Estimation of Breast Percent Density from Digital Breast Tomosynthesis Images

1
Estimation of Breast Percent Density from Digital
Breast Tomosynthesis Images

• Raymond J. Acciavatti
• University of Pennsylvania

2
Risks Factors for Breast Cancer

• Standard Risk Factors for Breast Cancer
• Age
• Age at menarche
• Age at birth of first live child
• Number of previous benign biopsies
• Number of 1st degree relatives with breast cancer
  
  

• Breast density is an independent risk factor.
• There is a 4- to 6-fold relative risk of cancer
  in very dense breasts compared to mainly adipose
  breasts.
• Mammographically, density is quantified as
  percent density (PD), percentage of total breast
  area occupied by dense tissue.

3
Digital Breast Tomosynthesis, DBT
Compression Plate
Breast
Detector

• We compare PD in DBT and DM.

4
PD Estimation by Thresholding

• Cumulus software

DM
5
PD Estimation by Thresholding

• Cumulus software
• Manual segmentation of pectoral muscle

DM
6
PD Estimation by Thresholding

• Cumulus software
• Manual segmentation of pectoral muscle
• Thresholding of
• Breast outline

DM
7
PD Estimation by Thresholding

• Cumulus software
• Manual segmentation of pectoral muscle
• Thresholding of
• Breast outline
• Dense tissue

DM
8
PD Estimation by Thresholding

• Cumulus software
• Manual segmentation of pectoral muscle
• Thresholding of
• Breast outline
• Dense tissue

PDMAreaDense/AreaBreast
DM
9
PD Analysis of DBT Reconstructed Images

• PD definition extended to volumetric DBT

10
PD Analysis of DBT Reconstructed Images

• PD definition extended to volumetric DBT
• DBT uses a limited number of projections, which
  introduces reconstruction artifacts, e.g.
  out-of-focus densities.

11
PD Analysis of DBT Reconstructed Images

• Manual analysis of all DBT slices is
  time-prohibitive.
• We have fitted thresholds through a small number
  of manually selected slices.
• We tested fitting methods using simulated DBT
  phantom images.
• Five methods were selected and applied on
  clinical images of 10 women with previously
  diagnosed or suspected cancer.

Central Phantom DBT Slice
Central Clinical DBT Slice
12
Five Methods of Dense Tissue Segmentation

• Average thresholds from 30, 50, and 70 slice.
• Quadratic fit through thresholds from 30, 50,
  70 slice.
• Quadratic fit through thresholds from 10, 50,
  90 slice.
• Quadratic fit through thresholds from 10, 30,
  50, 70, 90 slice.
• Quadratic fit using five slices, with PD forced
  to 0 at breast edge.

13
Five Methods of Dense Tissue Segmentation

• Average thresholds from 30, 50, and 70 slice.
• Quadratic fit through thresholds from 30, 50,
  70 slice.
• Quadratic fit through thresholds from 10, 50,
  90 slice.
• Quadratic fit through thresholds from 10, 30,
  50, 70, 90 slice.
• Quadratic fit using five slices, with PD forced
  to 0 at breast edge.

14
Five Methods of Dense Tissue Segmentation

• Average thresholds from 30, 50, and 70 slice.
• Quadratic fit through thresholds from 30, 50,
  70 slice.
• Quadratic fit through thresholds from 10, 50,
  90 slice.
• Quadratic fit through thresholds from 10, 30,
  50, 70, 90 slice.
• Quadratic fit using five slices, with PD forced
  to 0 at breast edge.

15
Five Methods of Dense Tissue Segmentation

• Average thresholds from 30, 50, and 70 slice.
• Quadratic fit through thresholds from 30, 50,
  70 slice.
• Quadratic fit through thresholds from 10, 50,
  90 slice.
• Quadratic fit through thresholds from 10, 30,
  50, 70, 90 slice.
• Quadratic fit using five slices, with PD forced
  to 0 at breast edge.

16
Five Methods of Dense Tissue Segmentation

• Average thresholds from 30, 50, and 70 slice.
• Quadratic fit through thresholds from 30, 50,
  70 slice.
• Quadratic fit through thresholds from 10, 50,
  90 slice.
• Quadratic fit through thresholds from 10, 30,
  50, 70, 90 slice.
• Quadratic fit using five slices, with PD forced
  to 0 at breast edge.

17
Statistical Analysis

• We validated the results of our fitting methods
  by computing
• Relative error in PDT,3D compared to the ground
  truth in the phantoms.
• Pearson correlation coefficient

• Kappa coefficient
• The Wilcoxon Signed-Rank test of statistically
  significant difference between the fitting
  methods.

18
Phantom Results
Relative Error ()
19
Phantom Results
Relative Error ()
20
Phantom Segmentation Image Results
Segmented Image
Central Slice
21
Clinical Segmentation Image Results
Central Slice
Segmented Image
22
Clinical Results
Global Threshold y
1.38x 0.06 , R2 0.65 Quad. . Three Central
Slices y1.22x 0.09, R2
0.52 Quad., Three Spaced Slices y
1.37x 0.11, R2 0.60 Quad., Five Slices
y 1.35x - 0.11, R2
0.60 Quad., PD0 at Breast Edges y 1.25x
0.10, R2 0.61
PDT,3D
PDM
23
Statistical Comparisons of DBT with DM
We observed substantial agreement between PDT,3D
and PDM.
24
Statistical Comparisons of DBT with DM

• Wilcoxon Signed-Rank Test Results
• No fitting methods are statistically different
  from each other except

25
Statistical Comparisons of DBT with DM

• Wilcoxon Signed-Rank Test Results
• No fitting methods are statistically different
  from each other except
• Global thresholding vs. all other methods
  (plt0.005).

26
Statistical Comparisons of DBT with DM

• Wilcoxon Signed-Rank Test Results
• No fitting methods are statistically different
  from each other except
• Global thresholding vs. all other methods
  (plt0.005)
• Quad., 3 central slices vs. Quad., w/ PD0 at
  breast edges (plt0.01).

27
Statistical Comparisons of DBT with DM

• Wilcoxon Signed-Rank Test Results
• No fitting methods are statistically different
  from each other except
• Global thresholding vs. all other methods
  (plt0.005)
• Quad., 3 central slices vs. Quad., w/ PD0 at
  breast edges (plt0.01).
• Quad., 5 slices vs. Quad., w/ PD0 at breast
  edges (plt0.01).

28
Conclusions and Future Direction

• Substantial agreement between PDT,3D and PDM was
  observed, suggesting that PD is robust to
  changes in acquisition.
• Due to limitations of DBT reconstruction, no
  final solution for dense tissue segmentation is
  presented.
• In future months, the breast cancer research lab
  will investigate more complex segmentation
  techniques, including analysis of local
  statistics.

29
Acknowledgment

• Thank you to my sponsor Dr. Predrag Bakic for his
  thorougness and expertise.
• Thank you to Dr. Cuping Zhang for preparing the
  phantom images, and to Dr. Despina Kontos for
  providing DBT and DM comparison images.
• Thank you to the lab of Dr. Maidment for offering
  feedback in lab meeting.

30
Thank you for your attention!