6/9/2017 Using deep learning to delineate pathological correlates of mammographic breast density from diagnostic image-guided breast biopsies No disclosures Maeve Mullooly PhD MPH Cancer Prevention Fellow Division of Cancer Epidemiology and Genetics maeve.mullooly@nih.gov 2 Molecular pathology of mammographic breast density (MBD) 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 Sherman ME, et al. Breast Disease; 2014 3 4 Khodr ZG, et al. CEBP; 2014 1
6/9/2017 Image analysis of breast tissue histology can differentiate � Hypothesis: Assessment of breast tissue sections using between epithelial and stromal morphological characteristics automated machine learning approaches may characterize important features of mammographic breast density � Automated digital pathology has allowed additional in depth, objective assessments of tissue � Objective: To identify breast tissue histologic features associated morphology with mammographic breast density � Image based algorithm to identify unique morphologic phenotypes predictive of breast cancer prognosis 5 6 Beck AH, et al. Science Translational Medicine; 2011 Quantitative measures of breast density using Single X-ray The NCI Breast Radiology Evaluation And Study of Absorptiometry (SXA) Tissues (BREAST) Stamp Project Stamp Act Fund Biopsy site Breast Cancer Surveillance Consortium MBD assessment was conducted using craniocaudal views of pre-biopsy digital mammograms of the ipsilateral breast: Cross-sectional molecular epidemiological study of MBD � MBD-volumetric: Overall volumetric MBD was assessed on pre-biopsy digital Women aged 40-65 years referred to diagnostic image-guided breast biopsy due � mammograms using a density phantom to abnormal mammogram (2007 – 2010; n=1,227 eligible women) MBD-localized: Localized peri-lesional MBD was measured in a standard volume Participant characteristics, biologic specimens (breast biopsy tissues) � surrounding the biopsy target Gierach G, et al. CEBP; 2014 8 7 Gierach G, et al. CEBP; 2014 2
6/9/2017 Overview of study design Utilizing H&E stained non-target breast biopsies to characterize histologic features Analytical population (n=798) Training set (67%) Testing set (33%) (n=534) (n=264) 1. Trained model 4. Applied and assessed 2. Extracted features performance Whole slide image (WSI) 3. Built MBD prediction model (H&E) MBD-volumetric MBD-localized 9 Gierach G, et al. CEBP; 2014 Analytical population Analytical population (n=798) (n=798) Training set (67%) Testing set (33%) Training set (67%) (n=264) (n=534) (n=534) 1. Trained and applied model to whole slide images to generate map of epithelium, stroma and fat (training set, n=534 women) Convolutional neural network model Whole slide image Bejnordi BE, et al. IEEE International Symposium on 12 Biomedical Imaging (ISBI) conference proceedings; 2017 3
6/9/2017 1. An example of the CNN output classification of biopsy breast Analytical population epithelial, stroma and fat tissue (n=798) 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) 2. 37 features were extracted from model output Feature Extraction Convolutional neural network model Whole slide image 13 14 2. Examples of features extracted from the CNN: global tissue 2. A total of 37 features were extracted from the model output amount 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 Statistics of number of neighbors for each node and distances between nodes (Delaunay Spatial arrangement of Triangulation) the epithelial regions Statistics of areas of Voronoi cells, and the area ratio between actual epithelial region and its zone of influence (ZOI) (Area-Voronoi diagram) 15 16 4
6/9/2017 2. Examples of features extracted from the CNN: spatial 2. Examples of features extracted from the CNN: spatial arrangement of the epithelial regions arrangement of the epithelial regions Epithelial structures very Epithelial structures far Each epithelial area has high Each epithelial area has low close to each other (low) from each other (high) average number of neighbors average number of neighbors node_dist_mean: distances of each epithelial area to all num_neighbors_mean: Mean of the average number of neighbors it's neighboring ones based on Delaunay Triangulation each epithelial area has based on Delaunay Triangulation 17 18 Features most strongly associated with volumetric MBD (%FGV) Analytical population (n=798) Training set (67%) Testing set (33%) (n=534) (n=264) 1. Trained and applied model to whole slide images to generate map of epithelium, stroma and fat 4. Applied model (step 3) and (training set, n=534 women) assessed performance 2. 37 features were extracted from model output 3. A random forest regressor was trained using these features to predict MD measurements Feature Extraction Convolutional neural network model Whole slide image *Spearman correlation between predicted and measured MBD in testing set 5
6/9/2017 Features most strongly associated with localized peri-lesional Summary and next steps MBD (%FGV) � Automated image analysis identified features of breast tissue predictive of overall 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 *Spearman correlation between predicted and measured MBD in testing set 22 Acknowledgements Implications of this study Radboud University, Netherlands BREAST Stamp Project � Babak Ehteshami Bejnordi UCSF � Expanding the application of automated image analysis to diagnostic breast � Jeroen van der Laak � John Shepherd � Jeff Wang biopsy specimens may increase understanding of the underlying biology of UVM MBD National Cancer Institute � Berta Geller � Gretchen Gierach � Pam Vacek � Ruth Pfeiffer � Donald Weaver � Sharon Fan � Continued improvements in CNN modeling may allow for extraction of novel � Brian Sprague � Maya Palakal breast cancer biomarkers from archival H&E breast tissue sections UVM Fletcher Allen Health Care � Manila Hada � Sally Herschorn � Louise Brinton � Jason Johnson � Identification of unique breast tissue features that can predict breast cancer Mayo Clinic risk among women with high and low breast density will be important to � Mark Sherman improve breast cancer risk prediction among the increasing number of women undergoing diagnostic breast biopsy Harvard Medical School � Andrew Beck (PathAI) 23 24 6
6/9/2017 maeve.mullooly@nih.gov cpfp.cancer.gov www.cancer.gov 26 Characteristics of study participants Characteristics of study participants 27 28 7
6/9/2017 Characteristics of study participants 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) *BMI: body mass index; MHT: menopausal hormone therapy 29 30 Features most strongly associated with overall volumetric percent MD Features most strongly associated with localized MD 31 32 8
6/9/2017 2. Examples of features extracted from the CNN: spatial Relationship between CNN features and MBD measurements arrangement of the epithelial regions Epithelial structures: far from Epithelial structures: small and Volumetric MBD Localized MBD each other (high) close to each other (low) Feature Definition P-score P-score Global tissue amount: Stroma amount Area of stroma normalized to >0.90 >0.90 total tissue area Epithelial amount Area of epithelial tissue 0.01 0.25 normalized to total tissue area Delaunay Triangulation: Node distance Average distances of each 0.58 0.75 epithelial area to all of it's neighboring ones based on Delaunay Triangluation Area-Voronoi diagram: Vor ZOI area Area ratio of each epithelial 0.26 0.53 region to its Voronoi region Vor_area mean: Mean area of the Voronoi cells Assessed using lasso regression 33 34 Classification of biopsy breast tissue into epithelium, stroma 2. Examples of features extracted from the CNN: spatial and fat arrangement of the epithelial regions num_neighbors_mean: Mean of the average number of neighbors each epithelial area has based on Delaunay triangulation 35 36 9
Recommend
More recommend
