Dataset Information

A Deep Learning Approach to Re-create Raw Full-Field Digital Mammograms for Breast Density and Texture Analysis.

ABSTRACT:

Purpose

To develop a computational approach to re-create rarely stored for-processing (raw) digital mammograms from routinely stored for-presentation (processed) mammograms.

Materials and methods

In this retrospective study, pairs of raw and processed mammograms collected in 884 women (mean age, 57 years ± 10 [standard deviation]; 3713 mammograms) from October 5, 2017, to August 1, 2018, were examined. Mammograms were split 3088 for training and 625 for testing. A deep learning approach based on a U-Net convolutional network and kernel regression was developed to estimate the raw images. The estimated raw images were compared with the originals by four image error and similarity metrics, breast density calculations, and 29 widely used texture features.

Results

In the testing dataset, the estimated raw images had small normalized mean absolute error (0.022 ± 0.015), scaled mean absolute error (0.134 ± 0.078) and mean absolute percentage error (0.115 ± 0.059), and a high structural similarity index (0.986 ± 0.007) for the breast portion compared with the original raw images. The estimated and original raw images had a strong correlation in breast density percentage (Pearson r = 0.946) and a strong agreement in breast density grade (Cohen κ = 0.875). The estimated images had satisfactory correlations with the originals in 23 texture features (Pearson r ≥ 0.503 or Spearman ρ ≥ 0.705) and were well complemented by processed images for the other six features.

Conclusion

This deep learning approach performed well in re-creating raw mammograms with strong agreement in four image evaluation metrics, breast density, and the majority of 29 widely used texture features.Keywords: Mammography, Breast, Supervised Learning, Convolutional Neural Network (CNN), Deep learning algorithms, Machine Learning AlgorithmsSee also the commentary by Chan in this issue.Supplemental material is available for this article.©RSNA, 2021.

SUBMITTER: Shu H

PROVIDER: S-EPMC8328112 | biostudies-literature |

REPOSITORIES: biostudies-literature

ACCESS DATA

Similar Datasets

Project description:IntroductionMammographic density has been established as a strong risk factor for breast cancer, primarily using digitized film mammograms. Full-field digital mammography (FFDM) is replacing film mammography, has different properties than film, and provides both raw and processed clinical display representation images. We evaluated and compared FFDM raw and processed breast density measures and their associations with breast cancer.MethodsA case-control study of 180 cases and 180 controls matched by age, postmenopausal hormone use, and screening history was conducted. Mammograms were acquired from a General Electric Senographe 2000D FFDM unit. Percent density (PD) was assessed for each FFDM representation using the operator-assisted Cumulus method. Reproducibility within image type (n = 80) was assessed using Lin's concordance correlation coefficient (rc). Correlation of PD between image representations (n = 360) was evaluated using Pearson's correlation coefficient (r) on the continuous measures and the weighted kappa statistic (?) for quartiles. Conditional logistic regression was used to estimate odds ratios (ORs) for the PD and breast cancer associations for both image representations with 95% confidence intervals. The area under the receiver operating characteristic curve (AUC) was used to assess the discriminatory accuracy.ResultsPercent density from the two representations provided similar intra-reader reproducibility (rc= 0.92 for raw and rc= 0.87 for processed images) and was correlated (r = 0.82 and ? = 0.64). When controlling for body mass index, the associations of quartiles of PD with breast cancer and discriminatory accuracy were similar for the raw (OR: 1.0 (ref.), 2.6 (1.2 to 5.4), 3.1 (1.4 to 6.8), 4.7 (2.1 to 10.6); AUC = 0.63) and processed representations (OR: 1.0 (ref.), 2.2 (1.1 to 4.1), 2.2 (1.1 to 4.4), 3.1 (1.5 to 6.6); AUC = 0.64).ConclusionsPercent density measured with an operator-assisted method from raw and processed FFDM images is reproducible and correlated. Both percent density measures provide similar associations with breast cancer.

Project description:We developed a semi-automated tool to assess mammographic density (MD), a phenotype risk marker for breast cancer (BC), in full-field digital images and evaluated its performance testing its reproducibility, comparing our MD estimates with those obtained by visual inspection and using Cumulus, verifying their association with factors that influence MD, and studying the association between MD measures and subsequent BC risk. Three radiologists assessed MD using DM-Scan, the new tool, on 655 processed images (craniocaudal view) obtained in two screening centers. Reproducibility was explored computing pair-wise concordance correlation coefficients (CCC). The agreement between DM-Scan estimates and visual assessment (semi-quantitative scale, 6 categories) was quantified computing weighted kappa statistics (quadratic weights). DM-Scan and Cumulus readings were compared using CCC. Variation of DM-Scan measures by age, body mass index (BMI) and other MD modifiers was tested in regression mixed models with mammographic device as a random-effect term. The association between DM-Scan measures and subsequent BC was estimated in a case-control study. All BC cases in screening attendants (2007-2010) at a center with full-field digital mammography were matched by age and screening year with healthy controls (127 pairs). DM-Scan was used to blindly assess MD in available mammograms (112 cases/119 controls). Unconditional logistic models were fitted, including age, menopausal status and BMI as confounders. DM-Scan estimates were very reliable (pairwise CCC: 0.921, 0.928 and 0.916). They showed a reasonable agreement with visual MD assessment (weighted kappa ranging 0.79-0.81). DM-Scan and Cumulus measures were highly concordant (CCC ranging 0.80-0.84), but ours tended to be higher (4%-5% on average). As expected, DM-Scan estimates varied with age, BMI, parity and family history of BC. Finally, DM-Scan measures were significantly associated with BC (p-trend=0.005). Taking MD<7% as reference, OR per categories of MD were: OR7%-17%=1.32 (95% CI=0.59-2.99), OR17%-28%=2.28 (95% CI=1.03-5.04) and OR>=29%=3.10 (95% CI=1.35-7.14). Our results confirm that DM-Scan is a reliable tool to assess MD in full-field digital mammograms.

Project description:IntroductionMammographic density, the white radiolucent part of a mammogram, is a marker of breast cancer risk and mammographic sensitivity. There are several means of measuring mammographic density, among which are area-based and volumetric-based approaches. Current volumetric methods use only unprocessed, raw mammograms, which is a problematic restriction since such raw mammograms are normally not stored. We describe fully automated methods for measuring both area and volumetric mammographic density from processed images.MethodsThe data set used in this study comprises raw and processed images of the same view from 1462 women. We developed two algorithms for processed images, an automated area-based approach (CASAM-Area) and a volumetric-based approach (CASAM-Vol). The latter method was based on training a random forest prediction model with image statistical features as predictors, against a volumetric measure, Volpara, for corresponding raw images. We contrast the three methods, CASAM-Area, CASAM-Vol and Volpara directly and in terms of association with breast cancer risk and a known genetic variant for mammographic density and breast cancer, rs10995190 in the gene ZNF365. Associations with breast cancer risk were evaluated using images from 47 breast cancer cases and 1011 control subjects. The genetic association analysis was based on 1011 control subjects.ResultsAll three measures of mammographic density were associated with breast cancer risk and rs10995190 (p<0.025 for breast cancer risk and p<1 × 10(-6) for rs10995190). After adjusting for one of the measures there remained little or no evidence of residual association with the remaining density measures (p>0.10 for risk, p>0.03 for rs10995190).ConclusionsOur results show that it is possible to obtain reliable automated measures of volumetric and area mammographic density from processed digital images. Area and volumetric measures of density on processed digital images performed similar in terms of risk and genetic association.

Project description:The amount of fibroglandular tissue content in the breast as estimated mammographically, commonly referred to as breast percent density (PD%), is one of the most significant risk factors for developing breast cancer. Approaches to quantify breast density commonly focus on either semiautomated methods or visual assessment, both of which are highly subjective. Furthermore, most studies published to date investigating computer-aided assessment of breast PD% have been performed using digitized screen-film mammograms, while digital mammography is increasingly replacing screen-film mammography in breast cancer screening protocols. Digital mammography imaging generates two types of images for analysis, raw (i.e., "FOR PROCESSING") and vendor postprocessed (i.e., "FOR PRESENTATION"), of which postprocessed images are commonly used in clinical practice. Development of an algorithm which effectively estimates breast PD% in both raw and postprocessed digital mammography images would be beneficial in terms of direct clinical application and retrospective analysis.This work proposes a new algorithm for fully automated quantification of breast PD% based on adaptive multiclass fuzzy c-means (FCM) clustering and support vector machine (SVM) classification, optimized for the imaging characteristics of both raw and processed digital mammography images as well as for individual patient and image characteristics. Our algorithm first delineates the breast region within the mammogram via an automated thresholding scheme to identify background air followed by a straight line Hough transform to extract the pectoral muscle region. The algorithm then applies adaptive FCM clustering based on an optimal number of clusters derived from image properties of the specific mammogram to subdivide the breast into regions of similar gray-level intensity. Finally, a SVM classifier is trained to identify which clusters within the breast tissue are likely fibroglandular, which are then aggregated into a final dense tissue segmentation that is used to compute breast PD%. Our method is validated on a group of 81 women for whom bilateral, mediolateral oblique, raw and processed screening digital mammograms were available, and agreement is assessed with both continuous and categorical density estimates made by a trained breast-imaging radiologist.Strong association between algorithm-estimated and radiologist-provided breast PD% was detected for both raw (r = 0.82, p < 0.001) and processed (r = 0.85, p < 0.001) digital mammograms on a per-breast basis. Stronger agreement was found when overall breast density was assessed on a per-woman basis for both raw (r = 0.85, p < 0.001) and processed (0.89, p < 0.001) mammograms. Strong agreement between categorical density estimates was also seen (weighted Cohen's κ ≥ 0.79). Repeated measures analysis of variance demonstrated no statistically significant differences between the PD% estimates (p > 0.1) due to either presentation of the image (raw vs processed) or method of PD% assessment (radiologist vs algorithm).The proposed fully automated algorithm was successful in estimating breast percent density from both raw and processed digital mammographic images. Accurate assessment of a woman's breast density is critical in order for the estimate to be incorporated into risk assessment models. These results show promise for the clinical application of the algorithm in quantifying breast density in a repeatable manner, both at time of imaging as well as in retrospective studies.

Project description:Optimization of exposure parameters (target, filter, and kVp) in digital mammography necessitates maximization of the image signal-to-noise ratio (SNR), while simultaneously minimizing patient dose. The goal of this study is to compare, for each of the major commercially available full field digital mammography (FFDM) systems, the impact of the selection of technique factors on image SNR and radiation dose for a range of breast thickness and tissue types. This phantom study is an update of a previous investigation and includes measurements on recent versions of two of the FFDM systems discussed in that article, as well as on three FFDM systems not available at that time. The five commercial FFDM systems tested, the Senographe 2000D from GE Healthcare, the Mammomat Novation DR from Siemens, the Selenia from Hologic, the Fischer Senoscan, and Fuji's 5000MA used with a Lorad M-IV mammography unit, are located at five different university test sites. Performance was assessed using all available x-ray target and filter combinations and nine different phantom types (three compressed thicknesses and three tissue composition types). Each phantom type was also imaged using the automatic exposure control (AEC) of each system to identify the exposure parameters used under automated image acquisition. The figure of merit (FOM) used to compare technique factors is the ratio of the square of the image SNR to the mean glandular dose. The results show that, for a given target/filter combination, in general FOM is a slowly changing function of kVp, with stronger dependence on the choice of target/filter combination. In all cases the FOM was a decreasing function of kVp at the top of the available range of kVp settings, indicating that higher tube voltages would produce no further performance improvement. For a given phantom type, the exposure parameter set resulting in the highest FOM value was system specific, depending on both the set of available target/filter combinations, and on the receptor type. In most cases, the AECs of the FFDM systems successfully identified exposure parameters resulting in FOM values near the maximum ones, however, there were several examples where AEC performance could be improved.