cbig
play

CBIG Computational Breast Imaging Group Quantitative imaging - PowerPoint PPT Presentation

Jan 20, 2024 •686 likes •1.25k views

CBIG Computational Breast Imaging Group Quantitative imaging phenotyping of breast cancer risk Aimilia Gastounioti, PhD Department of Radiology University of Pennsylvania 8th International Breast Density and Cancer Risk Assessment Workshop

  1. CBIG Computational Breast Imaging Group Quantitative imaging phenotyping of breast cancer risk Aimilia Gastounioti, PhD Department of Radiology University of Pennsylvania 8th International Breast Density and Cancer Risk Assessment Workshop June 8th, 2017

  2. Nothing to Disclose 8th International Breast Density and Cancer Risk Assessment Workshop June 8th, 2017

  3. Toward Precision Cancer Screening Shieh et al. ( Nat Rev Clin Oncol. 2016 8th International Breast Density and Cancer Risk Assessment Workshop June 8th, 2017

  4. Need for More Accurate Ways of Predicting Breast Cancer Risk The key role of imaging phenotypes 8th International Breast Density and Cancer Risk Assessment Workshop June 8th, 2017

  5. Wolfe’s Parenchymal Patterns DY NI PI P2 Lowest risk Highest risk Wolfe AJR 1976 8th International Breast Density and Cancer Risk Assessment Workshop June 8th, 2017

  6. Breast Density & Risk PD = 0% PD < 10% PD < 25% PD < 50% PD < 75% PD < 100% Breast Percent Density (PD%) Boyd et al. N Engl J Med. 2007 8th International Breast Density and Cancer Risk Assessment Workshop June 8th, 2017

  7. Breast Density & Risk Cumulus Established, independent risk factor McCormack et al . Cancer Epidemiol Biomarkers Prev. 2006 Eng et al. Breast Cancer Res. 2014 Sherratt et al. Breast Cancer Res. 2016 Improves risk assessment Quantra models Has shared genetic basis with Brentnall et al. Breast Cancer Res. 2015 breast cancer susceptibility Tice et al. Ann Intern Med. 2008 Volpara Stone et al. Cancer Res. 2015 Lindström et al. Nat Commun. 2014 Predicts both inherent risk and masking risk Krishnan et al. Breast Cancer Res. 2016 Strand et al. Int J Cancer 2017 Associated with tumor profile Bertrand et al. Cancer Epidemiol Biomarkers Prev. 2015 LIBRA 8th International Breast Density and Cancer Risk Assessment Workshop June 8th, 2017

  8. PD = 31% PD = 31% BIRADS = 3 BIRADS = 2 Gail 5 Yr = 0.7% Gail 5 Yr = 7.6% Gail Life = 3.6% Gail Life = 20.7% 8th International Breast Density and Cancer Risk Assessment Workshop June 8th, 2017

  9. Beyond Breast Density: Texture Features for Pattern Analysis Spatial relationship among gray levels 0 O 45 O 90 O 135 O 0.17 0 0 0 1 0 2 0 0 Low contrast 0 0.5 0 0 1 1 1 1 1 0.17 0 0 0 0 0 3 1 1 0 0.17 0 0 0 0 Gray-level co-occurrence matrix Run-length matrix 4 gray-level image for 0 O for 0 O High contrast Gastounioti et al. Breast Cancer Research 2016 (Review)

  10. Beyond Breast Density: Texture Features for Pattern Analysis Intrinsic patterns of image intensity Gray-level intensity distribution (texture roughness) Gastounioti et al. Breast Cancer Research 2016 (Review)

  11. Parenchymal texture patterns are indicative of genetic risk markers (BRCA1/2) Digitized Digital film mammograms mammograms Huo et al. Radiology 2002 Li et al . J Med Imag. 2014 Gastounioti et al. Breast Cancer Research 2016 (Review)

  12. Parenchymal texture patterns are predictive of cancer-case-control status Texture Feature OR (95% CI) Model adjusted for Age, BMI and breast PD Laws 1.27 (1.06, 1.54) Heine et al. J Natl Cancer Inst. 2012 Wei et al . Radiology 2011 Markovian 1.26 (1.07, 1.47) Run Length 1.26 (1.03, 1.54) Wavelet 1.24 (1.05, 1.46) Fourier 1.31 (1.08, 1.60) Power law 1.32 (1.09, 1.60) Manduca et al. Cancer Epidemiol Biomarkers Prev. 2009 Häberle et al. Breast Cancer Res. 2012 Gastounioti et al. Breast Cancer Research 2016 (Review)

  13. Associations of parenchymal texture features for specific cancer subtypes Malkov et al. Breast Cancer Res. 2016 Gastounioti et al. Breast Cancer Research 2016 (Review)

  14. Lattice-based Parenchymal Texture Analysis Histogram 95 th mean Entropy Run-length Emphasis Fractal dimension Spatial Lattice Windows Zheng et al. Med Phys. 2015, Keller et al. J Med Imag. 2015

  15. Limitations Non standardized way for feature extraction: • breast sampling Film versus digital mammography • feature parameterization Effects of image acquisition settings • vendor • image format • kVp, mAs, etc. Lack of anatomical correspondences Gastounioti et al. Breast Cancer Research 2016 (Review)

  16. For presentation For processing For presentation For processing (Processed) (Raw) (Processed) (Raw) Gastounioti et al. Medical Physics 2016

  17. Are there differences between image-derived measures from raw and processed digital mammograms? Gastounioti et al. Medical Physics 2016

  18. Automated Quantitative Measurements DA = 𝐵 𝑒𝑓𝑜𝑡𝑓 𝑢𝑗𝑡𝑡𝑣𝑓 𝑄𝐸 = 𝐵 𝑒𝑓𝑜𝑡𝑓 𝑢𝑗𝑡𝑡𝑣𝑓 𝐵 𝑐𝑠𝑓𝑏𝑡𝑢 … 2 Density Measures 29 Texture Features (LIBRA) (histogram, co-occurrence, run-length, structural) Gastounioti et al. Medical Physics 2016

  19. Study Population 8,458 Pairs of MLO-view Raw and Processed Digital Mammograms GE Senographe Essential/Hologic Selenia Dimensions Entire 1 Yr screening cohort 10,739 women MLO images available (Sept. 2010 - Aug. 2011) in both formats No history of breast cancer 4,389 women 4,278 women Unilateral or Bilateral Exclude image breast images available artifacts MLO: medio-lateral oblique Gastounioti et al. Medical Physics 2016

  20. Feature measurements are significantly different, yet strongly or moderately correlated, between raw and processed images. … Gastounioti et al. Medical Physics 2016

  21. Differences depend on the feature, the vendor, and image acquisition settings. … Gastounioti et al. Medical Physics 2016

  22. Differences depend on the feature, the vendor, and image acquisition settings. T1-T28 T1-T28 Modification of the linear model slope by woman- and system-specific factors Gastounioti et al. Medical Physics 2016

  23. Differences depend on the feature, the vendor, and image acquisition settings. T1-T28 T1-T28 Modification of the linear model slope by woman- and system-specific factors Gastounioti et al. Medical Physics 2016

  24. Potential Implications of Such Differences Feature correlations for processed images Feature correlations for processed images Feature correlations for raw images Feature correlations for raw images Gastounioti et al. Medical Physics 2016

  25. Potential Implications of Such Differences Bilateral feature symmetry Bilateral feature symmetry for processed images for processed images Bilateral feature symmetry for raw images Bilateral feature symmetry for raw images Gastounioti et al. Medical Physics 2016

  26. Identifying Robust Texture Features Fractal dimension Local binary pattern Histogram skewness ✓ Strongly correlated ✓ Slight modification of the linear model slope by woman- and system-specific factors Gastounioti et al. Medical Physics 2016

  27. Texture Analysis: The value of considering breast anatomy. 8th International Breast Density and Cancer Risk Assessment Workshop June 8th, 2017

  28. Largely Variable Breast Morphology 8th International Breast Density and Cancer Risk Assessment Workshop June 8th, 2017

  29. Is inherent risk uniformly expressed in the breast parenchyma? Interval Screen-detected Cancers Cancers 18% 28% 61% 65% Breast regions that show a significant difference CBA : central breast area in cancer-case-control classification scores UOA : upper-outer area Meeson et al. Br J Radiol. 2003 Karemore et al. Phys Med Biol. 2014 8th International Breast Density and Cancer Risk Assessment Workshop June 8th, 2017

  30. Breast-anatomy-driven texture analysis Dense vs. fatty tissue Anatomical Weight segmentation Anatomically-oriented texture feature extraction Weighted Breast landmarks texture feature and sub-regions summarization Gastounioti et al. SPIE Medical Imaging 2017, RSNA 2016

  31. Anatomically-oriented polar grid Gastounioti et al. SPIE Medical Imaging 2017, RSNA 2016

  32. 1 2 34 Texture Feature Maps … W 1 Each region is assigned a 0.5 different weight 0 1 Weighted 2 : : Texture Signature : 34 mean std Gastounioti et al. SPIE Medical Imaging 2017, RSNA 2016

  33. Preliminary Evaluation in a Cancer-case-Control Dataset Raw (“For Processing”) MLO -view Digital Mammograms of 424 women GE Healthcare Senographe 2000D / Senographe DS 1:3 age & side-matched 106 318 cancer cases controls Unaffected breasts Women with of women diagnosed with negative screening mammograms and unilateral breast cancer confirmed negative 1-year follow-up MLO: medio-lateral oblique Gastounioti et al. SPIE Medical Imaging 2017, RSNA 2016

  34. Comparisons against simpler texture analysis which does not incorporate the notion of breast anatomy* 1 2 3 : : : 34 mean std Equal weights in Regular grid to texture feature summarization sample the breast * Zheng et al. Med Phys. 2015 Gastounioti et al. SPIE Medical Imaging 2017, RSNA 2016

  35. Incorporating breast anatomy enhances texture associations with breast cancer. Breast-anatomy-driven approach Zheng et al. Med Phys. 2015 AUC = 0.87 AUC =0.80 95% CI [0.79 0.94] 95% CI [0.71 0.85] DeLong’s test p = 0.041 17% of cases correctly reclassified upwards 4% of controls correctly reclassified downwards Gastounioti et al. SPIE Medical Imaging 2017, RSNA 2016

  36. Intrinsic radiomic phenotypes of breast parenchymal complexity and their associations to breast density Work in progress (1 R01 CA207084)

Recommend

More recommend