Capability and reliability of deep learning models to make density predictions on low-dose mammograms.

Purpose: Breast density is associated with the risk of developing cancer and can be automatically estimated using deep learning models from digital mammograms. Our aim is to evaluate the capacity and reliability of such models to predict density from low-dose mammograms taken to enable risk estimates for younger women. Approach: We trained deep learning models on standard-dose and simulated low-dose mammograms. The models were then tested on a mammography dataset with paired standard- and low-dose images. The effect of different factors (including age, density, and dose ratio) on the differences between predictions on standard and low doses is analyzed. Methods to improve performance are assessed, and factors that reduce the model quality are demonstrated. Results: We showed that, although many factors have no significant effect on the quality of low-dose density prediction, both density and breast area have an impact. The correlation between density predictions on low- and standard-dose images of breasts with the largest breast area is 0.985 (0.949 to 0.995), whereas that with the smallest is 0.882 (0.697 to 0.961). We also demonstrated that averaging across craniocaudal-mediolateral oblique (CC-MLO) images and across repeatedly trained models can improve predictive performance. Conclusions: Low-dose mammography can be used to produce density and risk estimates that are comparable to standard-dose images. Averaging across CC-MLO and model predictions should improve this performance. The model quality is reduced when making predictions on denser and smaller breasts.

Ethnic differences in the manifestation of early-onset type 2 diabetes.

INTRODUCTION: We undertook phenotypic characterization of early-onset and late-onset type 2 diabetes (T2D) in adult black African and white European populations with recently diagnosed T2D to explore ethnic differences in the manifestation of early-onset T2D. RESEARCH DESIGN AND METHODS: Using the Uganda Diabetes Phenotype study cohort of 500 adult Ugandans and the UK StartRight study cohort of 714 white Europeans with recently diagnosed islet autoantibody-negative T2D, we compared the phenotypic characteristics of participants with early-onset T2D (diagnosed at <40 years) and late-onset T2D (diagnosed at ≥40 years). RESULTS: One hundred and thirty-four adult Ugandans and 113 white Europeans had early-onset T2D. Compared with late-onset T2D, early-onset T2D in white Europeans was significantly associated with a female predominance (52.2% vs 39.1%, p=0.01), increased body mass index (mean (95% CI) 36.7 (35.2-38.1) kg/m2 vs 33.0 (32.4-33.6) kg/m2, p<0.001), waist circumference (112.4 (109.1-115.6) cm vs 108.8 (107.6-110.1) cm, p=0.06), and a higher frequency of obesity (82.3% vs 63.4%, p<0.001). No difference was seen with the post-meal C-peptide levels as a marker of beta-cell function (mean (95% CI) 2130.94 (1905.12-2356.76) pmol/L vs 2039.72 (1956.52-2122.92), p=0.62).In contrast, early-onset T2D in Ugandans was associated with less adiposity (mean (95% CI) waist circumference 93.1 (89.9-96.3) cm vs 97.4 (95.9-98.8) cm, p=0.006) and a greater degree of beta-cell dysfunction (120 min post-glucose load C-peptide mean (95% CI) level 896.08 (780.91-1011.24) pmol/L vs 1310.10 (1179.24-1440.95) pmol/L, p<0.001), without female predominance (53.0% vs 57.9%, p=0.32) and differences in the body mass index (mean (95% CI) 27.3 (26.2-28.4) kg/m2 vs 27.9 (27.3-28.5) kg/m2, p=0.29). CONCLUSIONS: These differences in the manifestation of early-onset T2D underscore the need for ethnic-specific and population-specific therapeutic and preventive approaches for the condition.

Breast density prediction from low and standard dose mammograms using deep learning: effect of image resolution and model training approach on prediction quality.

Purpose. To improve breast cancer risk prediction for young women, we have developed deep learning methods to estimate mammographic density from low dose mammograms taken at approximately 1/10th of the usual dose. We investigate the quality and reliability of the density scores produced on low dose mammograms focussing on how image resolution and levels of training affect the low dose predictions.Methods. Deep learning models are developed and tested, with two feature extraction methods and an end-to-end trained method, on five different resolutions of 15,290 standard dose and simulated low dose mammograms with known labels. The models are further tested on a dataset with 296 matching standard and real low dose images allowing performance on the low dose images to be ascertained.Results. Prediction quality on standard and simulated low dose images compared to labels is similar for all equivalent model training and image resolution versions. Increasing resolution results in improved performance of both feature extraction methods for standard and simulated low dose images, while the trained models show high performance across the resolutions. For the trained models the Spearman rank correlation coefficient between predictions of standard and low dose images at low resolution is 0.951 (0.937 to 0.960) and at the highest resolution 0.956 (0.942 to 0.965). If pairs of model predictions are averaged, similarity increases.Conclusions. Deep learning mammographic density predictions on low dose mammograms are highly correlated with standard dose equivalents for feature extraction and end-to-end approaches across multiple image resolutions. Deep learning models can reliably make high quality mammographic density predictions on low dose mammograms.

Artificial Intelligence for Image-Based Breast Cancer Risk Prediction Using Attention.

Accurate prediction of individual breast cancer risk paves the way for personalised prevention and early detection. The incorporation of genetic information and breast density has been shown to improve predictions for existing models, but detailed image-based features are yet to be included despite correlating with risk. Complex information can be extracted from mammograms using deep-learning algorithms, however, this is a challenging area of research, partly due to the lack of data within the field, and partly due to the computational burden. We propose an attention-based Multiple Instance Learning (MIL) model that can make accurate, short-term risk predictions from mammograms taken prior to the detection of cancer at full resolution. Current screen-detected cancers are mixed in with priors during model development to promote the detection of features associated with risk specifically and features associated with cancer formation, in addition to alleviating data scarcity issues. MAI-risk achieves an AUC of 0.747 [0.711, 0.783] in cancer-free screening mammograms of women who went on to develop a screen-detected or interval cancer between 5 and 55 months, outperforming both IBIS (AUC 0.594 [0.557, 0.633]) and VAS (AUC 0.649 [0.614, 0.683]) alone when accounting for established clinical risk factors.

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.

Automatic assessment of mammographic density using a deep transfer learning method.

Purpose: Mammographic breast density is one of the strongest risk factors for cancer. Density assessed by radiologists using visual analogue scales has been shown to provide better risk predictions than other methods. Our purpose is to build automated models using deep learning and train on radiologist scores to make accurate and consistent predictions. Approach: We used a dataset of almost 160,000 mammograms, each with two independent density scores made by expert medical practitioners. We used two pretrained deep networks and adapted them to produce feature vectors, which were then used for both linear and nonlinear regression to make density predictions. We also simulated an "optimal method," which allowed us to compare the quality of our results with a simulated upper bound on performance. Results: Our deep learning method produced estimates with a root mean squared error (RMSE) of 8.79 ± 0.21 . The model estimates of cancer risk perform at a similar level to human experts, within uncertainty bounds. We made comparisons between different model variants and demonstrated the high level of consistency of the model predictions. Our modeled "optimal method" produced image predictions with a RMSE of between 7.98 and 8.90 for cranial caudal images. Conclusion: We demonstrated a deep learning framework based upon a transfer learning approach to make density estimates based on radiologists' visual scores. Our approach requires modest computational resources and has the potential to be trained with limited quantities of data.

Exploring the application of deep learning methods for polygenic risk score estimation.

Background: Polygenic risk scores (PRS) summarise genetic information into a single number with multiple clinical and research uses. Machine learning (ML) has revolutionised a diverse set of fields, however, the impact of ML on genomics in general, and PRSs in particular, has been less significant. We explore how ML can improve the generation of PRSs. Methods: We train ML models on known PRSs using UK Biobank data. We explore whether the models can recreate human programmed PRSs, including using a single model to generate multiple PRSs, and the difficulty in using ML for PRS generation. We also investigate how ML can compensate for missing data and the constraints on performance. Results: We demonstrate almost perfect generation of PRSs, including when using one model to predict multiple scores, and with little loss of performance with reduced quantity of training data. For an example set of missing SNPs the MLP produces predictions that enable separation of cases from population samples with an area under the receiver operating characteristic curve of 0.847 (95% CI: 0.828-0.864) compared to 0.798 (95% CI: 0.779-0.818) for the PRS. We provide evidence that input information is the limiting factor of further improvement. Conclusions: ML can accurately generate PRSs, including with one model for multiple PRSs. The models are transferable and have high longevity. With certain missing SNPs the ML models can statistically significantly improve on normal PRS generation. Models trained are probably at the edge of performance and further improvements likely require use of additional input data.

Mammographic breast density and breast cancer risk in the Saudi population: a case-control study using visual and automated methods.

OBJECTIVE: This study aims to establish risk of breast cancer based on breast density among Saudi women and to compare cancer prediction using different breast density methods. METHODS: 1140 pseudonymised screening mammograms from Saudi females were retrospectively collected. Breast density was assessed using Breast Imaging Reporting and Data System (BI-RADS) density categories and visual analogue scale (VAS) of 285 cases and 855 controls matched on age and body mass index. In a subset of 160 cases and 480 controls density was estimated by two automated methods, Volpara Density™ and predicted VAS (pVAS). Odds ratios (ORs) between the highest and second categories in BI-RADS and Volpara density grades, and highest vs lowest quartiles in VAS, pVAS and Volpara Density™, were estimated using conditional logistic regression. RESULTS: For BI-RADS, the OR was 6.69 (95% CI 2.79-16.06) in the highest vs second category and OR = 4.78 (95% CI 3.01-7.58) in the highest vs lowest quartile for VAS. In the subset, VAS was the strongest predictor OR = 7.54 (95% CI 3.86-14.74), followed by pVAS using raw images OR = 5.38 (95% CI 2.68-10.77) and Volpara Density ™ OR = 3.55, (95% CI 1.86-6.75) for highest vs lowest quartiles. The matched concordance index for VAS was 0.70 (95% CI 0.65-0.75) demonstrating better discrimination between cases and controls than all other methods. CONCLUSION: Increased mammographic density was strongly associated with risk of breast cancer among Saudi women. Radiologists' visual assessment of breast density is superior to automated methods. However, pVAS and Volpara Density ™ also significantly predicted breast cancer risk based on breast density. ADVANCES IN KNOWLEDGE: Our study established an association between breast density and breast cancer in a Saudi population and compared the performance of automated methods. This provides a stepping-stone towards personalised screening using automated breast density methods.

Rank Selection in Nonnegative Matrix Factorization using Minimum Description Length.

Nonnegative matrix factorization (NMF) is primarily a linear dimensionality reduction technique that factorizes a nonnegative data matrix into two smaller nonnegative matrices: one that represents the basis of the new subspace and the second that holds the coefficients of all the data points in that new space. In principle, the nonnegativity constraint forces the representation to be sparse and parts based. Instead of extracting holistic features from the data, real parts are extracted that should be significantly easier to interpret and analyze. The size of the new subspace selects how many features will be extracted from the data. An effective choice should minimize the noise while extracting the key features. We propose a mechanism for selecting the subspace size by using a minimum description length technique. We demonstrate that our technique provides plausible estimates for real data as well as accurately predicting the known size of synthetic data. We provide an implementation of our code in a Matlab format.