Advancing Artificial Intelligence to Meet Breast Imaging Needs
2022; Radiological Society of North America; Volume: 303; Issue: 1 Linguagem: Inglês
10.1148/radiol.213101
ISSN1527-1315
Autores Tópico(s)Artificial Intelligence in Healthcare and Education
ResumoHomeRadiologyVol. 303, No. 1 PreviousNext Reviews and CommentaryFree AccessEditorialAdvancing Artificial Intelligence to Meet Breast Imaging NeedsLiane E. Philpotts Liane E. Philpotts Author AffiliationsFrom the Department of Radiology and Biomedical Imaging, Yale School of Medicine, 333 Cedar St, PO Box 208042, New Haven, CT 06520.Address correspondence to the author (e-mail: [email protected]).Liane E. Philpotts Published Online:Jan 18 2022https://doi.org/10.1148/radiol.213101MoreSectionsPDF ToolsImage ViewerAdd to favoritesCiteTrack CitationsPermissionsReprints ShareShare onFacebookTwitterLinked In See also the article by Shoshan et al in this issue.Liane Philpotts, MD, is professor of radiology and biomedical imaging at Yale School of Medicine. She is a fellow of the American College of Radiology and the Society of Breast Imaging. She has been a member of National Institutes of Health Study Sections, served on American College of Radiology, Society of Breast Imaging and National Comprehensive Cancer Network committees, and been an editor and long-standing reviewer for Radiology. She coauthored the first textbook on digital breast tomosynthesis.Download as PowerPointOpen in Image Viewer The psychology of mammography interpretation is an interesting topic and pertinent to understanding how humans interact with artificial intelligence (AI). Whereas there has been an explosion of AI literature in the last few years, there are relatively few studies that actually assess the real-world clinical experience of radiologists with AI (1). Most AI algorithms are unlikely to be used as stand-alone modalities but rather as complementary tools for radiologists to use concurrently. The prospective use of AI is an essential step before its widespread clinical use. Lessons learned from the rollout of computer-aided detection (CAD) technology in the early 2000s showed that actual clinical experience did not live up to expectations. Laboratory reader and observer studies of CAD suggested that the tool could dramatically help radiologists find more cancers and avoid recalls. However, experience showed that once CAD was integrated into widespread clinical use, many radiologists' recall rates actually increased, and cancer detection decreased (2,3). With high rates of false alarms and less-than-optimal cancer detection, many radiologists ignored CAD flags or, conversely, relied too much on them. Like CAD, could AI possibly have a detrimental effect on radiologist performance from such so-called automation bias by depending too heavily on the AI interpretation (4)?Digital breast tomosynthesis (DBT) technology has proven to be a benefit to screening mammography. DBT reduces unnecessary recalls while also improving cancer detection. But rather than the four two-dimensional (2D) images that comprise routine bilateral screening mammography, DBT examinations often contain hundreds of images. Generally viewed in 1-mm sections, interpretation requires scrolling through large stacks of images to find subtle cancers and differentiate areas of asymmetry from superimposed fibroglandular tissue. This is considerably more time consuming than reading 2D images and can lead to fatigue and potentially decreased acuity. With shortages of radiologists and increasing volumes of images to be interpreted, methods to improve workflow while maintaining accuracy are welcomed.Most AI for mammography has been developed with 2D mammography platforms. These will likely have the same limitations that human reading of 2D mammography has compared with DBT: Areas of dense tissue can mask cancers or create false-positive findings. Given that DBT provides a better mammogram and will likely replace 2D mammography, AI algorithms on the basis of DBT represent the direction AI development should be focused.In this issue of Radiology, Shoshan et al (5) describe an AI model for assessing DBT examinations that could potentially identify 39% of screening examinations as normal. This could substantially reduce workload if the AI system is used as a stand-alone method and those examinations are not reviewed by a radiologist. Because there are roughly five cancers per 1000 screening examinations, in general only one examination of every 200 will harbor cancer. The vast majority of screening examinations are normal. It seems logical that many of these could be safely eliminated by an AI system, alleviating a large burden of screening volume. In addition to eliminating the subset of normal examinations, the authors found the AI model could potentially decrease 25% of screening recalls. It is important to emphasize that this was a modeling study. The authors used the radiologists' original readings to compare with the AI output. This model assumes stable reader performance for the remaining 61% cases (after the 39% normal cases are filtered out). Whether radiologist performance would remain the same knowing that that AI classified those cases as not normal is questionable and an area that needs further assessment.To date, there have been few studies reporting AI for DBT. Raya-Povedano et al (6), in a cohort of almost 16 000 digital and DBT mammograms, showed a potential work reduction of up to 70% with an AI-assisted system for both digital mammography and DBT screening. Conant et al (7) reported a reader study of 24 radiologists reading 260 DBT mammograms with and without an AI system. In a cancer-enriched cohort, they obtained encouraging results of statistically significant improvement in the performance of improved cancer detection, reduced recalls, and decreased reading times. Pinto et al (8) recently reported a retrospective cancer-enriched observer study that used single-view mediolateral oblique wide-angle DBT interpreted with and without the use of AI. They showed an improvement in radiologists' sensitivity without a statistically significant change in specificity or reading time with the use of AI. Will these initial results for DBT with AI be reproduced or are they examples of the so-called laboratory effect and will not translate to clinical practices?Prospective studies are necessary to determine how the use of AI will translate to the clinical world. Like CAD, AI for DBT needs to be trialed extensively in a variety of clinical settings to understand the downstream effects. What effect will the AI triaging of cases as described in this study actually have in real-world clinical practice? What will the potential removal of the normal cases from the reading batch have on radiologists? It is likely the AI is going to flag many of the less dense mammograms as being normal, leaving the cases with more complex tissue patterns as potentially abnormal. Eliminating a proportion of normal, nondense cases is therefore going to leave the more challenging cases for the radiologist to read. While batch reading screening mammograms, breast radiologists are always relieved when encountering a nondense case. Images in fatty breasts are considerably easier to interpret than images in dense breasts, and they provide something of a breather. Without them, reading mammograms of only dense breasts is akin to doing high-intensity sprinting intervals without the rests in-between; it would quickly become exhausting. Furthermore, radiologist awareness that AI flagged the remaining cases as potentially abnormal may actually increase the time involved to review those cases and lead to more recalls. This may negate the benefit of the reduction of the workload from the cases flagged as normal.A recent study of clinical use of an AI system for 2D mammography, however, showed promise for improved screening experience by using AI (9). By using both an AI-triage platform and a scoring system, the AI system sorted cases by the level of urgency. Compared with traditional CAD, the AI CAD had a 71% reduction in flags per examination. The three radiologists involved perceived greater ease of reading the batched cases, particularly those triaged as not suspicious for cancer. Additional larger multisite data will be needed to ascertain whether the results from this small single practice trial are reproduced on a broader scale.Of note, in the study by Shoshan et al, the sensitivity of the AI model, although noninferior, was slightly less than that of the radiologists. A review of the missed cancer cases revealed that most were truly occult. Nonetheless, four true cancer cases detected by the radiologist were missed by the AI system. Missing any cancer case is a cause for concern. A recent meta-analysis of AI algorithms found 34 of 36 AI systems evaluated were actually less accurate than a single radiologist, and all were less accurate than double reading by two radiologists (10). If radiologists are aware that the system is not 100% correct, will they trust the AI information? Several studies have shown that women are willing to accept false-positive findings so as to not have a breast cancer missed. Will patients accept that their images may only be interpreted by a computer? It is known that breast cancers can manifest with subtle findings that are often only detected because of change from previous mammograms. These nuanced perceptions are something only a radiologist can assess. For these reasons, is it reasonable to think that we can use an AI system to eliminate some cases from human viewing? In reality, radiologists will likely still have to fully evaluate most DBT scans to detect those few subtle cancers, despite what the AI tells us.AI is already pervasive in many aspects of modern life, and radiology is no exception. As radiologists, our relationship with AI will certainly evolve over time. Radiologists need to embrace and help mold the technology that is being developed. We can only hope continued efforts such as the study by Shoshan et al (5) make progress toward improving both radiologists' workflow and confidence in interpretation as well as patient outcomes.Disclosures of Conflicts of Interest: L.E.P. No relevant relationships.References1. Lui X, Faes L, Kale AU, et al. A comparison of deep learning performance against health-care professionals in detecting diseases from medical imaging: a systematic review and meta-analysis. Lancet Digit Health 2019;1(6):e271–e297 [Published correction appears in Lancet Digit Health 2019;1(7):e334.]. Crossref, Medline, Google Scholar2. Philpotts LE. Can computer-aided detection be detrimental to mammographic interpretation? Radiology 2009;253(1):17–22. Link, Google Scholar3. Lehman CD, Wellman RD, Buist DSM, et al. Diagnostic accuracy of digital screening mammography with and without computer-aided detection. JAMA Intern Med 2015;175(11):1828–1837. Crossref, Medline, Google Scholar4. Le EPV, Wang Y, Huang Y, Hickman S, Gilbert FJ. Artificial intelligence in breast imaging. Clin Radiol 2019;74(5):357–366. Crossref, Medline, Google Scholar5. Shoshan Y, Bakalo R, Gilboa-Solomon F, et al. An Artificial Intelligence Model for Reducing Workload in Breast Cancer Screening with Digital Breast Tomosynthesis. Radiology 2022;303(1):69–77. Google Scholar6. Raya-Povedano JL, Romero-Martín S, Elías-Cabot E, Gubern-Mérida A, Rodríguez-Ruiz A, Álvarez-Benito M. AI-based Strategies to Reduce Workload in Breast Cancer Screening with Mammography and Tomosynthesis: A Retrospective Evaluation. Radiology 2021;300(1):57–65. Link, Google Scholar7. Conant EF, Toledano AY, Periaswamy S, et al. Improving accuracy and efficiency with concurrent use of artificial intelligence for digital breast tomosynthesis. Radiol Artif Intell 2019;1(4):e180096. Link, Google Scholar8. Pinto MC, Rodriguez-Ruiz A, Pedersen K, et al. Impact of Artificial Intelligence Decision Support Using Deep Learning on Breast Cancer Screening Interpretation with Single-View Wide-Angle Digital Breast Tomosynthesis. Radiology 2021;300(3):529–536. Link, Google Scholar9. Tartar M, Le L, Watanabe AT, Enomoto AJ. Artificial intelligence support for mammography: in-practice clinical experience. J Am Coll Radiol 2021;18(11):1510–1513. Crossref, Medline, Google Scholar10. Freeman K, Geppert J, Stinton C, et al. Use of artificial intelligence for image analysis in breast cancer screening programmes: systematic review of test accuracy. BMJ 2021;374n1872. Crossref, Medline, Google ScholarArticle HistoryReceived: Dec 6 2021Revision requested: Dec 8 2021Revision received: Dec 8 2021Accepted: Dec 10 2021Published online: Jan 18 2022Published in print: Apr 2022 FiguresReferencesRelatedDetailsCited ByIntelligence artificielle : Place dans le dépistage du cancer du sein en FranceIsabelleThomassin-Naggara, LucCeugnart, AnneTardivon, LaurentVerzaux, CorinneBalleyguier, PatriceTaourel, BrigitteSeradour2022 | Bulletin du Cancer, Vol. 109, No. 7-8The Impact of Dense Breasts on the Stage of Breast Cancer at Diagnosis: A Review and Options for Supplemental ScreeningPaula B.Gordon2022 | Current Oncology, Vol. 29, No. 5Accompanying This ArticleArtificial Intelligence for Reducing Workload in Breast Cancer Screening with Digital Breast TomosynthesisJan 18 2022RadiologyRecommended Articles Implementation of Synthesized Two-dimensional Mammography in a Population-based Digital Breast Tomosynthesis Screening ProgramRadiology2016Volume: 281Issue: 3pp. 730-736Artificial Intelligence for Reducing Workload in Breast Cancer Screening with Digital Breast TomosynthesisRadiology2022Volume: 303Issue: 1pp. 69-77Strengths and Weaknesses of Synthetic Mammography in ScreeningRadioGraphics2017Volume: 37Issue: 7pp. 1913-1927The Coming of Age of Breast Tomosynthesis in ScreeningRadiology2019Volume: 291Issue: 1pp. 31-33Does Reader Performance with Digital Breast Tomosynthesis Vary according to Experience with Two-dimensional Mammography?Radiology2017Volume: 283Issue: 2pp. 371-380See More RSNA Education Exhibits Implementation of Synthetic Mammography: Benefits, Drawbacks, and PitfallsDigital Posters2022Artificial Intelligence in Breast Imaging: Past, Present, and FutureDigital Posters2020New Horizons: Artificial Intelligence (AI) In Digital Breast Tomosynthesis (DBT)Digital Posters2021 RSNA Case Collection Breast edemaRSNA Case Collection2021Invasive ductal carcinoma as developing asymmetryRSNA Case Collection2021Asymmetric lactational change RSNA Case Collection2021 Vol. 303, No. 1 Metrics Altmetric Score PDF download
Referência(s)