results of consulting a panel of experts and delivering a score based on the

degree to which the experts would agree on a particular diagnosis would help

reduce melanoma misdiagnosis and subsequently improve patient care.

The advent of digital pathology has brought the revolution in machine learn-

ing and artiﬁcial intelligence to bear on a variety of tasks common to pathology

labs. Campanella et al. [2] trained a model in a weakly-supervised framework

that did not require pixel-level annotations to classify prostate cancer and val-

idated on ∼10,000 WSIs sourced from multiple countries. This represented a

considerable advancement towards a system capable of use in clinical practice

for prostate cancer. However, some degree of human-in-the-loop curation was

performed on their data set, including manual quality control such as post-hoc

removal of slides with pen ink from the study. Pantanowitz et al. [17] used pixel-

wise annotations to develop a model trained on ∼550 WSIs that distinguishes

high-grade from low-grade prostate cancer. In dermatopathology, the model de-

veloped in [22] classiﬁed skin lesion specimens between six morphology-based

groups (including melanoma), was tested on ∼5099 WSIs, provided automated

quality control to remove WSI patches with pen ink or blur, and also demon-

strated that use of conﬁdence thresholding could provide a high accuracy.

Melanoma Concordance Regression 3

The recent application of deep learning to digital pathology has predomi-

nately leveraged the use of pre-trained WSI tile representations, usually obtained

by using feature extractors pre-trained on the ImageNet [6] data set. The features

learned by such pre-training is dominated by features present in natural-scene

images, which are not guaranteed to generalize to histopathology images. Such

representations can limit the reported performance metrics and aﬀect model ro-

bustness. It has been shown in [13] that self-supervised pre-training on WSIs

improved the downstream performance in identifying metastastic breast cancer.

In this work, we present a deep learning regression model capable of predict-

ing from WSIs the concordance rate of consulting a panel of experts on rendering

a case diagnosis of invasive melanoma or melanoma in-situ. The deep learn-

ing model learns meaningful feature representations from WSIs through self-

supervised pre-training, which are used to learn the concordance rate through

weakly-supervised training.

2 Methods

2.1 Data Collection and Characteristics

The melanoma concordance regression model was trained and evaluated on 1,412

specimens (consisting of 1,722 WSIs) from three distinct pathology labs. The

ﬁrst lab consists for 611 suspected melanoma specimens from a leading der-

matopathology lab in a top academic medical center (Department of Derma-

tology at University of Florida College of Medicine), denoted as University of

Florida. The second lab consisted of 605 suspected melanoma specimens dis-

tributed across North America, but re-scanned at The Department of Der-

matology at University of Florida College of Medicine, denoted as Florida -

External. The third lab consisted of 319 suspected specimens from Jeﬀerson

Dermatopathology Center, Department of Dermatology & Cutaneous Biology,

Thomas Jeﬀerson University denoted as Jeﬀerson. The WSIs consisted exclu-

sively of H&E-stained, formalin-ﬁxed, paraﬃn-embedded dermatopathology tis-

sue and were all scanned using a 3DHistech P250 High Capacity Slide Scanner

at an objective power of 20X, corresponding to 0.24µm/pixel. The diagnostic

categories present in our data set are summarized in Table 1.

The annotations for our data set were provided by at least three board-

certiﬁed pathologists who reviewed each melanocytic specimen. The ﬁrst review

was the original specimen diagnosis made via glass slide examination under a

microscope. At least two and up to four additional dermatopathologists indepen-

dently reviewed and rendered a diagnosis digitally for each melanocytic speci-

men. The patient’s year of birth and gender were provided with each specimen

upon review. Two dermatopathologists from the United States reviewed all 1,412

specimens in our data set and up to two additional dermatopathologists reviewed

a subset of our data set. A summary of the number of concordant reviews in this

study is given in Table 2.

The concordance reviews are converted to a concordance rate by calculating

the fraction of dermatopathologists who rendered a diagnosis of melanoma in-situ

文档加载中……请稍候！
如果长时间未打开，您也可以点击刷新试试。

下载文档到电脑，查找使用更方便

10 玖币 0人已下载

立即下载

摘要：

UsingWholeSlideImageRepresentationsfromSelf-SupervisedContrastiveLearningforMelanomaConcordanceRegression⋆SeanGrullon1,VaughnSpurrier1,JiayiZhao1,CoreyChivers1,YangJiang1,KiranMotaparthi2,JasonLee3,MichaelBonham1,andJuliannaIanni11Proscia,Inc.Philadelphia,UnitedStates2DepartmentofDermatology,Univers...

展开>> 收起<<

Using Whole Slide Image Representations from Self-Supervised Contrastive Learning for Melanoma Concordance Regression.pdf

共15页,预览3页

还剩页未读，继续阅读

声明：本站为文档C2C交易模式，即用户上传的文档直接被用户下载，本站只是中间服务平台，本站所有文档下载所得的收益归上传人(含作者)所有。玖贝云文库仅提供信息存储空间，仅对用户上传内容的表现方式做保护处理，对上载内容本身不做任何修改或编辑。若文档所含内容侵犯了您的版权或隐私，请立即通知玖贝云文库，我们立即给予删除！

Using Whole Slide Image Representations from Self-Supervised Contrastive Learning for Melanoma Concordance Regression

相关推荐

开通VIP享超值会员特权

作者详情

相关内容

热门标签

举报选择: