User login
A new way of using artificial intelligence (AI) can predict breast cancer 5 years in advance with impressive accuracy — and unlike previous AI models, we know how this one works.
The new AI system, called AsymMirai, simplifies previous models by solely comparing differences between right and left breasts to predict risk. It could potentially save lives, prevent unnecessary testing, and save the healthcare system money, its creators say.
“With traditional AI, you ask it a question and it spits out an answer, but no one really knows how it makes its decisions. It’s a black box,” said Jon Donnelly, a PhD student in the department of computer science at Duke University, Durham, North Carolina, and first author on a new paper in Radiology describing the model.
“With our approach, people know how the algorithm comes up with its output so they can fact-check it and trust it,” he said.
One in eight women will develop invasive breast cancer, and 1 in 39 will die from it. Mammograms miss about 20% of breast cancers. (The shortcomings of genetic screening and mammograms received extra attention recently when actress Olivia Munn disclosed that she’d been treated for an aggressive form of breast cancer despite a normal mammogram and a negative genetic test.)
The model could help doctors bring the often-abstract idea of AI to the bedside in a meaningful way, said radiologist Vivianne Freitas, MD, assistant professor of medical imaging at the University of Toronto.
“This marks a new chapter in the field of AI,” said Dr. Freitas, who authored an editorial lauding the new paper. “It makes AI more tangible and understandable, thereby improving its potential for acceptance.”
AI as a Second Set of Eyes
Mr. Donnelly described AsymMirai as a simpler, more transparent, and easier-to-use version of Mirai, a breakthrough AI model which made headlines in 2021 with its promise to determine with unprecedented accuracy whether a patient is likely to get breast cancer within the next 5 years.
Mirai identified up to twice as many future cancer diagnoses as the conventional risk calculator Tyrer-Cuzick. It also maintained accuracy across a diverse set of patients — a notable plus for two fields (AI and healthcare) notorious for delivering poorer results for minorities.
Tyrer-Cuzick and other lower-tech risk calculators use personal and family history to statistically calculate risk. Mirai, on the other hand, analyzes countless bits of raw data embedded in a mammogram to decipher patterns a radiologist’s eyes may not catch. Four images, including two angles from each breast, are fed into the model, which produces a score between 0 and 1 to indicate the person’s risk of getting breast cancer in 1, 3, or 5 years.
But even Mirai’s creators have conceded they didn’t know exactly how it arrives at that score — a fact that has fueled hesitancy among clinicians.
Study coauthor Fides Schwartz, MD, a radiologist at Brigham and Women’s Hospital, Boston, said researchers were able to crack the code on Mirai’s “black box,” finding that its scores were largely determined by assessing subtle differences between right breast tissue and left breast tissue.
Knowing this, the research team simplified the model to predict risk based solely on “local bilateral dissimilarity.” AsymMirai was born.
The team then used AsymMirai to look back at > 200,000 mammograms from nearly 82,000 patients. They found it worked nearly as well as its predecessor, assigning a higher risk to those who would go on to develop cancer 66% of the time (vs Mirai’s 71%). In patients where it noticed the same asymmetry multiple years in a row it worked even better, with an 88% chance of giving people who would develop cancer later a higher score than those who would not.
“We found that we can, with surprisingly high accuracy, predict whether a woman will develop cancer in the next 1-5 years based solely on localized differences between her left and right breast tissue,” said Mr. Donnelly.
Dr. Schwartz imagines a day when radiologists could use the model to help develop personalized screening strategies for patients. Doctors might advise those with higher scores to get screened more often than guidelines suggest, supplement mammograms with an MRI , and keep a close watch on trouble spots identified by AI.
“For people with really low risk, on the other hand, maybe we can save them an annual exam that’s not super pleasant and might not be necessary,” said Dr. Schwartz.
Cautious Optimism
Robert Smith, PhD, senior vice president of early cancer detection science at the American Cancer Society, noted that AI has been used for decades to try to reduce radiologists’ workload and improve diagnoses.
“But AI just never really lived up to its fullest potential,” Dr. Smith said, “quite often because it was being used as a crutch by inexperienced radiologists who, instead of interpreting the mammogram and then seeing what AI had to say ended up letting AI do most of the work which, frankly, just wasn’t that accurate.”
He’s hopeful that newer, more sophisticated iterations of AI medical imaging platforms (roughly 18-20 models are in development) can ultimately save women’s lives, particularly in areas where radiologists are in short supply.
But he believes it will be a long time before doctors, or their patients, are willing to risk postponing a mammogram based on an algorithm.
A version of this article appeared on Medscape.com.
A new way of using artificial intelligence (AI) can predict breast cancer 5 years in advance with impressive accuracy — and unlike previous AI models, we know how this one works.
The new AI system, called AsymMirai, simplifies previous models by solely comparing differences between right and left breasts to predict risk. It could potentially save lives, prevent unnecessary testing, and save the healthcare system money, its creators say.
“With traditional AI, you ask it a question and it spits out an answer, but no one really knows how it makes its decisions. It’s a black box,” said Jon Donnelly, a PhD student in the department of computer science at Duke University, Durham, North Carolina, and first author on a new paper in Radiology describing the model.
“With our approach, people know how the algorithm comes up with its output so they can fact-check it and trust it,” he said.
One in eight women will develop invasive breast cancer, and 1 in 39 will die from it. Mammograms miss about 20% of breast cancers. (The shortcomings of genetic screening and mammograms received extra attention recently when actress Olivia Munn disclosed that she’d been treated for an aggressive form of breast cancer despite a normal mammogram and a negative genetic test.)
The model could help doctors bring the often-abstract idea of AI to the bedside in a meaningful way, said radiologist Vivianne Freitas, MD, assistant professor of medical imaging at the University of Toronto.
“This marks a new chapter in the field of AI,” said Dr. Freitas, who authored an editorial lauding the new paper. “It makes AI more tangible and understandable, thereby improving its potential for acceptance.”
AI as a Second Set of Eyes
Mr. Donnelly described AsymMirai as a simpler, more transparent, and easier-to-use version of Mirai, a breakthrough AI model which made headlines in 2021 with its promise to determine with unprecedented accuracy whether a patient is likely to get breast cancer within the next 5 years.
Mirai identified up to twice as many future cancer diagnoses as the conventional risk calculator Tyrer-Cuzick. It also maintained accuracy across a diverse set of patients — a notable plus for two fields (AI and healthcare) notorious for delivering poorer results for minorities.
Tyrer-Cuzick and other lower-tech risk calculators use personal and family history to statistically calculate risk. Mirai, on the other hand, analyzes countless bits of raw data embedded in a mammogram to decipher patterns a radiologist’s eyes may not catch. Four images, including two angles from each breast, are fed into the model, which produces a score between 0 and 1 to indicate the person’s risk of getting breast cancer in 1, 3, or 5 years.
But even Mirai’s creators have conceded they didn’t know exactly how it arrives at that score — a fact that has fueled hesitancy among clinicians.
Study coauthor Fides Schwartz, MD, a radiologist at Brigham and Women’s Hospital, Boston, said researchers were able to crack the code on Mirai’s “black box,” finding that its scores were largely determined by assessing subtle differences between right breast tissue and left breast tissue.
Knowing this, the research team simplified the model to predict risk based solely on “local bilateral dissimilarity.” AsymMirai was born.
The team then used AsymMirai to look back at > 200,000 mammograms from nearly 82,000 patients. They found it worked nearly as well as its predecessor, assigning a higher risk to those who would go on to develop cancer 66% of the time (vs Mirai’s 71%). In patients where it noticed the same asymmetry multiple years in a row it worked even better, with an 88% chance of giving people who would develop cancer later a higher score than those who would not.
“We found that we can, with surprisingly high accuracy, predict whether a woman will develop cancer in the next 1-5 years based solely on localized differences between her left and right breast tissue,” said Mr. Donnelly.
Dr. Schwartz imagines a day when radiologists could use the model to help develop personalized screening strategies for patients. Doctors might advise those with higher scores to get screened more often than guidelines suggest, supplement mammograms with an MRI , and keep a close watch on trouble spots identified by AI.
“For people with really low risk, on the other hand, maybe we can save them an annual exam that’s not super pleasant and might not be necessary,” said Dr. Schwartz.
Cautious Optimism
Robert Smith, PhD, senior vice president of early cancer detection science at the American Cancer Society, noted that AI has been used for decades to try to reduce radiologists’ workload and improve diagnoses.
“But AI just never really lived up to its fullest potential,” Dr. Smith said, “quite often because it was being used as a crutch by inexperienced radiologists who, instead of interpreting the mammogram and then seeing what AI had to say ended up letting AI do most of the work which, frankly, just wasn’t that accurate.”
He’s hopeful that newer, more sophisticated iterations of AI medical imaging platforms (roughly 18-20 models are in development) can ultimately save women’s lives, particularly in areas where radiologists are in short supply.
But he believes it will be a long time before doctors, or their patients, are willing to risk postponing a mammogram based on an algorithm.
A version of this article appeared on Medscape.com.
A new way of using artificial intelligence (AI) can predict breast cancer 5 years in advance with impressive accuracy — and unlike previous AI models, we know how this one works.
The new AI system, called AsymMirai, simplifies previous models by solely comparing differences between right and left breasts to predict risk. It could potentially save lives, prevent unnecessary testing, and save the healthcare system money, its creators say.
“With traditional AI, you ask it a question and it spits out an answer, but no one really knows how it makes its decisions. It’s a black box,” said Jon Donnelly, a PhD student in the department of computer science at Duke University, Durham, North Carolina, and first author on a new paper in Radiology describing the model.
“With our approach, people know how the algorithm comes up with its output so they can fact-check it and trust it,” he said.
One in eight women will develop invasive breast cancer, and 1 in 39 will die from it. Mammograms miss about 20% of breast cancers. (The shortcomings of genetic screening and mammograms received extra attention recently when actress Olivia Munn disclosed that she’d been treated for an aggressive form of breast cancer despite a normal mammogram and a negative genetic test.)
The model could help doctors bring the often-abstract idea of AI to the bedside in a meaningful way, said radiologist Vivianne Freitas, MD, assistant professor of medical imaging at the University of Toronto.
“This marks a new chapter in the field of AI,” said Dr. Freitas, who authored an editorial lauding the new paper. “It makes AI more tangible and understandable, thereby improving its potential for acceptance.”
AI as a Second Set of Eyes
Mr. Donnelly described AsymMirai as a simpler, more transparent, and easier-to-use version of Mirai, a breakthrough AI model which made headlines in 2021 with its promise to determine with unprecedented accuracy whether a patient is likely to get breast cancer within the next 5 years.
Mirai identified up to twice as many future cancer diagnoses as the conventional risk calculator Tyrer-Cuzick. It also maintained accuracy across a diverse set of patients — a notable plus for two fields (AI and healthcare) notorious for delivering poorer results for minorities.
Tyrer-Cuzick and other lower-tech risk calculators use personal and family history to statistically calculate risk. Mirai, on the other hand, analyzes countless bits of raw data embedded in a mammogram to decipher patterns a radiologist’s eyes may not catch. Four images, including two angles from each breast, are fed into the model, which produces a score between 0 and 1 to indicate the person’s risk of getting breast cancer in 1, 3, or 5 years.
But even Mirai’s creators have conceded they didn’t know exactly how it arrives at that score — a fact that has fueled hesitancy among clinicians.
Study coauthor Fides Schwartz, MD, a radiologist at Brigham and Women’s Hospital, Boston, said researchers were able to crack the code on Mirai’s “black box,” finding that its scores were largely determined by assessing subtle differences between right breast tissue and left breast tissue.
Knowing this, the research team simplified the model to predict risk based solely on “local bilateral dissimilarity.” AsymMirai was born.
The team then used AsymMirai to look back at > 200,000 mammograms from nearly 82,000 patients. They found it worked nearly as well as its predecessor, assigning a higher risk to those who would go on to develop cancer 66% of the time (vs Mirai’s 71%). In patients where it noticed the same asymmetry multiple years in a row it worked even better, with an 88% chance of giving people who would develop cancer later a higher score than those who would not.
“We found that we can, with surprisingly high accuracy, predict whether a woman will develop cancer in the next 1-5 years based solely on localized differences between her left and right breast tissue,” said Mr. Donnelly.
Dr. Schwartz imagines a day when radiologists could use the model to help develop personalized screening strategies for patients. Doctors might advise those with higher scores to get screened more often than guidelines suggest, supplement mammograms with an MRI , and keep a close watch on trouble spots identified by AI.
“For people with really low risk, on the other hand, maybe we can save them an annual exam that’s not super pleasant and might not be necessary,” said Dr. Schwartz.
Cautious Optimism
Robert Smith, PhD, senior vice president of early cancer detection science at the American Cancer Society, noted that AI has been used for decades to try to reduce radiologists’ workload and improve diagnoses.
“But AI just never really lived up to its fullest potential,” Dr. Smith said, “quite often because it was being used as a crutch by inexperienced radiologists who, instead of interpreting the mammogram and then seeing what AI had to say ended up letting AI do most of the work which, frankly, just wasn’t that accurate.”
He’s hopeful that newer, more sophisticated iterations of AI medical imaging platforms (roughly 18-20 models are in development) can ultimately save women’s lives, particularly in areas where radiologists are in short supply.
But he believes it will be a long time before doctors, or their patients, are willing to risk postponing a mammogram based on an algorithm.
A version of this article appeared on Medscape.com.
<!--$RCSfile: InCopy_agile.xsl,v $ $Revision: 1.35 $-->
<!--$RCSfile: drupal.xsl,v $ $Revision: 1.7 $-->
<root generator="drupal.xsl" gversion="1.7"> <header> <fileName>167417</fileName> <TBEID>0C04F32A.SIG</TBEID> <TBUniqueIdentifier>MD_0C04F32A</TBUniqueIdentifier> <newsOrJournal>News</newsOrJournal> <publisherName>Frontline Medical Communications</publisherName> <storyname/> <articleType>2</articleType> <TBLocation>QC Done-All Pubs</TBLocation> <QCDate>20240322T105344</QCDate> <firstPublished>20240322T110548</firstPublished> <LastPublished>20240322T110548</LastPublished> <pubStatus qcode="stat:"/> <embargoDate/> <killDate/> <CMSDate>20240322T110547</CMSDate> <articleSource/> <facebookInfo/> <meetingNumber/> <byline>Lisa Marshall</byline> <bylineText>LISA MARSHALL</bylineText> <bylineFull>LISA MARSHALL</bylineFull> <bylineTitleText/> <USOrGlobal/> <wireDocType/> <newsDocType/> <journalDocType/> <linkLabel/> <pageRange/> <citation/> <quizID/> <indexIssueDate/> <itemClass qcode="ninat:text"/> <provider qcode="provider:imng"> <name>IMNG Medical Media</name> <rightsInfo> <copyrightHolder> <name>Frontline Medical News</name> </copyrightHolder> <copyrightNotice>Copyright (c) 2015 Frontline Medical News, a Frontline Medical Communications Inc. company. All rights reserved. This material may not be published, broadcast, copied, or otherwise reproduced or distributed without the prior written permission of Frontline Medical Communications Inc.</copyrightNotice> </rightsInfo> </provider> <abstract/> <metaDescription>A new way of using artificial intelligence (AI) can predict breast cancer 5 years in advance with impressive accuracy — and unlike previous AI models, we know h</metaDescription> <articlePDF/> <teaserImage/> <teaser>The model works to predict risk based solely on “local bilateral dissimilarity.” </teaser> <title>New Transparent AI Predicts Breast Cancer 5 Years Out</title> <deck/> <disclaimer/> <AuthorList/> <articleURL/> <doi/> <pubMedID/> <publishXMLStatus/> <publishXMLVersion>1</publishXMLVersion> <useEISSN>0</useEISSN> <urgency/> <pubPubdateYear/> <pubPubdateMonth/> <pubPubdateDay/> <pubVolume/> <pubNumber/> <wireChannels/> <primaryCMSID/> <CMSIDs/> <keywords/> <seeAlsos/> <publications_g> <publicationData> <publicationCode>fp</publicationCode> <pubIssueName/> <pubArticleType/> <pubTopics/> <pubCategories/> <pubSections/> </publicationData> <publicationData> <publicationCode>im</publicationCode> <pubIssueName/> <pubArticleType/> <pubTopics/> <pubCategories/> <pubSections/> </publicationData> <publicationData> <publicationCode>ob</publicationCode> <pubIssueName/> <pubArticleType/> <pubTopics/> <pubCategories/> <pubSections/> </publicationData> <publicationData> <publicationCode>oncr</publicationCode> <pubIssueName/> <pubArticleType/> <pubTopics/> <pubCategories/> <pubSections/> </publicationData> </publications_g> <publications> <term>15</term> <term>21</term> <term canonical="true">23</term> <term>31</term> </publications> <sections> <term>27970</term> <term canonical="true">39313</term> </sections> <topics> <term>263</term> <term>322</term> <term canonical="true">192</term> </topics> <links/> </header> <itemSet> <newsItem> <itemMeta> <itemRole>Main</itemRole> <itemClass>text</itemClass> <title>New Transparent AI Predicts Breast Cancer 5 Years Out</title> <deck/> </itemMeta> <itemContent> <p>A new way of using artificial intelligence (AI) can predict <span class="Hyperlink">breast cancer</span> 5 years in advance with impressive accuracy — and unlike previous AI models, we know how this one works.</p> <p>The new AI system, called AsymMirai, simplifies previous models by solely comparing differences between right and left breasts to predict risk. It could potentially save lives, prevent unnecessary testing, and save the healthcare system money, its creators say.<br/><br/>“With traditional AI, you ask it a question and it spits out an answer, but no one really knows how it makes its decisions. It’s a black box,” said Jon Donnelly, a PhD student in the department of computer science at Duke University, Durham, North Carolina, and first author <span class="Hyperlink"><a href="https://pubs.rsna.org/doi/10.1148/radiol.232780">on a new paper</a></span> in <em>Radiology</em> describing the model.<br/><br/>“With our approach, people know how the algorithm comes up with its output so they can fact-check it and trust it,” he said.<br/><br/>One in eight women will develop invasive breast cancer, and 1 in 39 will die from it. Mammograms miss about <span class="Hyperlink">20%</span> of breast cancers. (The shortcomings of genetic screening and mammograms received extra attention recently when actress Olivia Munn <span class="Hyperlink"><a href="https://www.instagram.com/p/C4dXfrULDdJ/?igsh=NzBmMjdhZWRiYQ%3D%3D&img_index=1">disclosed</a></span> that she’d been treated for an aggressive form of breast cancer despite a normal <span class="Hyperlink">mammogram</span> and a negative genetic test.)<br/><br/>The model could help doctors bring the often-abstract idea of AI to the bedside in a meaningful way, said radiologist Vivianne Freitas, MD, assistant professor of medical imaging at the University of Toronto.<br/><br/>“This marks a new chapter in the field of AI,” said Dr. Freitas, who authored an <span class="Hyperlink"><a href="https://pubs.rsna.org/doi/10.1148/radiol.240474">editorial</a></span> lauding the new paper. “It makes AI more tangible and understandable, thereby improving its potential for acceptance.”<br/><br/></p> <h2>AI as a Second Set of Eyes</h2> <p>Mr. Donnelly described AsymMirai as a simpler, more transparent, and easier-to-use version of Mirai, a breakthrough AI model which made headlines in 2021 with its promise to determine with unprecedented accuracy whether a patient is likely to get breast cancer within the next 5 years.<br/><br/><span class="Hyperlink"><a href="https://www.science.org/doi/10.1126/scitranslmed.aba4373">Mirai identified</a></span> up to twice as many future cancer diagnoses as the conventional risk calculator Tyrer-Cuzick. It also maintained accuracy across a diverse set of patients — a notable plus for two fields (AI and healthcare) notorious for delivering poorer results for minorities.<br/><br/>Tyrer-Cuzick and other lower-tech risk calculators use personal and family history to statistically calculate risk. Mirai, on the other hand, analyzes countless bits of raw data embedded in a mammogram to decipher patterns a radiologist’s eyes may not catch. Four images, including two angles from each breast, are fed into the model, which produces a score between 0 and 1 to indicate the person’s risk of getting breast cancer in 1, 3, or 5 years.<br/><br/>But <span class="Hyperlink"><a href="https://www.washingtonpost.com/technology/2021/12/21/mammogram-artificial-intelligence-cancer-prediction/">even Mirai’s creators</a></span> have conceded they didn’t know exactly how it arrives at that score — a fact that has fueled hesitancy among clinicians.<br/><br/>Study coauthor Fides Schwartz, MD, a radiologist at Brigham and Women’s Hospital, Boston, said researchers were able to crack the code on Mirai’s “black box,” finding that its scores were largely determined by assessing subtle differences between right breast tissue and left breast tissue.<br/><br/>Knowing this, the research team simplified the model to predict risk based solely on “local bilateral dissimilarity.” AsymMirai was born.<br/><br/>The team then used AsymMirai to look back at > 200,000 mammograms from nearly 82,000 patients. They found it worked nearly as well as its predecessor, assigning a higher risk to those who would go on to develop cancer 66% of the time (vs Mirai’s 71%). In patients where it noticed the same asymmetry multiple years in a row it worked even better, with an 88% chance of giving people who would develop cancer later a higher score than those who would not.<br/><br/>“We found that we can, with surprisingly high accuracy, predict whether a woman will develop cancer in the next 1-5 years based solely on localized differences between her left and right breast tissue,” said Mr. Donnelly.</p> <p>Dr. Schwartz imagines a day when radiologists could use the model to help develop personalized screening strategies for patients. Doctors might advise those with higher scores to get screened more often than guidelines suggest, supplement mammograms with an MRI , and keep a close watch on trouble spots identified by AI.<br/><br/>“For people with really low risk, on the other hand, maybe we can save them an annual exam that’s not super pleasant and might not be necessary,” said Dr. Schwartz.<br/><br/></p> <h2>Cautious Optimism</h2> <p>Robert Smith, PhD, senior vice president of early cancer detection science at the American Cancer Society, noted that AI has been used for decades to try to reduce radiologists’ workload and improve diagnoses.<br/><br/>“But AI just never really lived up to its fullest potential,” Dr. Smith said, “quite often because it was being used as a crutch by inexperienced radiologists who, instead of interpreting the mammogram and then seeing what AI had to say ended up letting AI do most of the work which, frankly, just wasn’t that accurate.”<br/><br/>He’s hopeful that newer, more sophisticated iterations of AI medical imaging platforms (roughly 18-20 models are in development) can ultimately save women’s lives, particularly in areas where radiologists are in short supply.<br/><br/>But he believes it will be a long time before doctors, or their patients, are willing to risk postponing a mammogram based on an algorithm.<br/><br/></p> <p> <em>A version of this article appeared on <span class="Hyperlink"><a href="https://www.medscape.com/viewarticle/new-transparent-ai-predicts-breast-cancer-5-years-out-2024a100059l">Medscape.com</a></span>.</em> </p> </itemContent> </newsItem> <newsItem> <itemMeta> <itemRole>teaser</itemRole> <itemClass>text</itemClass> <title/> <deck/> </itemMeta> <itemContent> </itemContent> </newsItem> </itemSet></root>