Spatio-temporal trend of BIRADS classification suggestive of malignancy: a national analysis of mammograms, 2013-2017.

INTRODUCTION: Promoting breast cancer (BC) detection in women by means of mammography is a viable strategy to reduce the number of diagnoses at clinically advanced stages and mortality. OBJECTIVES: To describe the results reported by mammography studies in women, carried out nationally during 2013-2017, and to analyze the spatiotemporal trend of Breast Imaging Reporting and Data System (BIRADS) categories suggestive of malignancy by State. METHOD: Longitudinal, analytical design that included information on mammography studies of women according to age group (< 40 and ≥ 40), evaluated in units of the Ministry of Health of Mexico during 2013-2017. The frequency of BIRADS categories and a standardized rate suggestive of malignancy (categories 4 and 5) were estimated in women aged ≥ 40 years, and spatial statistics were used to analyze the trend by State. RESULTS: A total of 3,659,151 mammograms were analyzed, 98.5 % in women aged ≥ 40 years. The malignancy-suggestive rate decreased from 38.3 (2013) to 31 (2017) per 100,000 women aged ≥ 40 years; however, the risk of detection increased up to 13 times in ten States. CONCLUSIONS: Although the risk of detection in categories suggestive of malignancy decreased at the national level, some States need to reinforce the application of BC detection programs through mammography and increase the participation of the target population.

INTRODUCCIÓN: Promover la detección de cáncer de mama (CaMa) en mujeres mediante mastografía es una estrategia viable para disminuir los diagnósticos en fases clínicamente avanzadas y la mortalidad. OBJETIVOS: Describir los resultados reportados por estudios de mastografía en mujeres realizados a nivel nacional durante 2013-2017 y analizar la tendencia espaciotemporal de categorías BIRADS (Breast Imaging Reporting and Data System) sugestivas de malignidad por Estado. MÉTODO: Diseño analítico longitudinal que incluyó información sobre estudios de mastografía de mujeres según grupo de edad (< 40 e ≥ 40), valoradas en unidades de la Secretaría de Salud, México, durante 2013-2017. Se estimó la frecuencia de categorías según BIRADS, tasa estandarizada sugestiva de malignidad (categorías 4 y 5) en mujeres ≥ 40 años y se utilizó estadística espacial para analizar la tendencia por Estado. RESULTADOS: Se analizaron 3,659,151 mastografías, el 98.5 % en mujeres ≥ 40 años. La tasa sugestiva de malignidad disminuyó de 38.3 (2013) a 31 (2017) por 100 mil mujeres ≥ 40 años; sin embargo, el riesgo de detección aumentó hasta 13 veces en diez Estados. CONCLUSIONES: Aunque el riesgo de detección en categorías sugestivas de malignidad disminuyó a nivel nacional, algunos Estados requieren reforzar la aplicación de programas de detección del CaMa mediante mastografía e incrementar la participación de la población blanco.

Impact of image compression on deep learning-based mammogram classification.

Image compression is used in several clinical organizations to help address the overhead associated with medical imaging. These methods reduce file size by using a compact representation of the original image. This study aimed to analyze the impact of image compression on the performance of deep learning-based models in classifying mammograms as "malignant"-cases that lead to a cancer diagnosis and treatment-or "normal" and "benign," non-malignant cases that do not require immediate medical intervention. In this retrospective study, 9111 unique mammograms-5672 normal, 1686 benign, and 1754 malignant cases were collected from the National Cancer Center in the Republic of Korea. Image compression was applied to mammograms with compression ratios (CRs) ranging from 15 to 11 K. Convolutional neural networks (CNNs) with three convolutional layers and three fully-connected layers were trained using these images to classify a mammogram as malignant or not malignant across a range of CRs using five-fold cross-validation. Models trained on images with maximum CRs of 5 K had an average area under the receiver operating characteristic curve (AUROC) of 0.87 and area under the precision-recall curve (AUPRC) of 0.75 across the five folds and compression ratios. For images compressed with CRs of 10 K and 11 K, model performance decreased (average 0.79 in AUROC and 0.49 in AUPRC). Upon generating saliency maps that visualize the areas each model views as significant for prediction, models trained on less compressed (CR < = 5 K) images had maps encapsulating a radiologist's label, while models trained on images with higher amounts of compression had maps that missed the ground truth completely. In addition, base ResNet18 models pre-trained on ImageNet and trained using compressed mammograms did not show performance improvements over our CNN model, with AUROC and AUPRC values ranging from 0.77 to 0.87 and 0.52 to 0.71 respectively when trained and tested on images with maximum CRs of 5 K. This paper finds that while training models on images with increased the robustness of the models when tested on compressed data, moderate image compression did not substantially impact the classification performance of DL-based models.

Tendencia espacio-temporal de clasificación BIRADS sugestiva de malignidad: un análisis nacional de mastografías, 2013-2017 / Spatio-temporal trend of BIRADS classification suggestive of malignancy: a national analysis of mammograms, 2013-2017

Resumen Introducción: Promover la detección de cáncer de mama (CaMa) en mujeres mediante mastografía es una estrategia viable para disminuir los diagnósticos en fases clínicamente avanzadas y la mortalidad. Objetivos: Describir los resultados reportados por estudios de mastografía en mujeres realizados a nivel nacional durante 2013-2017 y analizar la tendencia espaciotemporal de categorías BIRADS (Breast Imaging Reporting and Data System) sugestivas de malignidad por Estado. Método: Diseño analítico longitudinal que incluyó información sobre estudios de mastografía de mujeres según grupo de edad (< 40 e ≥ 40), valoradas en unidades de la Secretaría de Salud, México, durante 2013-2017. Se estimó la frecuencia de categorías según BIRADS, tasa estandarizada sugestiva de malignidad (categorías 4 y 5) en mujeres ≥ 40 años y se utilizó estadística espacial para analizar la tendencia por Estado. Resultados: Se analizaron 3,659,151 mastografías, el 98.5 % en mujeres ≥ 40 años. La tasa sugestiva de malignidad disminuyó de 38.3 (2013) a 31 (2017) por 100 mil mujeres ≥ 40 años; sin embargo, el riesgo de detección aumentó hasta 13 veces en diez Estados. Conclusiones: Aunque el riesgo de detección en categorías sugestivas de malignidad disminuyó a nivel nacional, algunos Estados requieren reforzar la aplicación de programas de detección del CaMa mediante mastografía e incrementar la participación de la población blanco.

Abstract Introduction: Promoting breast cancer (BC) detection in women by means of mammography is a viable strategy to reduce the number of diagnoses at clinically advanced stages and mortality. Objectives: To describe the results reported by mammography studies in women, carried out nationally during 2013-2017, and to analyze the spatiotemporal trend of Breast Imaging Reporting and Data System (BIRADS) categories suggestive of malignancy by State. Method: Longitudinal, analytical design that included information on mammography studies of women according to age group (< 40 and ≥ 40), evaluated in units of the Ministry of Health of Mexico during 2013-2017. The frequency of BIRADS categories and a standardized rate suggestive of malignancy (categories 4 and 5) were estimated in women aged ≥ 40 years, and spatial statistics were used to analyze the trend by State. Results: A total of 3,659,151 mammograms were analyzed, 98.5 % in women aged ≥ 40 years. The malignancy-suggestive rate decreased from 38.3 (2013) to 31 (2017) per 100,000 women aged ≥ 40 years; however, the risk of detection increased up to 13 times in ten States. Conclusions: Although the risk of detection in categories suggestive of malignancy decreased at the national level, some States need to reinforce the application of BC detection programs through mammography and increase the participation of the target population.

Deep feature-based automatic classification of mammograms.

Breast cancer has the second highest frequency of death rate among women worldwide. Early-stage prevention becomes complex due to reasons unknown. However, some typical signatures like masses and micro-calcifications upon investigating mammograms can help diagnose women better. Manual diagnosis is a hard task the radiologists carry out frequently. For their assistance, many computer-aided diagnosis (CADx) approaches have been developed. To improve upon the state of the art, we proposed a deep ensemble transfer learning and neural network classifier for automatic feature extraction and classification. In computer-assisted mammography, deep learning-based architectures are generally not trained on mammogram images directly. Instead, the images are pre-processed beforehand, and then they are adopted to be given as input to the ensemble model proposed. The robust features extracted from the ensemble model are optimized into a feature vector which are further classified using the neural network (nntraintool). The network was trained and tested to separate out benign and malignant tumors, thus achieving an accuracy of 0.88 with an area under curve (AUC) of 0.88. The attained results show that the proposed methodology is a promising and robust CADx system for breast cancer classification. Graphical Abstract Flow diagram of the proposed approach. Figure depicts the deep ensemble extracting the robust features with the final classification using neural networks.

Two-view digital breast tomosynthesis versus digital mammography in a population-based breast cancer screening programme (To-Be): a randomised, controlled trial.

BACKGROUND: Digital breast tomosynthesis is an advancement of mammography, and has the potential to overcome limitations of standard digital mammography. This study aimed to compare first-generation digital breast tomo-synthesis including two-dimensional (2D) synthetic mammograms versus digital mammography in a population-based screening programme. METHODS: BreastScreen Norway offers all women aged 50-69 years two-view (craniocaudal and mediolateral oblique) mammographic screening every 2 years and does independent double reading with consensus. We asked all 32 976 women who attended the programme in Bergen in 2016-17, to participate in this randomised, controlled trial with a parallel group design. A study-specific software was developed to allocate women to either digital breast tomosynthesis or digital mammography using a 1:1 simple randomisation method based on participants' unique national identity numbers. The interviewing radiographer did the randomisation by entering the number into the software. Randomisation was done after consent and was therefore concealed from both the women and the radiographer at the time of consent; the algorithm was not disclosed to radiographers during the recruitment period. All data needed for analyses were complete 12 months after the recruitment period ended. The primary outcome measure was screen-detected breast cancer, stratified by screening technique (ie, digital breast tomosynthesis and digital mammography). A log-binomial regression model was used to estimate the efficacy of digital breast tomosynthesis versus digital mammography, defined as the crude risk ratios (RRs) with 95% CIs for screen-detected breast cancer for women screened during the recruitment period. A per-protocol approach was used in the analyses. This trial is registered at ClinicalTrials.gov, number NCT02835625, and is closed to accrual. FINDINGS: Between, Jan 14, 2016, and Dec 31, 2017, 44 266 women were invited to the screening programme in Bergen, and 32 976 (74·5%) attended. After excluding women with breast implants and women who did not consent to participate, 29 453 (89·3%) were eligible for electronic randomisation. 14 734 women were allocated to digital breast tomosynthesis and 14 719 to digital mammography. After randomisation, women with a previous breast cancer were excluded (digital breast tomosynthesis group n=314, digital mammography group n=316), women with metastases from melanoma (digital breast tomosynthesis group n=1), and women who informed the radiographer about breast symptoms after providing consent (digital breast tomosynthesis group n=39, digital mammography group n=34). After exclusions, information from 28 749 women were included in the analyses (digital breast tomosynthesis group n=14 380, digital mammography group n=14 369). The proportion of screen-detected breast cancer among the screened women did not differ between the two groups (95 [0·66%, 0·53-0·79] of 14 380 vs 87 [0·61%, 0·48-0·73] of 14 369; RR 1·09, 95% CI 0·82-1·46; p=0·56). INTERPRETATION: This study indicated that digital breast tomosynthesis including synthetic 2D mammograms was not significantly different from standard digital mammography as a screening tool for the detection of breast cancer in a population-based screening programme. Economic analyses and follow-up studies on interval and consecutive round screen-detected breast cancers are needed to better understand the effect of digital breast tomosynthesis in population-based breast cancer screening. FUNDING: Cancer Registry of Norway, Department of Radiology at Haukeland University Hospital, University of Oslo, and Research Council of Norway.

A Probabilistic Model to Support Radiologists' Classification Decisions in Mammography Practice.

We developed a probabilistic model to support the classification decisions made by radiologists in mammography practice. Using the feature observations and Breast Imaging Reporting and Data System (BI-RADS) classifications from radiologists examining diagnostic and screening mammograms, we modeled their decisions to understand their judgments. Our model could help improve the decisions made by radiologists using their own feature observations and classifications while maintaining their observed sensitivities. Based on 112,433 mammographic cases from 36,111 patients and 13 radiologists at 2 separate institutions with a 1.1% prevalence of malignancy, we trained a probabilistic Bayesian network (BN) to estimate the malignancy probabilities of lesions. For each radiologist, we learned an observed probabilistic threshold within the model. We compared the sensitivity and specificity of each radiologist against the BN model using either their observed threshold or the standard 2% threshold recommended by BI-RADS. We found significant variability among the radiologists' observed thresholds. By applying the observed thresholds, the BN model showed a 0.01% (1 case) increase in false negatives and a 28.9% (3612 cases) reduction in false positives. When using the standard 2% BI-RADS-recommended threshold, there was a 26.7% (47 cases) increase in false negatives and a 47.3% (5911 cases) reduction in false positives. Our results show that we can significantly reduce screening mammography false positives with a minimal increase in false negatives. We find that learning radiologists' observed thresholds provides valuable information regarding the conservativeness of clinical practice and allows us to quantify the variability in sensitivity across and within institutions. Our model could provide support to radiologists to improve their performance and consistency within mammography practice.

Variability of Breast Density Classification Between US and UK Radiologists.

PURPOSE: To assess whether subjective breast density categorization remains the most useful way to categorize mammographic breast density and whether variations exist across geographic regions with differing national legislation. METHODS: Breast radiologists from two countries (UK, USA) were voluntarily recruited to review sets of anonymized mammographic images (n = 180) and additional repeated images (n = 70), totaling 250 images, to subjectively rate breast density according to the Breast Imaging Reporting and Data system (BI-RADS) categorization. Images were reviewed using standardized viewing conditions and Ziltron software. Inter-rater reliability was analyzed using the Kappa test. RESULTS: The US radiologists (n = 25) judged fewer images as being "mostly fatty" than UK radiologists (n = 24), leading a greater number of images classified in the higher BI-RADS categories, particularly in BI-RADS 3. Overall agreement for all data sets was k = 0.654 indicating substantial agreement between the two cohorts. When the data were split into BI-RADS categories, the level of agreement varied from fair to substantial. CONCLUSION: Variations in how radiologists from the USA and UK classify breast density was established, especially when the data were divided into breast density categories. This variation supports the need for a reliable breast density assessment method to enhance the individualized supplemental screening pathways for dense breasts. The use of two-scale categorization method demonstrated improved agreement. ADVANCES IN KNOWLEDGE: Larger sample of radiologists from different breast density jurisdictions confirms international subjective variability in density categorization and improved agreement with the two-scale (low, high) categorization. With this variability, a standardized and automated breast density assessment shows to be timely.

Career caseload predicts interobserver agreement on the final classification of a mammogram.

INTRODUCTION: Differences in radiologists' experience can potentially introduce interobserver variability in reading mammograms. This work investigated the effect of radiologists' experience on agreement on mammographic final classification. METHODS: This was a cross-sectional study. Seventeen radiologists were asked to provide their final impression on 60 mammogram cases. Experience parameters included breast subspecialty, years reading mammograms, cases read per year and career caseload. Career caseload was calculated by multiplying years reading mammograms by the average number of cases read per year. The interobserver agreement was calculated using Cohen kappa (κ). The difference in κ between radiologists' groups was compared using the independent-sample t-test and analysis of variance. RESULTS: The average interobserver agreement was 0.25 (fair). A small difference was found in favour of breast radiologists against general radiologists (κ = 0.21 and 0.29, respectively, P = 0.019). Years reading mammograms and cases read per year did not seem to significantly affect the interobserver agreement (P = 0.056 and 0.273 respectively). Radiologist who had career caseload of at least 2500 cases showed significantly higher consistency than those who read less. κ for radiologists who had career caseload of 2500-4000 cases and >4000 cases was 0.33 and 0.28, respectively, whereas for <2500 κ was 0.17 (P = 0.001). CONCLUSION: A fair level of interobserver agreement on the final classification of a mammogram was demonstrated. Career caseload was the most important experience parameter to associate with the interobserver agreement. Training strategies aiming to increase radiologists' career caseload may be beneficial.

Comparative Benefit-to-Radiation Risk Ratio of Molecular Breast Imaging, Two-Dimensional Full-Field Digital Mammography with and without Tomosynthesis, and Synthetic Mammography with Tomosynthesis.

Purpose: To apply previously published benefit-to-risk ratio methods for mammography and molecular breast imaging (MBI) risk estimates to an expanded range of mammographic screening techniques, compressed breast thicknesses, and screening views. Materials and Methods: Only previously published estimates were used; therefore, this study was exempt from the requirement to obtain institutional review board approval. Benefit-to-risk ratios were calculated as the ratio of breast cancer deaths averted and lives lost to screening over 10-year intervals starting at age 40 years for MBI, two-dimensional (2D) full-field digital mammography (FFDM) alone, 2D FFDM with synthetic mammography, and 2D FFDM with tomosynthesis for two-, four-, and five-view screening mammography and compressed breast thicknesses of 20-29 mm, 50-59 mm, and 80-89 mm. Results: Central estimates of the benefit-to-risk ratios ranged from 3 to 179 for screening mammography and from 5 to 9 for MBI. Benefit-to-risk ratios for MBI were inferior to those for mammography for most scenarios, but MBI may be performed at an equal or superior benefit-to-risk ratio for women aged 40-59 years with a compressed breast thickness of at least 80 mm and for those undergoing mammographic screening examinations with four or five views per breast. The benefit-to-risk ratios across all ages with use of tomosynthesis plus 2D FFDM as a screening examination were 45% lower than those for tomosynthesis plus synthetic mammography. Conclusion: Benefit-to-risk ratios for MBI are within the lower range of those for mammography when accounting for variation in mammography technique, compressed breast thickness, and age. Benefit-to-risk ratios of synthetic mammography plus tomosynthesis are superior to those of tomosynthesis plus 2D FFDM.Keywords: Breast, Mammography, Molecular Imaging, Molecular Imaging-Cancer, Radiation Safety, Radionuclide Studies, Screening, Tomosynthesis© RSNA, 2019See also the commentary by Hruska in this issue.

Automatic Labeling of Special Diagnostic Mammography Views from Images and DICOM Headers.

Applying state-of-the-art machine learning techniques to medical images requires a thorough selection and normalization of input data. One of such steps in digital mammography screening for breast cancer is the labeling and removal of special diagnostic views, in which diagnostic tools or magnification are applied to assist in assessment of suspicious initial findings. As a common task in medical informatics is prediction of disease and its stage, these special diagnostic views, which are only enriched among the cohort of diseased cases, will bias machine learning disease predictions. In order to automate this process, here, we develop a machine learning pipeline that utilizes both DICOM headers and images to predict such views in an automatic manner, allowing for their removal and the generation of unbiased datasets. We achieve AUC of 99.72% in predicting special mammogram views when combining both types of models. Finally, we apply these models to clean up a dataset of about 772,000 images with expected sensitivity of 99.0%. The pipeline presented in this paper can be applied to other datasets to obtain high-quality image sets suitable to train algorithms for disease detection.

Assessment of mammographic breast density after sleeve gastrectomy.

BACKGROUND: Mammographic breast density (BD) is an independent risk factor for breast cancer. The effects of bariatric surgery on BD are unknown. OBJECTIVES: To investigate BD changes after sleeve gastrectomy (SG). SETTING: University hospital, United States. METHODS: Fifty women with mammograms before and after SG performed from 2009 to 2015 were identified after excluding patients with a history of breast cancer, hormone replacement, and/or breast surgery. Patient age, menopausal status, co-morbidities, hemoglobin A1C, and body mass index were collected. Craniocaudal mammographic views before and after SG were interpreted by a blinded radiologist and analyzed by software to obtain breast imaging reporting and data system density categories, breast area, BD, and absolute dense breast area (ADA). Analyses were performed using χ2, McNemar's test, t test, and linear regressions. RESULTS: Radiologist interpretation revealed a significant increase in breast imaging reporting and data system B+C category (68% versus 54%; P = .0095) and BD (9.8 ± 7.4% versus 8.3 ± 6.4%; P = .0006) after SG. Software analyses showed a postoperative decrease in breast area (75,398.9 ± 22,941.2 versus 90,655.9 ± 25,621.0 pixels; P < .0001) and ADA (7287.1 ± 3951.3 versus 8204.6 ± 4769.9 pixels; P = .0314) with no significant change in BD. Reduction in ADA was accentuated in postmenopausal patients. Declining breast area was directly correlated with body mass index reduction (R2 = .4495; P < 0.0001). Changes in breast rather than whole body adiposity better explained ADA reduction. Neither diabetes status nor changes in hemoglobin A1C correlated with changes in ADA. CONCLUSIONS: ADA decreases after SG, particularly in postmenopausal patients. Software-generated ADA may be more accurate than radiologist-estimated BD or breast imaging reporting and data system for capturing changes in dense breast tissue after SG.

Proposing New RadLex Terms by Analyzing Free-Text Mammography Reports.

After years of development, the RadLex terminology contains a large set of controlled terms for the radiology domain, but gaps still exist. We developed a data-driven approach to discover new terms for RadLex by mining a large corpus of radiology reports using natural language processing (NLP) methods. Our system, developed for mammography, discovers new candidate terms by analyzing noun phrases in free-text reports to extend the mammography part of RadLex. Our NLP system extracts noun phrases from free-text mammography reports and classifies these noun phrases as "Has Candidate RadLex Term" or "Does Not Have Candidate RadLex Term." We tested the performance of our algorithm using 100 free-text mammography reports. An expert radiologist determined the true positive and true negative RadLex candidate terms. We calculated precision/positive predictive value and recall/sensitivity metrics to judge the system's performance. Finally, to identify new candidate terms for enhancing RadLex, we applied our NLP method to 270,540 free-text mammography reports obtained from three academic institutions. Our method demonstrated precision/positive predictive value of 0.77 (159/206 terms) and a recall/sensitivity of 0.94 (159/170 terms). The overall accuracy of the system is 0.80 (235/293 terms). When we ran our system on the set of 270,540 reports, it found 31,800 unique noun phrases that are potential candidates for RadLex. Our data-driven approach to mining radiology reports can identify new candidate terms for expanding the breast imaging lexicon portion of RadLex and may be a useful approach for discovering new candidate terms from other radiology domains.

A comparison of five methods of measuring mammographic density: a case-control study.

BACKGROUND: High mammographic density is associated with both risk of cancers being missed at mammography, and increased risk of developing breast cancer. Stratification of breast cancer prevention and screening requires mammographic density measures predictive of cancer. This study compares five mammographic density measures to determine the association with subsequent diagnosis of breast cancer and the presence of breast cancer at screening. METHODS: Women participating in the "Predicting Risk Of Cancer At Screening" (PROCAS) study, a study of cancer risk, completed questionnaires to provide personal information to enable computation of the Tyrer-Cuzick risk score. Mammographic density was assessed by visual analogue scale (VAS), thresholding (Cumulus) and fully-automated methods (Densitas, Quantra, Volpara) in contralateral breasts of 366 women with unilateral breast cancer (cases) detected at screening on entry to the study (Cumulus 311/366) and in 338 women with cancer detected subsequently. Three controls per case were matched using age, body mass index category, hormone replacement therapy use and menopausal status. Odds ratios (OR) between the highest and lowest quintile, based on the density distribution in controls, for each density measure were estimated by conditional logistic regression, adjusting for classic risk factors. RESULTS: The strongest predictor of screen-detected cancer at study entry was VAS, OR 4.37 (95% CI 2.72-7.03) in the highest vs lowest quintile of percent density after adjustment for classical risk factors. Volpara, Densitas and Cumulus gave ORs for the highest vs lowest quintile of 2.42 (95% CI 1.56-3.78), 2.17 (95% CI 1.41-3.33) and 2.12 (95% CI 1.30-3.45), respectively. Quantra was not significantly associated with breast cancer (OR 1.02, 95% CI 0.67-1.54). Similar results were found for subsequent cancers, with ORs of 4.48 (95% CI 2.79-7.18), 2.87 (95% CI 1.77-4.64) and 2.34 (95% CI 1.50-3.68) in highest vs lowest quintiles of VAS, Volpara and Densitas, respectively. Quantra gave an OR in the highest vs lowest quintile of 1.32 (95% CI 0.85-2.05). CONCLUSIONS: Visual density assessment demonstrated a strong relationship with cancer, despite known inter-observer variability; however, it is impractical for population-based screening. Percentage density measured by Volpara and Densitas also had a strong association with breast cancer risk, amongst the automated measures evaluated, providing practical automated methods for risk stratification.

Similar to Adiponectin, Serum Levels of Osteocalcin are Associated with Mammographic Breast Density in Postmenopausal Women.

OBJECTIVE: Breast cancer is the most common type of cancer in Canadian women and worldwide. Mammographic density is a well-established breast cancer risk. Recent evidence suggested inverse correlations among adiponectin, osteocalcin, and the risk developing breast cancer. The objective of the study was to evaluate the relationship between breast density and adiponectin and osteocalcin concentrations. METHODS: A cross-sectional study was performed in 239 women, age range 40 to 60. Mammographic density, serum adiponectin, and osteocalcin levels were measured. According to the Wolfe method, participants were divided into those with low-risk and high-risk pattern mammograms. RESULTS: The study population included 107 premenopausal and 132 postmenopausal women. Parameters were no different between women with low-risk and high-risk patterns. In obese postmenopausal women, the high-risk pattern mammogram group had significantly higher values of adiponectin and osteocalcin compared with the low-risk pattern group. Multiple linear regression analyses showed that adiponectin and osteocalcin levels were associated with high-risk pattern mammograms. CONCLUSION: Adiponectin and osteocalcin levels were directly associated with high-risk pattern mammograms in obese postmenopausal women. These results do not support the use of adipokines as biomarkers; nevertheless, the most important factor is to assess the risk through breast density.

Summary measures of agreement and association between many raters' ordinal classifications.

PURPOSE: Interpretation of screening tests such as mammograms usually require a radiologist's subjective visual assessment of images, often resulting in substantial discrepancies between radiologists' classifications of subjects' test results. In clinical screening studies to assess the strength of agreement between experts, multiple raters are often recruited to assess subjects' test results using an ordinal classification scale. However, using traditional measures of agreement in some studies is challenging because of the presence of many raters, the use of an ordinal classification scale, and unbalanced data. METHODS: We assess and compare the performances of existing measures of agreement and association as well as a newly developed model-based measure of agreement to three large-scale clinical screening studies involving many raters' ordinal classifications. We also conduct a simulation study to demonstrate the key properties of the summary measures. RESULTS: The assessment of agreement and association varied according to the choice of summary measure. Some measures were influenced by the underlying prevalence of disease and raters' marginal distributions and/or were limited in use to balanced data sets where every rater classifies every subject. Our simulation study indicated that popular measures of agreement and association are prone to underlying disease prevalence. CONCLUSIONS: Model-based measures provide a flexible approach for calculating agreement and association and are robust to missing and unbalanced data as well as the underlying disease prevalence.

Malignancy Detection on Mammography Using Dual Deep Convolutional Neural Networks and Genetically Discovered False Color Input Enhancement.

Breast cancer is the most prevalent malignancy in the US and the third highest cause of cancer-related mortality worldwide. Regular mammography screening has been attributed with doubling the rate of early cancer detection over the past three decades, yet estimates of mammographic accuracy in the hands of experienced radiologists remain suboptimal with sensitivity ranging from 62 to 87% and specificity from 75 to 91%. Advances in machine learning (ML) in recent years have demonstrated capabilities of image analysis which often surpass those of human observers. Here we present two novel techniques to address inherent challenges in the application of ML to the domain of mammography. We describe the use of genetic search of image enhancement methods, leading us to the use of a novel form of false color enhancement through contrast limited adaptive histogram equalization (CLAHE), as a method to optimize mammographic feature representation. We also utilize dual deep convolutional neural networks at different scales, for classification of full mammogram images and derivative patches combined with a random forest gating network as a novel architectural solution capable of discerning malignancy with a specificity of 0.91 and a specificity of 0.80. To our knowledge, this represents the first automatic stand-alone mammography malignancy detection algorithm with sensitivity and specificity performance similar to that of expert radiologists.

Pareto-optimal multi-objective dimensionality reduction deep auto-encoder for mammography classification.

BACKGROUND AND OBJECTIVE: Feature reduction is an essential stage in computer aided breast cancer diagnosis systems. Multilayer neural networks can be trained to extract relevant features by encoding high-dimensional data into low-dimensional codes. Optimizing traditional auto-encoders works well only if the initial weights are close to a proper solution. They are also trained to only reduce the mean squared reconstruction error (MRE) between the encoder inputs and the decoder outputs, but do not address the classification error. The goal of the current work is to test the hypothesis that extending traditional auto-encoders (which only minimize reconstruction error) to multi-objective optimization for finding Pareto-optimal solutions provides more discriminative features that will improve classification performance when compared to single-objective and other multi-objective approaches (i.e. scalarized and sequential). METHODS: In this paper, we introduce a novel multi-objective optimization of deep auto-encoder networks, in which the auto-encoder optimizes two objectives: MRE and mean classification error (MCE) for Pareto-optimal solutions, rather than just MRE. These two objectives are optimized simultaneously by a non-dominated sorting genetic algorithm. RESULTS: We tested our method on 949 X-ray mammograms categorized into 12 classes. The results show that the features identified by the proposed algorithm allow a classification accuracy of up to 98.45%, demonstrating favourable accuracy over the results of state-of-the-art methods reported in the literature. CONCLUSIONS: We conclude that adding the classification objective to the traditional auto-encoder objective and optimizing for finding Pareto-optimal solutions, using evolutionary multi-objective optimization, results in producing more discriminative features.

Interobserver variability in upgraded and non-upgraded BI-RADS 3 lesions.

AIM: To evaluate interobserver variability in the assessment of Breast Imaging-Reporting and Data System (BI-RADS) 3 mammographic lesions, and to determine if the initial evaluation of upgraded BI-RADS 3 lesions was appropriate. MATERIALS AND METHODS: Retrospective review of the mammography database (1/1/2004-12/31/2008) identified 1,188 screen-detected BI-RADS 3 lesions, 60 (5.1%) were upgraded to BI-RADS 4/5 during surveillance (cases). Cases were matched to 60 non-upgraded BI-RADS 3 lesions (controls) by lesion type, laterality, and year. Available studies were assessed separately by two radiologists blinded to outcomes. RESULTS: Eighty-two studies were available (43 cases, eight malignancies, and 39 controls). Reader 1 assessed 18/82 (22%) as BI-RADS 0, 13 cases, five controls; 35/82 (42.7%) as BI-RADS 2, 11 cases, 24 controls; 7/82 (8.5%) BI-RADS 3, four cases, three controls; 22/82 BI-RADS 4, 15 cases, seven controls. Reader 2 assessed 8/82 (9.8%) as BI-RADS 0, four cases, four controls; 27 (32.9%) BI-RADS 2, 11 cases, 16 controls; 33 (40.2%) BI-RADS 3, 19 cases, 14 controls; 14 (17%) BI-RADS 4, nine cases, five controls. For cancers, reader 1 assessed two BI-RADS 0, one BI-RADS 2, one BI-RADS 3, and four BI-RADS 4; reader 2 assessed two BI-RADS 2, four BI-RADS 3, and two BI-RADS 4. Reasons for BI-RADS 0 assessment included incomplete mammographic views, lack of ultrasound, and failure to include the lesion on follow-up imaging. Reasons for BI-RADS 4 assessment included suspicious morphology or instability. CONCLUSION: There is much interobserver variability in the assessment of BI-RADS 3 lesions. Many BI-RADS 3 lesions were judged as incompletely evaluated on blinded review.

Differential Data Augmentation Techniques for Medical Imaging Classification Tasks.

Data augmentation is an essential part of training discriminative Convolutional Neural Networks (CNNs). A variety of augmentation strategies, including horizontal flips, random crops, and principal component analysis (PCA), have been proposed and shown to capture important characteristics of natural images. However, while data augmentation has been commonly used for deep learning in medical imaging, little work has been done to determine which augmentation strategies best capture medical image statistics, leading to more discriminative models. This work compares augmentation strategies and shows that the extent to which an augmented training set retains properties of the original medical images determines model performance. Specifically, augmentation strategies such as flips and gaussian filters lead to validation accuracies of 84% and 88%, respectively. On the other hand, a less effective strategy such as adding noise leads to a significantly worse validation accuracy of 66%. Finally, we show that the augmentation affects mass generation.