The effect of variable labels on deep learning models trained to predict breast density

Purpose. High breast density is associated with reduced efficacy of mammographic screening and increased risk of developing breast cancer. Accurate and reliable automated density estimates can be used for direct risk prediction and passing density related information to further predictive models. Expert reader assessments of density show a strong relationship to cancer risk but also inter-reader variation. The effect of label variability on model performance is important when considering how to utilise automated methods for both research and clinical purposes.Methods. We utilise subsets of images with density labels from the same 13 readers and 12 reader pairs, and train a deep transfer learning model which is used to assess how label variability affects the mapping from representation to prediction. We then create two end-to-end models: one that is trained on averaged labels across the reader pairs and the second that is trained using individual reader scores, with a novel alteration to the objective function. The combination of these two end-to-end models allows us to investigate the effect of label variability on the model representation formed.Results. We show that the trained mappings from representations to labels are altered considerably by the variability of reader scores. Training on labels with distribution variation removed causes the Spearman rank correlation coefficients to rise from 0.751 ± 0.002 to either 0.815 ± 0.026 when averaging across readers or 0.844 ± 0.002 when averaging across images. However, when we train different models to investigate the representation effect we see little difference, with Spearman rank correlation coefficients of 0.846 ± 0.006 and 0.850 ± 0.006 showing no statistically significant difference in the quality of the model representation with regard to density prediction.Conclusions. We show that the mapping between representation and mammographic density prediction is significantly affected by label variability. However, the effect of the label variability on the model representation is limited.

Application of Deep Learning in Breast Cancer Imaging

This review gives an overview of the current state of deep learning research in breast cancer imaging. Breast imaging plays a major role in detecting breast cancer at an earlier stage, as well as monitoring and evaluating breast cancer during treatment. The most commonly used modalities for breast imaging are digital mammography, digital breast tomosynthesis, ultrasound and magnetic resonance imaging. Nuclear medicine imaging techniques are used for detection and classification of axillary lymph nodes and distant staging in breast cancer imaging. All of these techniques are currently digitized, enabling the possibility to implement deep learning (DL), a subset of Artificial intelligence, in breast imaging. DL is nowadays embedded in a plethora of different tasks, such as lesion classification and segmentation, image reconstruction and generation, cancer risk prediction, and prediction and assessment of therapy response. Studies show similar and even better performances of DL algorithms compared to radiologists, although it is clear that large trials are needed, especially for ultrasound and magnetic resonance imaging, to exactly determine the added value of DL in breast cancer imaging. Studies on DL in nuclear medicine techniques are only sparsely available and further research is mandatory. Legal and ethical issues need to be considered before the role of DL can expand to its full potential in clinical breast care practice.

Identifying Women With Mammographically- Occult Breast Cancer Leveraging GAN-Simulated Mammograms

Our objective is to show the feasibility of using simulated mammograms to detect mammographically-occult (MO) cancer in women with dense breasts and a normal screening mammogram who could be triaged for additional screening with magnetic resonance imaging (MRI) or ultrasound. We developed a Conditional Generative Adversarial Network (CGAN) to simulate a mammogram with normal appearance using the opposite mammogram as the condition. We used a Convolutional Neural Network (CNN) trained on Radon Cumulative Distribution Transform (RCDT) processed mammograms to detect MO cancer. For training CGAN, we used screening mammograms of 1366 women. For MO cancer detection, we used screening mammograms of 333 women (97 MO cancer) with dense breasts. We simulated the right mammogram for normal controls and the cancer side for MO cancer cases. We created two RCDT images, one from a real mammogram pair and another from a real-simulated mammogram pair. We finetuned a VGG16 on resulting RCDT images to classify the women with MO cancer. We compared the classification performance of the CNN trained on fused RCDT images, CNN_Fused to that of trained only on real RCDT images, CNN_Real, and to that of trained only on simulated RCDT images, CNN_Simulated. The test AUC for CNN_Fused was 0.77 with a 95% confidence interval (95CI) of [0.71, 0.83], which was statistically better (p-value < 0.02) than the CNN_Real AUC of 0.70 with a 95CI of [0.64, 0.77] and CNN_Simulated AUC of 0.68 with a 95CI of [0.62, 0.75]. It showed that CGAN simulated mammograms can help MO cancer detection.

RICE: A method for quantitative mammographic image enhancement

We introduce Region of Interest Contrast Enhancement (RICE) to identify focal densities in mammograms. It aims to help radiologists: 1) enhancing the contrast of mammographic images; and 2) detecting regions of interest (such as focal densities) that are candidate masses potentially masked behind dense parenchyma. Cancer masking is an unsolved issue, particularly in breast density categories BI-RADS C and D. RICE suppresses normal breast parenchyma in order to highlight focal densities. Unlike methods that enhance mammograms by modifying the dynamic range of an image; RICE relies on the actual tissue composition of the breast. It segments Volumetric Breast Density (VBD) maps into smaller regions and then applies a recursive mechanism to estimate the 'neighbourhood' for each segment. The method then subtracts and updates the neighbourhood, or the encompassing tissue, from each piecewise constant component of the breast image. This not only enhances the appearance of a candidate mass but also helps in estimating the mass density. In extensive experiments, RICE enhances focal densities in all breast density types including the most challenging category BI-RADS D. Suitably adapted, RICE can be used as a precursor to any computer-aided diagnostics and detection system.

Virtual clinical trials in medical imaging: a review

The accelerating complexity and variety of medical imaging devices and methods have outpaced the ability to evaluate and optimize their design and clinical use. This is a significant and increasing challenge for both scientific investigations and clinical applications. Evaluations would ideally be done using clinical imaging trials. These experiments, however, are often not practical due to ethical limitations, expense, time requirements, or lack of ground truth. Virtual clinical trials (VCTs) (also known as in silico imaging trials or virtual imaging trials) offer an alternative means to efficiently evaluate medical imaging technologies virtually. They do so by simulating the patients, imaging systems, and interpreters. The field of VCTs has been constantly advanced over the past decades in multiple areas. We summarize the major developments and current status of the field of VCTs in medical imaging. We review the core components of a VCT: computational phantoms, simulators of different imaging modalities, and interpretation models. We also highlight some of the applications of VCTs across various imaging modalities.

Detecting mammographically occult cancer in women with dense breasts using deep convolutional neural network and Radon Cumulative Distribution Transform

We have applied the Radon Cumulative Distribution Transform (RCDT) as an image transformation to highlight the subtle difference between left and right mammograms to detect mammographically occult (MO) cancer in women with dense breasts and negative screening mammograms. We developed deep convolutional neural networks (CNNs) as classifiers for estimating the probability of having MO cancer. We acquired screening mammograms of 333 women (97 unilateral MO cancer) with dense breasts and at least two consecutive mammograms and used the immediate prior mammograms, which radiologists interpreted as negative. We used fivefold cross validation to divide our dataset into a training and independent test sets with ratios of 0.8:0.2. We set aside 10% of the training set as a validation set. We applied RCDT on the left and right mammograms of each view. We applied inverse Radon transform to represent the resulting RCDT images in the image domain. We then fine-tuned a VGG16 network pretrained on ImageNet using the resulting images per each view. The CNNs achieved mean areas under the receiver operating characteristic (AUC) curve of 0.73 (standard error, SE = 0.024) and 0.73 (SE = 0.04) for the craniocaudal and mediolateral oblique views, respectively. We combined the scores from two CNNs by training a logistic regression classifier and it achieved a mean AUC of 0.81 (SE = 0.032). In conclusion, we showed that inverse Radon-transformed RCDT images contain information useful for detecting MO cancers and that deep CNNs could learn such information.

Investigating the feasibility of stratified breast cancer screening using a masking risk predictor

BACKGROUND

Women with dense breasts face a double risk for breast cancer; they are at a higher risk for development of breast cancer than those with less dense breasts, and there is a greater chance that mammography will miss detection of a cancer in dense breasts due to the masking effect of surrounding fibroglandular tissue. These women may be candidates for supplemental screening. In this study, a masking risk model that was previously developed is tested on a cohort of cancer-free women to assess potential efficiency of stratification.

METHODS

Three masking risk models based on (1) BI-RADS density, (2) volumetric breast density (VBD), and (3) a combination of VBD and detectability were applied to stratify the mammograms of 1897 cancer-free women. The fraction of cancer-free women whose mammograms were deemed by the algorithm to be masked and who would be considered for supplemental imaging was computed as was the corresponding fraction in a screened population of interval (masked) cancers that would be potentially detected by supplemental imaging.

RESULTS

Of the models tested, the combined VBD/detectability model offered the highest efficiency for stratification to supplemental imaging. It predicted that 725 supplemental screens would be performed per interval cancer potentially detected, at an operating point that allowed detection of 64% of the interval cancers. In comparison, stratification based on the upper two BI-RADS density categories required 1117 supplemental screenings per interval cancer detected to capture 64% of interval cancers.

CONCLUSION

The combined VBD/detectability models perform better than BI-RADS and offer a continuum of operating points, suggesting that this model may be effective in guiding a stratified screening environment.

Deep learning networks find unique mammographic differences in previous negative mammograms between interval and screen-detected cancers: a case-case study

Background

To determine if mammographic features from deep learning networks can be applied in breast cancer to identify groups at interval invasive cancer risk due to masking beyond using traditional breast density measures.

Methods

Full-field digital screening mammograms acquired in our clinics between 2006 and 2015 were reviewed. Transfer learning of a deep learning network with weights initialized from ImageNet was performed to classify mammograms that were followed by an invasive interval or screen-detected cancer within 12 months of the mammogram. Hyperparameter optimization was performed and the network was visualized through saliency maps. Prediction loss and accuracy were calculated using this deep learning network. Receiver operating characteristic (ROC) curves and area under the curve (AUC) values were generated with the outcome of interval cancer using the deep learning network and compared to predictions from conditional logistic regression with errors quantified through contingency tables.

Results

Pre-cancer mammograms of 182 interval and 173 screen-detected cancers were split into training/test cases at an 80/20 ratio. Using Breast Imaging-Reporting and Data System (BI-RADS) density alone, the ability to correctly classify interval cancers was moderate (AUC = 0.65). The optimized deep learning model achieved an AUC of 0.82. Contingency table analysis showed the network was correctly classifying 75.2% of the mammograms and that incorrect classifications were slightly more common for the interval cancer mammograms. Saliency maps of each cancer case found that local information could highly drive classification of cases more than global image information.

Conclusions

Pre-cancerous mammograms contain imaging information beyond breast density that can be identified with deep learning networks to predict the probability of breast cancer detection.

Prediction of Cancer Masking in Screening Mammography Using Density and Textural Features

RATIONALE AND OBJECTIVES

High mammographic density reduces the diagnostic accuracy of screening mammography due to masking of tumors, resulting in possible delayed diagnosis and missed cancers. Women with high masking risk could be preselected for alternative screening regimens less susceptible to masking. In this study, various models to predict masking status are presented based on biometric and image-based parameters.

MATERIALS AND METHODS

For a cohort of 67 nonscreen-detected (cancers detected via other means after a negative mammogram) and 147 screen-detected invasive cancers, quantitative volumetric breast density, BI-RADS density, and the distribution and appearance of dense tissue through statistical and texture metrics were measured. Age and Body Mass Index were recorded. Stepwise multivariate logistic regressions were computed to select those parameters that predicted nonscreen-detected cancers. Accuracy of the models was evaluated using the area under receiver operator characteristic curve (AUC).

RESULTS

Using BI-RADS density alone to predict masking risk yielded an AUC of 0.64 (95% confidence interval [0.57-0.70]). Age-adjusted BI-RADS density or volumetric breast density had AUCs of 0.72 [0.64-0.79] and 0.71 [0.62-0.78], respectively. A model extracted from the full pool of variables had an AUC of 0.75 [0.67-0.82].

CONCLUSION

The optimal model predicts masking more accurately than density alone, suggesting that texture metrics may be useful in models to guide a stratified screening strategy.

Derived mammographic masking measures based on simulated lesions predict the risk of interval cancer after controlling for known risk factors: a case-case analysis

PURPOSE

Women with radiographically dense or texturally complex breasts are at increased risk for interval cancer, defined as cancers diagnosed after a normal screening examination. The purpose of this study was to create masking measures and apply them to identify interval risk in a population of women who experienced either screen-detected or interval cancers after controlling for breast density.

METHODS

We examined full-field digital screening mammograms acquired from 2006 to 2015. Examinations associated with 182 interval cancers were matched to 173 screen-detected cancers on age, race, exam date and time since last imaging examination. Local Image Quality Factor (IQF) values were calculated and used to create IQF maps that represented mammographic masking. We used various statistics to define global masking measures of these maps. Association of these masking measures with interval cancer vs screen-detected cancer was estimated using conditional logistic regression in a univariate and adjusted model for Breast Imaging-Reporting and Data System (BI-RADS) density. Receiver operator curves were calculated in each case to compare specificity vs sensitivity, and area under those curves were generated. Proportion of screen-detected cancer was estimated for stratifications of IQF features.

RESULTS

Several masking features showed significant association with interval compared to screen-detected cancers after adjusting for BI-RADS density (up to P = 2.52E-6), and the 10th percentile of the IQF value (P = 1.72E-3) showed the strongest improvement in the area under the receiver operator curve, increasing from 0.65 using only BI-RADS density to 0.69. The highest masking group had a 32% proportion of screen-detected cancers while the low masking group had a 69% proportion.

CONCLUSIONS

We conclude that computer vision methods using model observers may improve quantifying the probability of breast cancer detection beyond using breast density alone.

Localized mammographic density is associated with interval cancer and large breast cancer: a nested case-control study

Background

High mammographic density is associated with breast cancer and with delayed detection. We have examined whether localized density, at the site of the subsequent cancer, is independently associated with being diagnosed with a large-sized or interval breast cancer.

Methods

Within a prospective cohort of 63,130 women, we examined 891 women who were diagnosed with incident breast cancer. For 386 women, retrospective localized density assessment was possible. The main outcomes were interval cancer vs. screen-detected cancer and large (> 2 cm) vs. small cancer. In negative screening mammograms, overall and localized density were classified reflecting the BI-RADS standard. Density concordance probabilities were estimated through multinomial regression. The associations between localized density and the two outcomes were modeled through logistic regression, adjusted for overall density, age, body mass index, and other characteristics.

Results

The probabilities of concordant localized density were 0.35, 0.60, 0.38, and 0.32 for overall categories “A,” “B,” “C,” and “D.” Overall density was associated with large cancer, comparing density category D to A with OR 4.6 (95%CI 1.8–11.6) and with interval cancer OR 31.5 (95%CI 10.9–92) among all women. Localized density was associated with large cancer at diagnosis with OR 11.8 (95%CI 2.7–51.8) among all women and associated with first-year interval cancer with OR 6.4 (0.7 to 58.7) with a significant linear trend p = 0.027.

Conclusions

Overall density often misrepresents localized density at the site where cancer subsequently arises. High localized density is associated with interval cancer and with large cancer. Our findings support the continued effort to develop and examine computer-based measures of localized density for use in personalized breast cancer screening.

Electronic supplementary material

The online version of this article (10.1186/s13058-019-1099-y) contains supplementary material, which is available to authorized users.

Feasibility of estimating volumetric breast density from mammographic x-ray spectra using a cadmium telluride photon-counting detector

PURPOSE

Mammographic density of glandular breast tissue has a masking effect that can reduce lesion detection accuracy and is also a strong risk factor for breast cancer. Therefore, accurate quantitative estimation of breast density is clinically important. In this study, we investigate experimentally the feasibility of quantifying volumetric breast density with spectral mammography using a CdTe-based photon-counting detector.

METHODS

To demonstrate proof-of-principle, this study was carried out using the single pixel Amptek XR-100T-CdTe detector. The total number of x rays recorded by the detector from a single pencil-beam projection through 50%/50% of adipose/glandular mass fraction-equivalent phantoms was measured. Material decomposition assuming two, four, and eight energy bins was then applied to characterize the inspected phantom into adipose and glandular using log-likelihood estimation, taking into account the polychromatic source, the detector response function, and the energy-dependent attenuation.

RESULTS

Measurement tests were carried out for different doses, kVp settings, and different breast sizes. For dose of 1 mGy and above, the percent relative root mean square (RMS) errors of the estimated breast density was measured below 7% for all three phantom studies. It was also observed that some decrease in RMS errors was achieved using eight energy bins. For 3 and 4 cm thick phantoms, performance at 40 and 45 kVp showed similar performance. However, it was observed that 45 kVp showed better performance for a phantom thickness of 6 cm at low dose levels due to increased statistical variation at lower photon count levels with 40 kVp.

CONCLUSION

The results of the current study suggest that photon-counting spectral mammography systems using CdTe detectors have the potential to be used for accurate quantification of volumetric breast density on a pixel-to-pixel basis, with an RMS error of less than 7%.

Qualitative Versus Quantitative Mammographic Breast Density Assessment: Applications for the US and Abroad

Mammographic breast density (MBD) has been proven to be an important risk factor for breast cancer and an important determinant of mammographic screening performance. The measurement of density has changed dramatically since its inception. Initial qualitative measurement methods have been found to have limited consistency between readers, and in regards to breast cancer risk. Following the introduction of full-field digital mammography, more sophisticated measurement methodology is now possible. Automated computer-based density measurements can provide consistent, reproducible, and objective results. In this review paper, we describe various methods currently available to assess MBD, and provide a discussion on the clinical utility of such methods for breast cancer screening.

Quantification of masking risk in screening mammography with volumetric breast density maps

Purpose

Fibroglandular tissue may mask breast cancers, thereby reducing the sensitivity of mammography. Here, we investigate methods for identification of women at high risk of a masked tumor, who could benefit from additional imaging.

Methods

The last negative screening mammograms of 111 women with interval cancer (IC) within 12 months after the examination and 1110 selected normal screening exams from women without cancer were used. From the mammograms, volumetric breast density maps were computed, which provide the dense tissue thickness for each pixel location. With these maps, three measurements were derived: (1) percent dense volume (PDV), (2) percent area where dense tissue thickness exceeds 1 cm (PDA), and (3) dense tissue masking model (DTMM). Breast density was scored by a breast radiologist using BI-RADS. Women with heterogeneously and extremely dense breasts were considered at high masking risk. For each masking measure, mammograms were divided into a high- and low-risk category such that the same proportion of the controls is at high masking risk as with BI-RADS.

Results

Of the women with IC, 66.1, 71.9, 69.2, and 63.0% were categorized to be at high masking risk with PDV, PDA, DTMM, and BI-RADS, respectively, against 38.5% of the controls. The proportion of IC at high masking risk is statistically significantly different between BI-RADS and PDA (p-value 0.022). Differences between BI-RADS and PDV, or BI-RADS and DTMM, are not statistically significant.

Conclusion

Measures based on density maps, and in particular PDA, are promising tools to identify women at high risk for a masked cancer.

Location- and lesion-dependent estimation of mammographic background tissue complexity

We specify a notion of perceived background tissue complexity (BTC) that varies with lesion shape, lesion size, and lesion location in the image. We propose four unsupervised BTC estimators based on: perceived pre and postlesion similarity of images, lesion border analysis (LBA; conspicuous lesion should be brighter than its surround), tissue anomaly detection, and local energy. The latter two are existing methods adapted for location- and lesion-dependent BTC estimation. For evaluation, we ask human observers to measure BTC (threshold visibility amplitude of a given lesion inserted) at specified locations in a mammogram. As expected, both human measured and computationally estimated BTC vary with lesion shape, size, and location. BTCs measured by different human observers are correlated ([Formula: see text]). BTC estimators are correlated to each other ([Formula: see text]) and less so to human observers ([Formula: see text]). With change in lesion shape or size, LBA estimated BTC changes in the same direction as human measured BTC. Proposed estimators can be generalized to other modalities (e.g., breast tomosynthesis) and used as-is or customized to a specific human observer, to construct BTC-aware model observers with applications, such as optimization of contrast-enhanced medical imaging systems and creation of a diversified image dataset with characteristics of a desired population.