6/9/2017 Using deep learning to delineate pathological correlates - - PowerPoint PPT Presentation

6/9/2017 1

Using deep learning to delineate pathological correlates of mammographic breast density from diagnostic image-guided breast biopsies

Maeve Mullooly PhD MPH Cancer Prevention Fellow Division of Cancer Epidemiology and Genetics maeve.mullooly@nih.gov

2

No disclosures

3

Molecular pathology of mammographic breast density (MBD)

MBD is one of the strongest risk factors for breast cancer; however the mechanisms underlying its relationship with risk are unclear As a large number of women with high MBD will not develop breast cancer, it is important to further understand histological characteristics of MBD to improve risk stratification Little is known about molecular underpinnings of MBD

4

Sherman ME, et al. Breast Disease; 2014 Khodr ZG, et al. CEBP; 2014

Molecular pathology of mammographic breast density (MBD)

6/9/2017 2

5

Image analysis of breast tissue histology can differentiate between epithelial and stromal morphological characteristics

Automated digital pathology has allowed additional in depth,

• bjective assessments of tissue

morphology Image based algorithm to identify unique morphologic phenotypes predictive of breast cancer prognosis

Beck AH, et al. Science Translational Medicine; 2011

6

Hypothesis: Assessment of breast tissue sections using automated machine learning approaches may characterize important features of mammographic breast density Objective: To identify breast tissue histologic features associated with mammographic breast density

7

• Cross-sectional molecular epidemiological study of MBD
• Women aged 40-65 years referred to diagnostic image-guided breast biopsy due

to abnormal mammogram (2007 – 2010; n=1,227 eligible women)

• Participant characteristics, biologic specimens (breast biopsy tissues)

Breast Cancer Surveillance Consortium

The NCI Breast Radiology Evaluation And Study of Tissues (BREAST) Stamp Project

Stamp Act Fund

Gierach G, et al. CEBP; 2014

Quantitative measures of breast density using Single X-ray Absorptiometry (SXA)

MBD assessment was conducted using craniocaudal views of pre-biopsy digital mammograms of the ipsilateral breast: MBD-volumetric: Overall volumetric MBD was assessed on pre-biopsy digital mammograms using a density phantom MBD-localized: Localized peri-lesional MBD was measured in a standard volume surrounding the biopsy target

8

Gierach G, et al. CEBP; 2014 Biopsy site

6/9/2017 3

9

Whole slide image (WSI) (H&E) MBD-volumetric MBD-localized

Utilizing H&E stained non-target breast biopsies to characterize histologic features

Gierach G, et al. CEBP; 2014 1. Trained model 2. Extracted features 3. Built MBD prediction model 4. Applied and assessed performance Training set (67%) (n=534) Testing set (33%) (n=264) Analytical population (n=798)

Overview of study design

Training set (67%) (n=534) Testing set (33%) (n=264) Analytical population (n=798)

12

Whole slide image Convolutional neural network model

Bejnordi BE, et al. IEEE International Symposium on Biomedical Imaging (ISBI) conference proceedings; 2017 Training set (67%) (n=534) 1. Trained and applied model to whole slide images to generate map of epithelium, stroma and fat (training set, n=534 women) Analytical population (n=798)

6/9/2017 4

13

• 1. An example of the CNN output classification of biopsy breast

epithelial, stroma and fat tissue

14

Feature Extraction Whole slide image Convolutional neural network model

1. Trained and applied model to whole slide images to generate map of epithelium, stroma and fat (training set, n=534 women) 2. 37 features were extracted from model output Training set (67%) (n=534) Analytical population (n=798)

15

• 2. A total of 37 features were extracted from the model output

Feature Category Feature Definition Global tissue amount Total area of epithelium, stroma, and fat and normalized areas of each tissue class by total tissue amount Morphology Statistics of area and eccentricity of epithelial regions Spatial arrangement of the epithelial regions Statistics of number of neighbors for each node and distances between nodes (Delaunay Triangulation) Statistics of areas of Voronoi cells, and the area ratio between actual epithelial region and its zone of influence (ZOI) (Area-Voronoi diagram)

16

• 2. Examples of features extracted from the CNN: global tissue

amount

6/9/2017 5

17

• 2. Examples of features extracted from the CNN: spatial

arrangement of the epithelial regions

Epithelial structures very close to each other (low) Epithelial structures far from each other (high) node_dist_mean: distances of each epithelial area to all it's neighboring ones based on Delaunay Triangulation

18

• 2. Examples of features extracted from the CNN: spatial

arrangement of the epithelial regions

Each epithelial area has high average number of neighbors num_neighbors_mean: Mean of the average number of neighbors each epithelial area has based on Delaunay Triangulation Each epithelial area has low average number of neighbors 1. Trained and applied model to whole slide images to generate map of epithelium, stroma and fat (training set, n=534 women) 2. 37 features were extracted from model output 3. A random forest regressor was trained using these features to predict MD measurements 4. Applied model (step 3) and assessed performance Training set (67%) (n=534) Testing set (33%) (n=264) Analytical population (n=798)

Feature Extraction Whole slide image Convolutional neural network model

Features most strongly associated with volumetric MBD (%FGV)

*Spearman correlation between predicted and measured MBD in testing set

6/9/2017 6

Features most strongly associated with localized peri-lesional MBD (%FGV)

*Spearman correlation between predicted and measured MBD in testing set

22

Summary and next steps

Automated image analysis identified features of breast tissue predictive of

• verall and localized MBD

Greater stromal tissue amount and spatial distribution patterns of epithelial regions had strong associations with MBD Ongoing work: investigate stromal properties of the tissue determine histological features that distinguish benign from malignant tissue in high vs. low MBD using CNN modeling

23

Implications of this study

Expanding the application of automated image analysis to diagnostic breast biopsy specimens may increase understanding of the underlying biology of MBD Continued improvements in CNN modeling may allow for extraction of novel breast cancer biomarkers from archival H&E breast tissue sections Identification of unique breast tissue features that can predict breast cancer risk among women with high and low breast density will be important to improve breast cancer risk prediction among the increasing number of women undergoing diagnostic breast biopsy

24

Acknowledgements

Radboud University, Netherlands

Babak Ehteshami Bejnordi Jeroen van der Laak

National Cancer Institute

Gretchen Gierach Ruth Pfeiffer Sharon Fan Maya Palakal Manila Hada Louise Brinton

Mayo Clinic

Mark Sherman

Harvard Medical School

Andrew Beck (PathAI)

BREAST Stamp Project UCSF

John Shepherd Jeff Wang

UVM

Berta Geller Pam Vacek Donald Weaver Brian Sprague UVM Fletcher Allen Health Care Sally Herschorn Jason Johnson

6/9/2017 7

cpfp.cancer.gov www.cancer.gov

maeve.mullooly@nih.gov

26 27

Characteristics of study participants

28

Characteristics of study participants

6/9/2017 8

29

Training (n=534) Testing (n=264) Mean (SD) Mean (SD) Age at mammogram (mean, years) 51 (SD: 6.8) 51 (SD: 6.8) BMI (mean, kg/m2) 26.8 (SD: 6.3) 26.6 (SD: 6.3) Age at menopause (mean, years) 48 (SD: 6.7) 48 (SD: 7.3) n (%) n (%) White race 488 (91) 242 (92) Premenopausal 293 (58) 146 (58) Parous 408 (77) 205 (79) MHT users 79 (15) 35 (14)

Characteristics of study participants

*BMI: body mass index; MHT: menopausal hormone therapy

30 31

Features most strongly associated with overall volumetric percent MD

32

Features most strongly associated with localized MD

6/9/2017 9

33

Relationship between CNN features and MBD measurements

Volumetric MBD Localized MBD Feature Definition P-score P-score Stroma amount Area of stroma normalized to total tissue area >0.90 >0.90 Epithelial amount Area of epithelial tissue normalized to total tissue area 0.01 0.25 Node distance Average distances of each epithelial area to all of it's neighboring ones based on Delaunay Triangluation 0.58 0.75 Vor ZOI area Area ratio of each epithelial region to its Voronoi region 0.26 0.53

Assessed using lasso regression

Global tissue amount: Delaunay Triangulation: Area-Voronoi diagram:

34

Vor_area mean: Mean area of the Voronoi cells

• 2. Examples of features extracted from the CNN: spatial

arrangement of the epithelial regions

Epithelial structures: far from each other (high) Epithelial structures: small and close to each other (low)

35

Classification of biopsy breast tissue into epithelium, stroma and fat

36

num_neighbors_mean: Mean of the average number of neighbors each epithelial area has based on Delaunay triangulation

• 2. Examples of features extracted from the CNN: spatial

arrangement of the epithelial regions

6/9/2017 10

37

Training (n=534) Testing (n=264) Mean (SD) Mean (SD) Age at mammogram (years) 51 (SD: 6.8) 51 (SD: 6.8) BMI (kg/m2) 26.8 (SD: 6.3) 26.6 (SD: 6.3) Volumetric MBD (%FGV) 38 (21) 38 (35) Localized MBD (%FGV) 43 (24) 47 (25) n (%) n (%) Pathologic diagnosis Benign 168 (32) 92 (35) Proliferative 204 (38) 101 (38) Proliferative with atypia 44 (8) 13 (5) In-situ (LCIS or DCIS) 47 (9) 28 (11) Invasive 71 (13) 30 (11)

Characteristics of study participants

*BMI: body mass index; FGV: fibroglandular volume