Transfer learning for the generalization of artificial intelligence in breast cancer detection: a case-control study.

BACKGROUND: Some researchers have questioned whether artificial intelligence (AI) systems maintain their performance when used for women from populations not considered during the development of the system. PURPOSE: To evaluate the impact of transfer learning as a way of improving the generalization of AI systems in the detection of breast cancer. MATERIAL AND METHODS: This retrospective case-control Finnish study involved 191 women diagnosed with breast cancer and 191 matched healthy controls. We selected a state-of-the-art AI system for breast cancer detection trained using a large US dataset. The selected baseline system was evaluated in two experimental settings. First, we examined our private Finnish sample as an independent test set that had not been considered in the development of the system (unseen population). Second, the baseline system was retrained to attempt to improve its performance in the unseen population by means of transfer learning. To analyze performance, we used areas under the receiver operating characteristic curve (AUCs) with DeLong's test. RESULTS: Two versions of the baseline system were considered: ImageOnly and Heatmaps. The ImageOnly and Heatmaps versions yielded mean AUC values of 0.82±0.008 and 0.88±0.003 in the US dataset and 0.56 (95% CI=0.50-0.62) and 0.72 (95% CI=0.67-0.77) when evaluated in the unseen population, respectively. The retrained systems achieved AUC values of 0.61 (95% CI=0.55-0.66) and 0.69 (95% CI=0.64-0.75), respectively. There was no statistical difference between the baseline system and the retrained system. CONCLUSION: Transfer learning with a small study sample did not yield a significant improvement in the generalization of the system.

Saliency of breast lesions in breast cancer detection using artificial intelligence.

The analysis of mammograms using artificial intelligence (AI) has shown great potential for assisting breast cancer screening. We use saliency maps to study the role of breast lesions in the decision-making process of AI systems for breast cancer detection in screening mammograms. We retrospectively collected mammograms from 191 women with screen-detected breast cancer and 191 healthy controls matched by age and mammographic system. Two radiologists manually segmented the breast lesions in the mammograms from CC and MLO views. We estimated the detection performance of four deep learning-based AI systems using the area under the ROC curve (AUC) with a 95% confidence interval (CI). We used automatic thresholding on saliency maps from the AI systems to identify the areas of interest on the mammograms. Finally, we measured the overlap between these areas of interest and the segmented breast lesions using Dice's similarity coefficient (DSC). The detection performance of the AI systems ranged from low to moderate (AUCs from 0.525 to 0.694). The overlap between the areas of interest and the breast lesions was low for all the studied methods (median DSC from 4.2% to 38.0%). The AI system with the highest cancer detection performance (AUC = 0.694, CI 0.662-0.726) showed the lowest overlap (DSC = 4.2%) with breast lesions. The areas of interest found by saliency analysis of the AI systems showed poor overlap with breast lesions. These results suggest that AI systems with the highest performance do not solely rely on localized breast lesions for their decision-making in cancer detection; rather, they incorporate information from large image regions. This work contributes to the understanding of the role of breast lesions in cancer detection using AI.

An in silico study on the detectability of field cancerization through parenchymal analysis of digital mammograms.

BACKGROUND: Parenchymal analysis has shown promising performance for the assessment of breast cancer risk through the characterization of the texture features of mammography images. However, the working principles behind this practice are yet not well understood. Field cancerization is a phenomenon associated with genetic and epigenetic alterations in large volumes of cells, putting them on a path of malignancy before the appearance of recognizable cancer signs. Evidence suggests that it can induce changes in the biochemical and optical properties of the tissue. PURPOSE: The aim of this work was to study whether the extended genetic mutations and epigenetic changes due to field cancerization, and the impact they have on the biochemistry of breast tissues are detectable in the radiological patterns of mammography images. METHODS: An in silico experiment was designed, which implied the development of a field cancerization model to modify the optical tissue properties of a cohort of 60 voxelized virtual breast phantoms. Mammography images from these phantoms were generated and compared with images obtained from their non-modified counterparts, that is, without field cancerization. We extracted 33 texture features from the breast area to quantitatively assess the impact of the field cancerization model. We analyzed the similarity and statistical equivalence of texture features with and without field cancerization using the t-test, Wilcoxon sign rank test and Kolmogorov-Smirnov test, and performed a discrimination test using multinomial logistic regression analysis with lasso regularization. RESULTS: With modifications of the optical tissue properties on 3.9% of the breast volume, some texture features started to fail to show equivalence (p < 0.05). At 7.9% volume modification, a high percent of texture features showed statistically significant differences (p < 0.05) and non-equivalence. At this level, multinomial logistic regression analysis of texture features showed a statistically significant performance in the discrimination of mammograms from breasts with and without field cancerization (AUC = 0.89, 95% CI: 0.75-1.00). CONCLUSIONS: These results support the idea that field cancerization is a feasible underlying working principle behind the distinctive performance of parenchymal analysis in breast cancer risk assessment.

Image retrieval-based parenchymal analysis for breast cancer risk assessment.

PURPOSE: This research on breast cancer risk assessment aims to develop models that predict the likelihood of breast cancer. In recent years, the computerized analysis of visual texture patterns in mammograms, namely parenchymal analysis, has shown great potential for risk assessment. However, the visual complexity and heterogeneity of visual patterns limit the performance of parenchymal analysis in large populations. In this work, we propose a method to create individualized risk assessment models based on the radiological visual appearance (radiomic phenotypes) of the mammograms. METHODS: We developed a content-based image retrieval system to stratify mammographic analysis according to the similarities of their radiomic phenotypes. We collected 1144 mammograms from 286 women following a case-control study design. We compared the classical parenchymal analysis with the proposed approach using the area under the ROC curve (AUC) with 95% confidence intervals (CI). Statistical significance was assessed using DeLong's test ( p < $p<$ 0.05). RESULTS: At a patient level, AUC values of 0.504 (95% CI: 0.398-0.611) with classical parenchymal analysis increased to 0.813 (95% CI: 0.734-0.892) when the radiomic phenotypes are incorporated with the proposed method. In risk estimation from individual, standard mammographic views, the highest performance was obtained with the mediolateral oblique view of the right breast (RMLO), with an AUC value of 0.727 (95% CI: 0.634-0.820). Differences in performance among views were statistically significant ( p < 0.05 $p<0.05$ ) CONCLUSIONS: These results indicate that the utilization of radiomic phenotypes increases the performance of computerized risk assessment based on parenchymal analysis of mammographic images. SIGNIFICANCE: The creation of individualized risk assessment models may be leveraged to target personalized screening and prevention recommendations according to the person's risk.

Methodology, clinical applications, and future directions of body composition analysis using computed tomography (CT) images: A review.

PURPOSE OF THE REVIEW: We aim to review the methods, current research evidence, and future directions in body composition analysis (BCA) with CT imaging. RECENT FINDINGS: CT images can be used to evaluate muscle tissue, visceral adipose tissue (VAT), and subcutaneous adipose tissue (SAT) compartments. Manual and semiautomatic segmentation methods are still the gold standards. The segmentation of skeletal muscle tissue and VAT and SAT compartments is most often performed at the level of the 3rd lumbar vertebra. A decreased amount of CT-determined skeletal muscle mass is a marker of impaired survival in many patient populations, including patients with most types of cancer, some surgical patients, and those admitted to the intensive care unit (ICU). Patients with increased VAT are more susceptible to impaired survival / worse outcomes; however, those patients who are critically ill or admitted to the ICU or who will undergo surgery appear to be exceptions. The independent significance of SAT is less well established. Recently, the roles of the CT-determined decrease of muscle mass and increased VAT area and epicardial adipose tissue (EAT) volume have been shown to predict a more debilitating course of illness in patients suffering from severe acute respiratory syndrome coronavirus 2 (COVID-19) infection. SUMMARY: The field of CT-based body composition analysis is rapidly evolving and shows great potential for clinical implementation.

Algorithms and methods for computerized analysis of mammography images in breast cancer risk assessment.

BACKGROUND AND OBJECTIVES: The computerized analysis of mammograms for the development of quantitative biomarkers is a growing field with applications in breast cancer risk assessment. Computerized image analysis offers the possibility of using different methods and algorithms to extract additional information from screening and diagnosis images to aid in the assessment of breast cancer risk. In this work, we review the algorithms and methods for the automated, computerized analysis of mammography images for the task mentioned, and discuss the main challenges that the development and improvement of these methods face today. METHODS: We review the recent progress in two main branches of mammography-based risk assessment: parenchymal analysis and breast density estimation, including performance indicators of most of the studies considered. Parenchymal analysis methods are divided into feature-based methods and deep learning-based methods; breast density methods are grouped into area-based, volume-based, and breast categorization methods. Additionally, we identify the challenges that these study fields currently face. RESULTS: Parenchymal analysis using deep learning algorithms are on the rise, with some studies showing high-performance indicators, such as an area under the receiver operating characteristic curve of up to 90. Methods for risk assessment using breast density report a wider variety of performance indicators; however, we can also identify that the approaches using deep learning methods yield high performance in each of the subdivisions considered. CONCLUSIONS: Both breast density estimation and parenchymal analysis are promising tools for the task of breast cancer risk assessment; deep learning methods have shown performance comparable or superior to the other considered methods. All methods considered face challenges such as the lack of objective comparison between them and the lack of access to datasets from different populations.

A New Benchmark and Method for the Evaluation of Chest Wall Detection in Digital Mammography.

CAD systems have shown good potential for improving breast cancer diagnosis and anomaly detection in mammograms. A basic enabling step for the utilization of CAD systems in mammographic analysis is the correct identification of the breast region. Therefore, several methods to segment the pectoral muscle in the medio-lateral oblique (MLO) mammographic view have been proposed in the literature. However, currently it is difficult to perform and objective comparison between different chest wall (CW) detection methods since they are often evaluated with different evaluation procedures, datasets and the implementations of the methods are not publicly available. For this reason, we propose a methodology to evaluate and compare the performance of CW detection methods using a publicly available dataset (INbreast). We also propose a new intensity-based method for automatic CW detection. We then utilize the proposed evaluation methodology to compare the performance of our CW detection algorithm with a state-of-the-art CW detection method. The performance was measured in terms of the Dice's coefficient similarity, the area error and mean contour distance. The proposed method achieves yielded the best results in all the performance measures.

A Comparison of Regions of Interest in Parenchymal Analysis for Breast Cancer Risk Assessment.

Computerized parenchymal analysis has shown potential to be utilized as an imaging biomarker to estimate the risk of breast cancer. Parenchymal analysis of digital mammograms is based on the extraction of computerized measures to build machine learning-based models for the prediction of breast cancer risk. However, the choice of the region of interest (ROI) for feature extraction within the breast remains an open problem. In this work we perform a comparison between five different methods suggested in the literature for automated ROI selection, including the whole breast (WB), the maximum squared (MS), the retro-areolar region (RA), the lattice-based (LB), and the polar-based (PB) selection methods. For the experiments, we built a retrospective dataset of 896 screening mammograms from 224 women (112 cases and 112 healthy controls). The performance of each ROI selection method was measured in terms of the area under the curve (AUC) values. The AUC values varied between 0.55 and 0.79 depending on the method and experimental settings. The best performance on an independent test set was achieved by the MS method (AUC of 0.59, 95% CI: 0.55-0.64). This method is fully-automated and does not require adjusting hyper-parameters. Based on our results, we prompt the use of the MS method for ROI selection in the computerized parenchymal analysis for breast cancer risk assessment.

Field cancerization in the understanding of parenchymal analysis of mammograms for breast cancer risk assessment.

In recent years, mammographic image analysis has shown great potential for breast cancer risk assessment. The aim of risk assessment is to predict how likely a woman is to develop breast cancer in the future. Several studies suggest that computerized parenchymal analysis of mammograms can be utilized as an independent imaging biomarker of breast cancer. Parenchymal analysis consists of the quantitative assessment of visual texture patterns in mammograms to infer the level of risk. In spite of substantial evidence of the association between parenchymal patterns and breast cancer risk, its biological foundations remain poorly understood. In this work, we draw a hypothesis that links the field cancerization (FC) with breast cancer risk assessment based on the parenchymal analysis. In the literature, the FC is interpreted as a biochemical anomaly amplification in otherwise healthy cells due to the effect of pre-cancerous transformed cells in surrounding regions. Our hypothesis is that these biochemical anomaly amplifications change the cellular micro-environment which, in turn, alter tissue responses to X-ray radiation. As a result, it is reasonable to think that these changes influence the interaction of X-rays with parenchymal - the functional - breast tissue thus enabling cancer prediction by analyzing X-ray images of the breast. We believe that our hypothesis provides an actionable explanation as to how computerized parenchymal analysis of apparently normal mammograms can be successfully utilized for the stratification of breast cancer risk.

Clinical evaluation of a fully-automated parenchymal analysis software for breast cancer risk assessment: A pilot study in a Finnish sample.

PURPOSE: To assess the association between breast cancer risk and mammographic parenchymal measures obtained using a fully-automated, publicly available software, OpenBreast. METHODS: This retrospective case-control study involved screening mammograms of asymptomatic women diagnosed with breast cancer between 2016 and 2017. The 114 cases were matched with corresponding healthy controls by birth and screening years and the mammographic system used. Parenchymal analysis was performed using OpenBreast, a software implementing a computerized parenchymal analysis algorithm. Breast percent density was measured with an interactive thresholding method. The parenchymal measures were Box-Cox transformed and adjusted for age and percent density. Changes in the odds ratio per standard deviation (OPERA) with 95% confidence intervals (CIs) and the area under the ROC curve (AUC) for parenchymal measures and percent densities were used to evaluate the discrimination between cases and controls. Differences in AUCs were assessed using DeLong's test. RESULTS: The adjusted OPERA value of parenchymal measures was 2.49 (95% CI: 1.79-3.47). Parenchymal measures using OpenBreast were more accurate (AUC = 0.779) than percent density (AUC = 0.609) in discriminating between cases and controls (p < 0.001). CONCLUSIONS: Parenchymal measures obtained with the evaluated software were positively associated with breast cancer risk and were more accurate than percent density in the prediction of risk.

Measurement challenge: protocol for international case-control comparison of mammographic measures that predict breast cancer risk.

INTRODUCTION: For women of the same age and body mass index, increased mammographic density is one of the strongest predictors of breast cancer risk. There are multiple methods of measuring mammographic density and other features in a mammogram that could potentially be used in a screening setting to identify and target women at high risk of developing breast cancer. However, it is unclear which measurement method provides the strongest predictor of breast cancer risk. METHODS AND ANALYSIS: The measurement challenge has been established as an international resource to offer a common set of anonymised mammogram images for measurement and analysis. To date, full field digital mammogram images and core data from 1650 cases and 1929 controls from five countries have been collated. The measurement challenge is an ongoing collaboration and we are continuing to expand the resource to include additional image sets across different populations (from contributors) and to compare additional measurement methods (by challengers). The intended use of the measurement challenge resource is for refinement and validation of new and existing mammographic measurement methods. The measurement challenge resource provides a standardised dataset of mammographic images and core data that enables investigators to directly compare methods of measuring mammographic density or other mammographic features in case/control sets of both raw and processed images, for the purposes of the comparing their predictions of breast cancer risk. ETHICS AND DISSEMINATION: Challengers and contributors are required to enter a Research Collaboration Agreement with the University of Melbourne prior to participation in the measurement challenge. The Challenge database of collated data and images are stored in a secure data repository at the University of Melbourne. Ethics approval for the measurement challenge is held at University of Melbourne (HREC ID 0931343.3).

Morphological Area Gradient: System-independent Dense Tissue Segmentation in Mammography Images.

Breast density has been identified as one of the strongest risk factors for breast cancer. However, the development of reliable and reproducible methods for the automatic dense tissue segmentation has been an important challenge. Due to the complexity of the acquisition process of mammography images, current approaches need to be calibrated for specific mammographic systems or require access to raw mammograms. In this work, we introduce the Morphological Area Gradient (MAG) as a generic measure for mammography images. MAG is generic in the sense that it does not need calibration or access to raw mammograms. At the core of MAG is the derivative of the area of segmented tissue with respect to the pixel intensity. We have found that the high-density regions can be automatically segmented by minimizing the MAG of a mammogram. To verify the performance of MAG, we collected 566 full-field digital mammograms using two different medical devices and a human expert manually annotated the high-density regions in each image. The proposed MAG method yields a median absolute error of 7.6% and a Dices similarity coefficient of 0.83, which are superior to other clinically validated state-of-the-art algorithms.

Do Mammographic Systems Affect the Performance of Computerized Parenchymal Analysis?

Early identification of women at high risk of developing breast cancer is fundamental for timely diagnosis and treatment. Recently, researchers have demonstrated that the computerized analysis of parenchymal (breast tissue) patterns in mammograms can be utilized to assess the risk level of patients. However, parenchymal analysis being an image-based biomarker, its performance may be affected by the acquisition parameters of the mammogram. Unfortunately, research on the effect of the mammographic system on the performance of parenchymal analysis is very scarce. In this paper, we implement a parenchymal analysis algorithm and study the effect of different mammographic systems on its performance. We show in a setting of 286 women that the use of different mammographic systems can yield differences of up to 24% in the area under the ROC curve. Results suggest the the construction of models for risk assessment based on parenchymal analysis should incorporate the imaging technologies into the analysis.

The Impact of Acquisition Dose on Quantitative Breast Density Estimation with Digital Mammography: Results from ACRIN PA 4006.

Purpose To investigate the impact of radiation dose on breast density estimation in digital mammography. Materials and Methods With institutional review board approval and Health Insurance Portability and Accountability Act compliance under waiver of consent, a cohort of women from the American College of Radiology Imaging Network Pennsylvania 4006 trial was retrospectively analyzed. All patients underwent breast screening with a combination of dose protocols, including standard full-field digital mammography, low-dose digital mammography, and digital breast tomosynthesis. A total of 5832 images from 486 women were analyzed with previously validated, fully automated software for quantitative estimation of density. Clinical Breast Imaging Reporting and Data System (BI-RADS) density assessment results were also available from the trial reports. The influence of image acquisition radiation dose on quantitative breast density estimation was investigated with analysis of variance and linear regression. Pairwise comparisons of density estimations at different dose levels were performed with Student t test. Agreement of estimation was evaluated with quartile-weighted Cohen kappa values and Bland-Altman limits of agreement. Results Radiation dose of image acquisition did not significantly affect quantitative density measurements (analysis of variance, P = .37 to P = .75), with percent density demonstrating a high overall correlation between protocols (r = 0.88-0.95; weighted κ = 0.83-0.90). However, differences in breast percent density (1.04% and 3.84%, P < .05) were observed within high BI-RADS density categories, although they were significantly correlated across the different acquisition dose levels (r = 0.76-0.92, P < .05). Conclusion Precision and reproducibility of automated breast density measurements with digital mammography are not substantially affected by variations in radiation dose; thus, the use of low-dose techniques for the purpose of density estimation may be feasible. (©) RSNA, 2016 Online supplemental material is available for this article.

Fully Automated Quantitative Estimation of Volumetric Breast Density from Digital Breast Tomosynthesis Images: Preliminary Results and Comparison with Digital Mammography and MR Imaging.

PURPOSE: To assess a fully automated method for volumetric breast density (VBD) estimation in digital breast tomosynthesis (DBT) and to compare the findings with those of full-field digital mammography (FFDM) and magnetic resonance (MR) imaging. MATERIALS AND METHODS: Bilateral DBT images, FFDM images, and sagittal breast MR images were retrospectively collected from 68 women who underwent breast cancer screening from October 2011 to September 2012 with institutional review board-approved, HIPAA-compliant protocols. A fully automated computer algorithm was developed for quantitative estimation of VBD from DBT images. FFDM images were processed with U.S. Food and Drug Administration-cleared software, and the MR images were processed with a previously validated automated algorithm to obtain corresponding VBD estimates. Pearson correlation and analysis of variance with Tukey-Kramer post hoc correction were used to compare the multimodality VBD estimates. RESULTS: Estimates of VBD from DBT were significantly correlated with FFDM-based and MR imaging-based estimates with r = 0.83 (95% confidence interval [CI]: 0.74, 0.90) and r = 0.88 (95% CI: 0.82, 0.93), respectively (P < .001). The corresponding correlation between FFDM and MR imaging was r = 0.84 (95% CI: 0.76, 0.90). However, statistically significant differences after post hoc correction (α = 0.05) were found among VBD estimates from FFDM (mean ± standard deviation, 11.1% ± 7.0) relative to MR imaging (16.6% ± 11.2) and DBT (19.8% ± 16.2). Differences between VDB estimates from DBT and MR imaging were not significant (P = .26). CONCLUSION: Fully automated VBD estimates from DBT, FFDM, and MR imaging are strongly correlated but show statistically significant differences. Therefore, absolute differences in VBD between FFDM, DBT, and MR imaging should be considered in breast cancer risk assessment.

Generation of all-in-focus images by noise-robust selective fusion of limited depth-of-field images.

The limited depth-of-field of some cameras prevents them from capturing perfectly focused images when the imaged scene covers a large distance range. In order to compensate for this problem, image fusion has been exploited for combining images captured with different camera settings, thus yielding a higher quality all-in-focus image. Since most current approaches for image fusion rely on maximizing the spatial frequency of the composed image, the fusion process is sensitive to noise. In this paper, a new algorithm for computing the all-in-focus image from a sequence of images captured with a low depth-of-field camera is presented. The proposed approach adaptively fuses the different frames of the focus sequence in order to reduce noise while preserving image features. The algorithm consists of three stages: 1) focus measure; 2) selectivity measure; 3) and image fusion. An extensive set of experimental tests has been carried out in order to compare the proposed algorithm with state-of-the-art all-in-focus methods using both synthetic and real sequences. The obtained results show the advantages of the proposed scheme even for high levels of noise.