Deep learning assessment of breast terminal duct lobular unit involution: Towards automated prediction of breast cancer risk

Abstract

Terminal duct lobular unit (TDLU) involution is the regression of milk-producing structures in the breast. Women with less TDLU involution are more likely to develop breast cancer. A major bottleneck in studying TDLU involution in large cohort studies is the need for labor-intensive manual assessment of TDLUs. We developed a computational pathology solution to automatically capture TDLU involution measures. Whole slide images (WSIs) of benign breast biopsies were obtained from the Nurses' Health Study. A set of 92 WSIs was annotated for acini, TDLUs and adipose tissue to train deep convolutional neural network (CNN) models for detection of acini, and segmentation of TDLUs and adipose tissue. These networks were integrated into a single computational method to capture TDLU involution measures including number of TDLUs per tissue area, median TDLU span and median number of acini per TDLU. We validated our method on 40 additional WSIs by comparing with manually acquired measures. Our CNN models detected acini with an F1 score of 0.73±0.07, and segmented TDLUs and adipose tissue with Dice scores of 0.84±0.13 and 0.87±0.04, respectively. The inter-observer ICC scores for manual assessments on 40 WSIs of number of TDLUs per tissue area, median TDLU span, and median acini count per TDLU were 0.71, 0.81 and 0.73, respectively. Intra-observer reliability was evaluated on 10/40 WSIs with ICC scores of >0.8. Inter-observer ICC scores between automated results and the mean of the two observers were: 0.80 for number of TDLUs per tissue area, 0.57 for median TDLU span, and 0.80 for median acini count per TDLU. TDLU involution measures evaluated by manual and automated assessment were inversely associated with age and menopausal status. We developed a computational pathology method to measure TDLU involution. This technology eliminates the labor-intensiveness and subjectivity of manual TDLU assessment, and can be applied to future breast cancer risk studies.

Fig 1. Examples of annotations for acini (A; annotated by blue squares), terminal duct lobular units (B) and adipose tissue (C).

Fig 2. Results of the acini detection (A), terminal duct lobular unit (B), and adipose tissue (C) segmentation algorithms.

The original images are in the left column, the middle column shows ground truth as annotated by human observers, and the detections and segmentations performed by the automated method are displayed in the right column.

Fig 3. Results of the acini detection, terminal duct lobular unit, and adipose tissue segmentation algorithms (B) overlaid on the original image (A).

Fig 4. Scatterplots of the association of quantitative terminal ductal lobular unit (TDLU) involution measures and age.

TDLU count per tissue area assessed using manual (A) and automated (B) method were significantly inversely correlated with age (p<0.01). Median TDLU span assessed manually (C) and with the automated method (D) was significantly inversely correlated with age (p<0.01 and p = 0.01). Median acini count per TDLU assessed using manual (E) and automated (F) assessment was also significantly inversely correlated with age (p<0.01). Acini count per tissue area assessed by the automated method was significantly inversely correlated with age (G; p<0.01). Median TDLU area assessed by the automated method was significantly inversely correlated with age (H; p<0.01).

Fig 5. Boxplots demonstrating the association of qualitative terminal ductal lobular unit involution measures and age.

(A) Women with predominantly type 1 lobules were significantly older than women with predominantly type 2 lobules (manual method: p<0.01; automated method: p = 0.01). No woman presented with predominately type 3 lobules. (B) Women with “Predominantly type 1, no type 3” lobules were significantly older than women with “Mixed lobules” (manual method p<0.01; automated method p<0.01). No woman was assessed as having “No type 1” lobules by the automated method. The manual qualitative measures were obtained by consensus vote. The boxplots show the median value, interquartile range (IQR), and 5th and 95th whiskers.