Artificial Intelligence Applications in Breast Imaging

Version	Summary	Created by	Modification	Content Size	Created at	Operation
1		Tara A. Retson	--	2626	2023-07-03 18:50:36	\|
2	layout	Jessie Wu	+ 7 word(s)	2633	2023-07-04 04:41:49	\|

This entry is adapted from the peer-reviewed paper 10.3390/diagnostics13132133

Artificial intelligence (AI) applications in mammography have gained significant popular attention; however, AI has the potential to revolutionize other aspects of breast imaging beyond simple lesion detection. AI has the potential to enhance risk assessment by combining conventional factors with imaging and improve lesion detection through a comparison with prior studies and considerations of symmetry. It also holds promise in ultrasound analysis and automated whole breast ultrasound, areas marked by unique challenges. AI’s potential utility also extends to administrative tasks such as Mammography Quality Standards Act (MQSA) compliance, scheduling, and protocoling, which can reduce the radiologists’ workload. However, adoption in breast imaging faces limitations in terms of data quality and standardization, generalizability, benchmarking performance, and integration into clinical workflows.

artificial intelligence breast imaging study comparison beyond mammography MQSA

1. Introduction

In the rapidly evolving field of medical imaging, artificial intelligence (AI) has emerged as a powerful tool with the potential to revolutionize diagnosis, quantitative tasks, and numerous aspects of clinical practice. Breast imaging has always been at the forefront of embracing and incorporating technological advances in radiology, and AI is no exception. The AI applications for mammography have been widely explored in both the research and commercial realms, resulting, at times, in public fanfare. For example, in 2020 a paper by McKinney et al., backed in part by Google, propelled breast imaging AI into the public spotlight, as it showed that an AI algorithm could outperform radiologists in predicting breast cancer on screening mammography ^[1]. Although most research and product development in breast imaging have focused on cancer detection or density assessment on screening mammography, a radiologist’s responsibilities extend beyond these tasks to include various modalities and clinical and administrative duties.

2. Integrating Information

2.1. Fusing Clinical Data and Imaging

Much like the multifaceted analysis of radiomics, a mammographer’s role requires the synthesis of many factors beyond images. Mammographers rarely look at imaging in isolation; rather, they adopt a more holistic approach, considering a patient’s clinical information alongside imaging to render a diagnosis. In contrast, most commercially available software, particularly current CAD applications, analyze images without incorporating clinical information or risk models. The concept of software that integrates clinical information is rapidly gaining traction in other fields of radiology, with studies that integrated imaging and medical records doubling between 2020 and 2021 ^[2]. Although currently most common in the study of neurological disorders, the successful fusion of image data with non-imaging data was demonstrated in basal cell carcinoma detection by Kharazmi et al., pulmonary embolism detection by Huang et al. (2020), and prostate cancer detection on MRI images fused with the level of prostate-specific antigen by Reda et al. ^[3]^[4]^[5].

Three different strategies known as early fusion, joint fusion, and late fusion, are described in several reviews, including Huang et al. and Mohsen et al. ^[2]^[5]. The optimal approach and the most important information from images and clinical data will be determined by the specific task, with the intuitive integration of clinical information poised to enhance algorithmic performance and improve clinical care. Mohsen et al.’s. review highlights the overall success of fusion strategies, with fusion studies outperforming single-modality approaches when applied to the same tasks ^[2]. For instance, in a study by Reda et al. the integration of clinical information with imaging was able to achieve 94% accuracy in diagnosing prostate cancer, compared to 88% accuracy when analyzing the imaging data alone ^[4].

2.2. Information from Prior Studies

Screening mammography is intended to be performed multiple times throughout a patient’s lifetime, with at least one comparison mammogram often available for interpretation in clinical practice. In fact, one study observed that over 90% of their exams included a comparison film ^[6]. Most radiologists interpret mammograms within the context of comparison with a prior, allowing them to discern static lesions versus evolving changes, and increasing their confidence in the assessment. However, some evidence suggests that prior imaging may increase the callback rate due to physiologic and normal positional changes between mammograms that may appear suspicious simply due to their difference in appearance between studies. For example, a study from Yankasas et al. ^[6] demonstrated that when comparison mammograms were available and had an apparent change, there was an increase in the false-positive interpretation rate. However, others have shown that prior imaging allows radiologists to identify more subtle changes that may represent the early stages of cancer development, resulting in increased sensitivity. A study by Hayward et al. showed a reduced recall and increased cancer detection rate when multiple prior mammograms were available ^[7], and a study by Burnside et al. showed a decrease in false positives, with the detection of cancer occurring at an earlier stage ^[8].

Much like prior imaging may enhance the capabilities of human radiologists, incorporating prior imaging has been proposed to improve the performance of AI. Several approaches to this challenge exist, with a newer type of AI network demonstrating a high level of success in comparing two medical images to determine similarities or differences between them ^[9]. This network architecture is novel for its use of two parallel and identical networks to analyze the features of comparison images separately (for example, the current image and the prior image), before adding an additional component that compares between the two. Investigators have recently employed such networks in medical tasks, such as determining osteoarthritis progression on sequential knee radiographs ^[9] and for retinopathy grading ^[10]. Within breast imaging, several recent articles, including a review by Loizidou et al., and a few commercial products have emerged that specifically utilize temporal changes in medical images for better diagnosis ^[11]. For instance, Bai et al. compared several different types of AI networks for cancer classification, finding that the best performance was achieved with a model capable of image comparisons ^[12]. Using a different novel technique based on image subtraction, a study by Loizidou et al. demonstrated 99% accuracy in distinguishing masses from normal tissue in their dataset ^[11].

In addition to comparing across time points, radiologists also consider similarities and differences between the right and left breast when analyzing a study. Organ symmetry has been effectively employed in image analysis of other body regions, such as the mastoid air cells for detection of mastoiditis ^[13]. The integration of breast symmetry information in the academic literature is still evolving, however. A study by Shimokawa et al. demonstrated the early promise of this technique, where a network comparing the symmetry of bilateral breasts improved cancer detection compared to more traditional neural network approaches ^[14].

2.3. Challenges to Information Integration: Interoperability and Data Security

The incorporation of patient clinical information with imaging and ability to compare with prior studies are not without challenges. A significant obstacle lies in the lack of interoperability among healthcare data systems. For example, to consider a patient’s cancer history an algorithm may need permission to access the medical record, which is likely a separate application from a different company. Inconsistencies also exist in terminologies, measurement units, and data entry formats, or may need to be derived from natural language in provider notes. The scope of information an algorithm has access to also raises concerns for data privacy and security, as each new integration offers another potential source for data breaches or hacks. A healthcare data breach costs upwards of USD 6.5 million on average, making security an ethical and financial concern ^[15]. Addressing these challenges has the potential to improve patient care and diagnosis but will necessitate a multifaceted approach that includes standardized data formats and enhanced data management protocols.

3. Reducing the Clinical Workload, and the Importance of Bringing Patients into the Discussion

The uses of AI in breast imaging also extend beyond diagnostic applications to address challenges in the clinical workflow. Although not currently implemented in clinical use, several studies have proposed AI for workload reduction through two primary mechanisms that involve changing the way cases are presented to the radiologist: (1) removing negative/normal cases from the worklist and (2) prioritizing abnormal cases that require prompt attention. For instance, an algorithm could analyze all screening mammography exams and automatically report those considered confidently normal, allowing radiologists to concentrate on more nuanced cases or complex diagnoses. In mammography, several groups have conducted retrospective simulations to assess workload reduction potential. Early studies demonstrated moderate benefits, with work by Rodreguez-Ruiz et al. showing a 17% reduction in studies while missing 1% of true positives, and work by Yala et al. showed no change in radiologist specificity and sensitivity while eliminating 19.3% of exams ^[16]. As algorithms have continued to improve, more recent research involving larger populations has shown the potential for more significant benefits. For example, a study by Shoshan et al. reported a workload reduction of 40%, with noninferior sensitivity and decreased recall rate ^[17]. Similarly, a large European study by Sharma et al. revealed a reading time reduction of nearly 45% while also reducing recalls ^[18]. By eliminating a portion of the normal exams, workload reduction algorithms have the potential to help address radiologist shortages and the potential to reduce the time patients spend waiting for anxiety-provoking results.

Despite the perceived benefit to radiologist workflows, the decision to rely on an algorithm for risk assessment or final patient diagnosis must ultimately consider several factors. Foremost among these is the level of patient comfort with reduced or absent input from a physician. The understanding of patient opinions regarding medical AI remains limited, and the landscape of AI and its integration into daily life continue to evolve. A recent meta-review by Young et al. highlighted the paucity and variability of existing studies, finding that studies examining patient attitudes were often of varying quality and were subject to selection bias ^[19]. Despite their overall conclusion that patients generally had a positive view of AI tools, they observed that many still prefer an element of human supervision. A breast-imaging-specific study by Lennox-Chugani ^[20] from England found an overall positive patient perception of using AI to read screening mammograms, with 50% of patients of screening age feeling positively and, interestingly, a slightly lower level of trust among women under screening age at 45%. The feasibility of a fully AI-based diagnosis, even for normal screening exams, should prioritize patient-centeredness and foster harmony among patients, radiologists, and administrators. Moreover, this group of stakeholders should be aware that opinions about AI-based diagnosis may change over time and maintain adaptability to ensure the best possible patient safety and comfort.

4. Reducing the Administrative Workload

4.1. Automating Mammography Quality Standards Act and General Quality Assurance

The potential of AI extends further into areas traditionally viewed as administrative, such as quality assurance tasks. Enacted in 1992, the Mammography Quality Standards Act (MQSA) mandates facilities to audit medical outcomes with the objective of establishing uniform quality standards in screening mammography. To maintain certification, MQSA requires the periodic submission of sample images to demonstrate service quality, which facilitates a comparison between an individual clinic’s performance and national-level statistics. The implementation of this audit has been shown to improve screening and diagnostic quality, with the short-interval performance feedback proving beneficial to both radiologists and technicians ^[21]. However, the MQSA audits necessitate the collection of a significant amount of information and the selection of appropriate images may be very time consuming, resulting in an increased administrative workload for radiologists and technologists. AI solutions have been proposed to assist with identifying images for MQSA submission, thereby reducing the administrative burden on breast imaging clinics. While there is a paucity of academic literature on this topic, several products have already entered this domain to help radiologists to streamline the process of acquiring data for EQUIP and other administrative workflows. It is noteworthy to mention that such applications for AI are relatively new, and only a limited number of products are commercially available to help radiology offices comply with MQSA and EQUIP regulations. As such, there are limited data on the actual performance and reliability of such products in clinics.

Beyond MQSA requirements, ensuring the quality of mammographic images is critical for accurate diagnoses. Poor-quality mammograms can have a significant impact on patient care, increasing radiation dose and delaying cancer detection ^[22]. While breast imaging phantoms can help guarantee the technical quality of mammographic equipment, human factors play a role in assuring the quality of the final image. Breast positioning has been identified as a leading cause of poor-quality images, with positioning errors contributing to misdiagnosis or missed detection of cancers ^[23]^[24]. ML-based solutions have been proposed to perform real-time quality control on acquired images, reducing the need for a technical repeat before the patient leaves the screening appointment. These types of solutions could track the performance of individual technologists, identifying areas for performance improvements, such as adequate compression or positioning, and allowing for continued and prompt feedback.

Although limited academic research has been conducted in this area, several companies have commercial products designed to automate quality-control tasks. For instance, Volpara Health reports a product that assesses factors such as position and compression on screening mammograms ^[25], CureMetrix, Inc has developed a product that aims to analyze a longitudinal set of studies from an institution to provide individualized quality statistics ^[26], and Densitas Health has a product for evaluating mammograms to flag poor-quality images and benchmark performance ^[27]. It is anticipated that additional products will continue to be developed that automatically ensure and maintain high-quality imaging, ultimately enhancing patient care and reducing the risk of misdiagnoses or delayed cancer detection.

4.2. Clinical Scheduling and Protocoling

Extending beyond the realm of imaging, AI’s efficiency can also be employed to streamline other areas of clinical operation. Several commercial AI applications have been developed to assist with clinical scheduling, aiming to optimize equipment and staff utilization. For example, algorithms have been developed to assess patient risk factors and predict the amount of time needed for a surgical case, thus enabling more accurate scheduling and utilization of operating rooms ^[28]. In radiology, scheduling applications may also focus on protocoling studies, such as MRIs, that have the potential to be performed differently based on the clinical scenario. Protocolling studies are essential but time-consuming and are estimated to take up to 6% of a radiologist’s time ^[29]^[30]. A study by Trivedi et al. used natural language processing (NLP), a type of machine learning, to assign contrast to musculoskeletal MRI protocols, and a study by Brown and Marotta used NLP to protocol brain MRIs, with both showing 83% accuracy ^[31]^[32]. A broader simulation study by Kalra et al. used NLP to protocol general CT and MRI studies, finding that nearly 70% of case protocols could be successfully automated ^[29]. As more advanced language models (such as large language models including ChatGPT) become available for study, complex clinical questions may be even better understood and translated by AI into their relevant imaging parameters. Reducing this workload could not only save time but potentially minimize interruptions and help ensure that a consistent and appropriate modality is being utilized to address the clinical question.

Appointment scheduling algorithms have also been developed, often based on the likelihood of a patient missing an appointment. This allows the schedulers to “overbook” or schedule multiple patients for the same appointment times with the expectation of attrition or cancelations. However, studies in other medical fields have found that AI scheduling may inadvertently contribute to healthcare disparities. For example, in one study, socioeconomic biases were inherent in the algorithm training data, resulting in some patients having inappropriately longer wait times ^[33]. ML applications are ultimately a reflection of their training data, and understanding that ML applications may perpetuate human biases is important to ensure that vulnerable populations are receiving equitable care. Multiple studies have underscored the imperative for a cautious and ethical approach towards creating AI models, with a clear focus on enhancing data diversity to ensure equitable health outcomes for all populations. For example, Mema and McGinty discussed the potential for AI to reduce health disparities in breast cancer care and highlight the need for active clinician engagement to reduce biases, Agarwal et al. discussed bias sources and proposed mitigation strategies in AI for healthcare, and Halamka et al. discussed discrimination relating to surgical care and proposed ways AI may help ^[34]^[35]^[36].

References

McKinney, S.M.; Sieniek, M.; Godbole, V.; Godwin, J.; Antropova, N.; Ashrafian, H.; Back, T.; Chesus, M.; Corrado, G.S.; Darzi, A.; et al. International Evaluation of an AI System for Breast Cancer Screening. Nature 2020, 577, 89–94.
Mohsen, F.; Ali, H.; El Hajj, N.; Shah, Z. Artificial Intelligence-Based Methods for Fusion of Electronic Health Records and Imaging Data. Sci. Rep. 2022, 12, 17981.
Kharazmi, P.; Kalia, S.; Lui, H.; Wang, Z.J.; Lee, T.K. A Feature Fusion System for Basal Cell Carcinoma Detection through Data-Driven Feature Learning and Patient Profile. Ski. Res. Technol. 2018, 24, 256–264.
Reda, I.; Khalil, A.; Elmogy, M.; Abou El-Fetouh, A.; Shalaby, A.; Abou El-Ghar, M.; Elmaghraby, A.; Ghazal, M.; El-Baz, A. Deep Learning Role in Early Diagnosis of Prostate Cancer. Technol. Cancer Res. Treat. 2018, 17, 1533034618775530.
Huang, S.-C.; Pareek, A.; Seyyedi, S.; Banerjee, I.; Lungren, M.P. Fusion of Medical Imaging and Electronic Health Records Using Deep Learning: A Systematic Review and Implementation Guidelines. NPJ Digit. Med. 2020, 3, 136.
Yankaskas, B.C.; May, R.C.; Matuszewski, J.; Bowling, J.M.; Jarman, M.P.; Schroeder, B.F. Effect of Observing Change from Comparison Mammograms on Performance of Screening Mammography in a Large Community-Based Population. Radiology 2011, 261, 762–770.
Hayward, J.H.; Ray, K.M.; Wisner, D.J.; Kornak, J.; Lin, W.; Joe, B.N.; Sickles, E.A. Improving Screening Mammography Outcomes Through Comparison With Multiple Prior Mammograms. AJR Am. J. Roentgenol. 2016, 207, 918–924.
Burnside, E.S.; Sickles, E.A.; Sohlich, R.E.; Dee, K.E. Differential Value of Comparison with Previous Examinations in Diagnostic Versus Screening Mammography. Am. J. Roentgenol. 2002, 179, 1173–1177.
Li, M.D.; Chang, K.; Bearce, B.; Chang, C.Y.; Huang, A.J.; Campbell, J.P.; Brown, J.M.; Singh, P.; Hoebel, K.V.; Erdoğmuş, D.; et al. Siamese Neural Networks for Continuous Disease Severity Evaluation and Change Detection in Medical Imaging. NPJ Digit. Med. 2020, 3, 48.
Nirthika, R.; Manivannan, S.; Ramanan, A. Siamese Network Based Fine Grained Classification for Diabetic Retinopathy Grading. Biomed. Signal. Process. Control 2022, 78, 103874.
Loizidou, K.; Skouroumouni, G.; Nikolaou, C.; Pitris, C. Automatic Breast Mass Segmentation and Classification Using Subtraction of Temporally Sequential Digital Mammograms. IEEE J. Transl. Eng. Health Med. 2022, 10, 1801111.
Bai, J.; Jin, A.; Wang, T.; Yang, C.; Nabavi, S. Feature Fusion Siamese Network for Breast Cancer Detection Comparing Current and Prior Mammograms. Med. Phys. 2022, 49, 3654–3669.
Lee, J.; Kang, B.J.; Kim, S.H.; Park, G.E. Evaluation of Computer-Aided Detection (CAD) in Screening Automated Breast Ultrasound Based on Characteristics of CAD Marks and False-Positive Marks. Diagnostics 2022, 12, 583.
Shimokawa, D.; Takahashi, K.; Kurosawa, D.; Takaya, E.; Oba, K.; Yagishita, K.; Fukuda, T.; Tsunoda, H.; Ueda, T. Deep Learning Model for Breast Cancer Diagnosis Based on Bilateral Asymmetrical Detection (BilAD) in Digital Breast Tomosynthesis Images. Radiol. Phys. Technol. 2023, 16, 20–27.
Seh, A.H.; Zarour, M.; Alenezi, M.; Sarkar, A.K.; Agrawal, A.; Kumar, R.; Ahmad Khan, R. Healthcare Data Breaches: Insights and Implications. Healthcare 2020, 8, 133.
Yala, A.; Schuster, T.; Miles, R.; Barzilay, R.; Lehman, C. A Deep Learning Model to Triage Screening Mammograms: A Simulation Study. Radiology 2019, 293, 38–46.
Shoshan, Y.; Bakalo, R.; Gilboa-Solomon, F.; Ratner, V.; Barkan, E.; Ozery-Flato, M.; Amit, M.; Khapun, D.; Ambinder, E.B.; Oluyemi, E.T.; et al. Artificial Intelligence for Reducing Workload in Breast Cancer Screening with Digital Breast Tomosynthesis. Radiology 2022, 303, 69–77.
Sharma, N.; Ng, A.Y.; James, J.J.; Khara, G.; Ambrozay, E.; Austin, C.C.; Forrai, G.; Fox, G.; Glocker, B.; Heindl, A.; et al. Retrospective Large-Scale Evaluation of an AI System as an Independent Reader for Double Reading in Breast Cancer Screening. medRxiv 2022.
Young, A.T.; Amara, D.; Bhattacharya, A.; Wei, M.L. Patient and General Public Attitudes towards Clinical Artificial Intelligence: A Mixed Methods Systematic Review. Lancet Digit. Health 2021, 3, e599–e611.
Lennox-Chhugani, N.; Chen, Y.; Pearson, V.; Trzcinski, B.; James, J. Women’s Attitudes to the Use of AI Image Readers: A Case Study from a National Breast Screening Programme. BMJ Health Care Inf. 2021, 28, e100293.
Hussain, S.; Omar, A.; Shah, B.A. The Breast Imaging Medical Audit: What the Radiologist Needs to Know. Contemp. Diagn. Radiol. 2021, 44, 1–5.
Huppe, A.I.; Overman, K.L.; Gatewood, J.B.; Hill, J.D.; Miller, L.C.; Inciardi, M.F. Mammography Positioning Standards in the Digital Era: Is the Status Quo Acceptable? AJR Am. J. Roentgenol. 2017, 209, 1419–1425.
Sweeney, R.-J.I.; Lewis, S.J.; Hogg, P.; McEntee, M.F. A Review of Mammographic Positioning Image Quality Criteria for the Craniocaudal Projection. Br. J. Radiol. 2018, 91, 20170611.
Taplin, S.H.; Rutter, C.M.; Finder, C.; Mandelson, M.T.; Houn, F.; White, E. Screening Mammography: Clinical Image Quality and the Risk of Interval Breast Cancer. AJR Am. J. Roentgenol. 2002, 178, 797–803.
VolparaEnterprise Will Help Compliance with FDA’s EQUIP. Available online: https://www.volparahealth.com/news/volparaenterprise-will-help-compliance-with-fdas-equip/ (accessed on 29 April 2023).
CureMetrix CmTriage. Available online: https://curemetrix.com/cm-triage-2/ (accessed on 14 February 2022).
MQSA Compliance|Mammography Quality Assurance. Densitas. Available online: https://densitashealth.com/solutions/quality/ (accessed on 29 April 2023).
Abbas, A.; Mosseri, J.; Lex, J.R.; Toor, J.; Ravi, B.; Khalil, E.B.; Whyne, C. Machine Learning Using Preoperative Patient Factors Can Predict Duration of Surgery and Length of Stay for Total Knee Arthroplasty. Int. J. Med. Inform. 2022, 158, 104670.
Kalra, A.; Chakraborty, A.; Fine, B.; Reicher, J. Machine Learning for Automation of Radiology Protocols for Quality and Efficiency Improvement. J. Am. Coll. Radiol. 2020, 17, 1149–1158.
Lau, W.; Aaltonen, L.; Gunn, M.; Yetisgen, M. Automatic Assignment of Radiology Examination Protocols Using Pre-Trained Language Models with Knowledge Distillation. AMIA Annu. Symp. Proc. 2021, 2021, 668.
Trivedi, H.; Mesterhazy, J.; Laguna, B.; Vu, T.; Sohn, J.H. Automatic Determination of the Need for Intravenous Contrast in Musculoskeletal MRI Examinations Using IBM Watson’s Natural Language Processing Algorithm. J. Digit. Imaging 2018, 31, 245–251.
Brown, A.D.; Marotta, T.R. A Natural Language Processing-Based Model to Automate MRI Brain Protocol Selection and Prioritization. Acad. Radiol. 2017, 24, 160–166.
Samorani, M.; Blount, L.G. Machine Learning and Medical Appointment Scheduling: Creating and Perpetuating Inequalities in Access to Health Care. Am. J. Public Health 2020, 110, 440–441.
Mema, E.; McGinty, G. The Role of Artificial Intelligence in Understanding and Addressing Disparities in Breast Cancer Outcomes. Curr. Breast Cancer Rep. 2020, 12, 168–174.
Agarwal, R.; Bjarnadottir, M.; Rhue, L.; Dugas, M.; Crowley, K.; Clark, J.; Gao, G. Addressing Algorithmic Bias and the Perpetuation of Health Inequities: An AI Bias Aware Framework. Health Policy Technol. 2023, 12, 100702.
Halamka, J.; Bydon, M.; Cerrato, P.; Bhagra, A. Addressing Racial Disparities in Surgical Care with Machine Learning. NPJ Digit. Med. 2022, 5, 152.

© Text is available under the terms and conditions of the Creative Commons Attribution (CC BY) license; additional terms may apply. By using this site, you agree to the Terms and Conditions and Privacy Policy.

Upload a video for this entry

Information

Subjects: Pathology

Contributors MDPI registered users' name will be linked to their SciProfiles pages. To register with us, please refer to https://encyclopedia.pub/register :

Tara A. Retson

Mohammad Eghtedari

View Times: 270

Update Date: 04 Jul 2023

Table of Contents

Video Upload Options

Confirm