Specifically, LIBRA delineates the air-breast boundary using automated gray-level thresholding, and extracts the pectoralis muscle region by applying a straight-line Hough transform To identify dense versus fatty tissue areas, LIBRA uses adaptive fuzzy c-means clustering to partition the breast into density clusters DCs of similar gray-level intensity, which are then aggregated into the final dense tissue segmentation. For landmarks d — e , the breast image was first rotated, and then flipped for right breasts, so that the pectoralis muscle was vertically aligned on the left side of all images.
Then, the nipple was approximated as the rightmost circular region of the breast bounded by a rectangle of maximum size 1. The parameter f , was implemented to range within [0. A polar grid of a radius unit equal to D , centered on the nipple, was then overlaid on the mammographic image Fig. The polar grid was fitted to the shape and size of the individual breast as well as to an approximation of the ductal distribution extending from the nipple posteriorly and perpendicular to the pectoralis muscle in a radial fashion.
Centering on the nipple allowed for denser sampling in the retro-areolar region of the breast, which typically contains more complex parenchymal tissue patterns than the breast areas closer to the pectoralis muscle Anatomical sampling of the breast. Polar grid fitted to the morphology of the particular breast and morphology-aligned orientations for texture feature calculations.
- Insomniac City: New York, Oliver, and Me;
- All Challenges.
- Radiology dataset.
- Bestselling Series.
Briefly, gray-level histogram features are common first-order statistics which describe the distribution of gray-level intensity. The co-occurrence features also consider the spatial relationships of pixel intensities in specified directions and are based on the gray-level co-occurrence matrix GLCM which encodes the relative frequency of neighboring intensity values. Run-length features capture the coarseness of texture in specified directions by measuring strings of consecutive pixels which have the same gray-level intensity along specific linear orientations.
Using the polar grid, the axes for calculating the co-occurrence and run-length features were aligned with the structure of the parenchymal tissue and ductal structures Fig. Finally, structural features capture the architectural composition of the parenchyma by characterizing the tissue complexity, the directionality of flow-like structures in the breast, and intensity variations between central and neighboring pixels.
A total of 34 feature maps were computed i. As the final step, we generated a weight map which assigned a weight to each region; this weight map was, then, region-wise multiplied to the 34 original texture feature maps to generate a set of 34 weighted texture feature maps. Our design of the weight map was motivated by studies associating the CBA and UOA with potential specific roles in breast cancer development 12 , 13 , and by work investigating biologic correlates of tissue composition with breast cancer development 11 , 18 , 19 , The relative importance of weights S and T when merged to the final weight map was defined by parameter c.
See Supplementary Note N2 for detailed definitions. Generating the weight of each region. The optimization of our anatomy-driven approach for parenchymal texture analysis and breast cancer prediction involved two inter-related tasks: automatically configuring the tunable texture analysis parameters D, f, a, b , and c and determining the most discriminatory subset of covariates out of the 68 available texture features. To this end, for different parameter combinations Supplementary Fig. S1 , the following steps were applied.
First, per-woman texture estimates were generated by averaging the corresponding bilateral breast texture signatures for each woman and were, then, z-score normalized. To remove features with little or no variability while also maintaining all different aspects of texture captured by the different features, we, then, identified pairs of features with absolute Pearson correlation greater than 0. Starting from the remaining features, elastic net regression 21 with nested cross-validation 22 was used to build a parsimonious logistic regression model with the most discriminatory subset of covariates, where model performance was evaluated using the area under the curve AUC of the receiver operating characteristic ROC see Supplementary Note N3 for feature optimization details.
The model corresponding to maximum cross-validated AUC i. Baseline and augmented logistic regression models were fitted to obtain estimates of AUC, odds ratios ORs , and statistical significance of predictor variables. For comparison, conventional lattice-based texture measurements 23 were also evaluated in terms of their ability to augment established risk factors in breast cancer risk prediction.
After excluding 17 features with low IQR, elastic net regression selected 30 of the 51 textural features for inclusion in this optimal model, including nine gray-level histogram, nine co-occurrence, eight run-length, and four structural features; among these, 14 represented mean values and 16 the variation i. Texture feature maps for four texture descriptors. Top row: weighted values on polar grid using the proposed breast-anatomy-driven approach with the optimal set of parameters.
Bottom row: non-weighted values on a regular lattice Our anatomy-driven features were also able to significantly augment all four baseline models fitted with breast density, BMI and age. Our findings suggest that incorporating breast anatomy information in mammographic phenotypes of parenchymal pattern could augment imaging markers of breast cancer risk with the potential to improve personalized breast cancer risk assessment.
Interestingly, the configuration of the weight map in our breast-anatomy-driven approach as indicated by the optimization experiments also suggests that the textural properties of different regions in the breast may contribute differently towards breast cancer risk, with dense tissue regions and the central breast quadrant having potentially a more important role.
Further, the heterogeneity in textural measurements within the breast may also be important, as more than half of the features selected as strongest covariates in our model represent the variation i. In addition, our anatomy-driven approach to breast parenchymal texture analysis outperformed parenchymal texture features assessed without the incorporation of factors describing breast anatomy structure and variability We postulate that the improvement observed in this preliminary comparison is due to including factors that capture the wide variety of breast morphology and anatomy, found not only across the screening population but also within a single woman due to differences in positioning for FFDM.
Incorporating such information, therefore, allows for also establishing anatomical correspondences across breasts of the same or different women, which in turn results in standardized imaging features and more comparable texture measurements. The observed improvement in discriminatory capacity may also be due to our polar grid, which allows for more granular texture measurements to be obtained in the retroareolar breast area where some of the most complex parenchymal tissue patterns usually appear 17 , and the ability to consider different contributions of the different sub-regions within the breast in the overall parenchymal texture signature.
With this improvement, our anatomy-driven approach achieved a promising performance in this challenging task of breast cancer risk prediction with prior screening mammograms. Overall, adding the breast-anatomy-driven features to baseline models with established breast cancer risk factors led to a significant increase in discriminatory capacity, suggesting a promising role in augmenting breast cancer risk assessment models. Similar conclusions have been reported in related studies 4 , 26 , where case-control classification models considering parenchymal textural features in addition to established risk factors and breast density achieved AUC values of 0.
Together, these findings consistently support independent associations of parenchymal texture with breast cancer and, therefore, create a strong argument for incorporating quantitative breast textural features in models estimating breast cancer risk. The improvement of breast cancer risk estimation models can have substantial clinical implications, as it would allow for more informed recommendations for supplementary breast cancer screening e.
Hence, the potential of parenchymal texture to leverage breast cancer risk assessment might ultimately affect the chance of early cancer detection or prevention in women categorized as being at low breast cancer risk based on conventional risk factors. Important limitations of our study must also be noted.
In addition, our analysis was confined to a fixed feature set and, although elastic-net regression was used to alleviate model over-fitting, our reported model performance may be over-estimated due to the relatively small sample size as a single-institution evaluation. Considering the reported substantial differences in textural measurements across image acquisition settings 27 , different FFDM representations, and vendors 28 , in our future studies we will plan to more thoroughly test the robustness of our method by incorporating multiple FFDM vendors from larger populations.
In addition, larger studies will allow us to more rigorously evaluate the added discriminatory capacity of such imaging biomarkers when considering additional demographic and clinical risk factors e. Finally, while 2D FFDM images were analyzed as a first step for the purposes of this proof-of-concept study, we ultimately envision extending our algorithm to volumetric texture analysis for digital breast tomosynthesis images also available for the studies in our study population , as tomosynthesis is increasingly being clinically implemented due to its reported improvements in sensitivity and specificity as compared to conventional 2D FFDM This new pseudo-3D imaging technology may, therefore, also result in superior imaging phenotypes of breast cancer risk.
In conclusion, our study provides evidence that incorporating breast anatomy strengthens the associations of mammographic parenchymal phenotypes with breast cancer risk and suggests that anatomy-driven measurements of parenchymal texture could complement current established risk factors and quantitative breast density measures. This additional information reflecting breast anatomy structure has the potential to further refine individualized risk assessment and, therefore, advance tailored screening and prevention strategies for breast cancer. The data generated during the current study are available from the corresponding author on reasonable request.
Hum, S. Current Breast Cancer Reports 8 , — Onega, T. Cancer , — McDonald, E. Clinical Diagnosis and Management of Breast Cancer. Journal of Nuclear Medicine 57 , 9S—16S Manduca, A. Texture features from mammographic images and risk of breast cancer. Cancer Epidemiol Biomarkers Prev 18 , — Li, H. Computerized analysis of mammographic parenchymal patterns on a large clinical dataset of full-field digital mammograms: robustness study with two high-risk datasets. Journal of Digital Imaging 25 , — Jon Sporring. Longin Jan Latecki.
Home Contact us Help Free delivery worldwide. Free delivery worldwide. Bestselling Series. Harry Potter. Popular Features. New Releases. Description Breast cancer is a major health problem in the Western world, where it is the most common cancer among women. Approximately 1 in 12 women will develop breast cancer during the course of their lives.
Over the past twenty years there have been a series of major advances in the manage- ment of women with breast cancer, ranging from novel chemotherapy and radiotherapy treatments to conservative surgery. The next twenty years are likely to see computerized image analysis playing an increasingly important role in patient management. As applications of image analysis go, medical applications are tough in general, and breast cancer image analysis is one of the toughest. There are many reasons for this: highly variable and irregular shapes of the objects of interest, changing imaging conditions, and the densely textured nature of the images.
Add to this the increasing need for quantitative informa- tion, precision, and reliability very few false positives , and the image pro- cessing challenge becomes quite daunting, in fact it pushes image analysis techniques right to their limits. Product details Format Paperback pages Dimensions x x Other books in this series. Add to basket. Image Structure Luc Florack. Medial Representations Kaleem Siddiqi. Robust Computer Vision Nicu Sebe. Scharfenberger, C. Chung, A. Khalvati, M. Shafiee , M. Haider, and A. Ben Daya, I. Chen, J. Yeow, and A. Cameron, A. Khalvati, A.
Wong , and M. Amelard, R.
Kazemzadeh , K. Pfisterer, B. Lui, D. Modhafar, A. Lin, D. Clausi , and A.
CVonline: Vision Related Books including Online Books and Book Support Sites
Kazemzadeh, F. Jin, S. Molladavoodi, M. Emelko, M. Gorbet, and A. Azimifar, and A. Li, F. Boroomand , N. Schwartz, S. Khalvati, F. Cameron , A. Fieguth , and M. R Medeiros, S. Scharcanski, and A. Wang, and M. Kazemzadeh , C. Jin, and Z. Wang, C. Scott, A. Glaister, J. Liu, C. Wong , A. Fieguth , H. Bie, and K. Cho, D. Clausi , J. Callaghan, and J. Scharfenberger , K. Fergani, K. Wong , and J. Wong , and C. Gallagher, K.
- Biomedical Imaging | Vision and Image Processing Lab | University of Waterloo.
- Nav view search.
- Reliving Karbala: Martyrdom in South Asian Memory.
- Learning the bash Shell: Unix Shell Programming (In a Nutshell (OReilly))!
- Tales from the Arabian Nights?
- Computational Imaging and Vision.
Boroomand, A. Wong , E. Cho , B. Ni, and K. Wong , and K. Cameron , and M. Gallagher, and J. Kumar, A. Fieguth , and K. Wang, X. Eichel, J.
Cliffs of Moher Tour
Wong , K. Fieguth , and H. Karimi, A. Hariri, and K. Moayed, A. Wong , V. Choh, and K. Irving, R. Genest, V. Choh, and N. Zhang, W. Wong , and P-H. Frankovich, M. Liu, Y. Mishra, A. Koff, D. Scharcanski, L. Mishra , K. Tung, F. Dunk, and J. Mishra , J. Clausi , and J. Bizheva, K.