Selecting a single checkbox alongside the words “Enable Auto ID” brings the power of artificial intelligence into routine clinical use in molecular imaging: an action that saves precious time and streamlines processes that would otherwise be tedious.
Photography by Steven Bridges & Lars Berg
Illustration by Joseph Schmidt-Klingenberg
Photography by Steven Bridges & Lars Berg
Illustration by Joseph Schmidt-Klingenberg
The ease of use physicians might one day take for granted stands in contrast to the years of development, validation, and testing performed by a dedicated team of scientists and physicians who created Lesion Scout with Auto ID.[a]
The clinical application—housed within the syngo®.via for Molecular Imaging reading solution—sits at the interface of medicine and artificial intelligence, and opens the door to a future in which care is delivered more efficiently and precisely. The power of Auto ID lies in its potential to dramatically speed workflow for physicians by automatically segmenting and classifying the uptake of 18F-FDG in whole-body PET/CT images as either pathologic or physiologic. Auto ID also enables physicians to calculate whole-body metabolic tumor volume (MTV) and total lesion glycolysis (TLG) within seconds.[b]
“For me, one of the motivations is the potential impact of this type of technology," says Ludovic Sibille, MS, the senior scientist at Siemens Healthineers who developed the algorithm that powers Auto ID. "It has a wide application and a large impact potentially on our customers and on radiology.”
The clinical application—housed within the syngo®.via for Molecular Imaging reading solution—sits at the interface of medicine and artificial intelligence, and opens the door to a future in which care is delivered more efficiently and precisely. The power of Auto ID lies in its potential to dramatically speed workflow for physicians by automatically segmenting and classifying the uptake of 18F-FDG in whole-body PET/CT images as either pathologic or physiologic. Auto ID also enables physicians to calculate whole-body metabolic tumor volume (MTV) and total lesion glycolysis (TLG) within seconds.[b]
“For me, one of the motivations is the potential impact of this type of technology," says Ludovic Sibille, MS, the senior scientist at Siemens Healthineers who developed the algorithm that powers Auto ID. "It has a wide application and a large impact potentially on our customers and on radiology.”
A collaborative journey in nuclear medicine
Collaboration has been an integral part of the journey that has made Auto ID possible. Physicians from the Department of Nuclear Medicine and the European Institute for Molecular Imaging at the University of Münster provided the clinical data that was used to train the artificial intelligence algorithm. In addition, Sibille and his colleagues at Siemens Healthineers have incorporated feedback on early prototypes from meetings and surveys of clinicians and management at institutions ranging from large academic medical centers to smaller regional hospitals.
“For me, one of the motivation is the potential impact of this type of technology. It has a wide application and a large impact potentially on our customers and on radiology.”
Siemens Healthineers
Carl von Gall, MD, product manager for oncology applications at Siemens Healthineers, recalls showing an early prototype of Auto ID to physicians at the 2019 Annual Congress of the European Association of Nuclear Medicine (EANM) and seeing their marvel.
“There’s this ‘aha’ moment when they would open up their eyes and say, ‘You just identified uptake that would need to be included or excluded, with a little help from me, in less than two minutes,'" he recalls, adding that the current iteration of Auto ID has brought the time down to approximately 10 seconds.[c] “They told us pretty directly that Auto ID will be something that they could use every day.”
“There’s this ‘aha’ moment when they would open up their eyes and say, ‘You just identified uptake that would need to be included or excluded, with a little help from me, in less than two minutes,'" he recalls, adding that the current iteration of Auto ID has brought the time down to approximately 10 seconds.[c] “They told us pretty directly that Auto ID will be something that they could use every day.”
Auto ID proposes a selection of foci to be excluded and included from findings creation. Data acquired under IRB-approved protocol.
Data on file.
Artificial intelligence in nuclear medicine grounded in clinical expertise
Auto ID automatically segments and classifies lesions using a type of deep learning algorithm known as a convolutional neural network (CNN). CNNs are often used in image recognition and computer vision, but Sibille explains that the immense amount of data associated with 3D PET/CT images made the development of a CNN that can accurately classify uptake particularly challenging.
Through an iterative process, Sibille worked to refine the parameters that comprise the algorithm. He trained, validated, and tested the algorithm against data grounded in the clinical expertise of Michael Schäfers, MD, and his colleagues at the University of Münster. To create a reference standard to train, validate, and test the algorithm, a board-certified nuclear medicine physician and a board-certified radiologist from the University of Münster analyzed whole-body PET/CT scans of 629 patients with lung cancer or lymphoma. The experts manually delineated foci with increased 18F-FDG uptake and performed more than 12,000 annotations. Like Sibille, Schäfers says he was motivated by the opportunity to have a meaningful impact on the future of medicine.
“There’s hype about artificial intelligence, but most of the approaches suffer from poor data quality or not enough data,” Schäfers says. “If you want to use AI as an expert system, to train people, to support people in their decisions, you have to make sure that the data, the ground truth, is not wrong from the beginning.”
Through an iterative process, Sibille worked to refine the parameters that comprise the algorithm. He trained, validated, and tested the algorithm against data grounded in the clinical expertise of Michael Schäfers, MD, and his colleagues at the University of Münster. To create a reference standard to train, validate, and test the algorithm, a board-certified nuclear medicine physician and a board-certified radiologist from the University of Münster analyzed whole-body PET/CT scans of 629 patients with lung cancer or lymphoma. The experts manually delineated foci with increased 18F-FDG uptake and performed more than 12,000 annotations. Like Sibille, Schäfers says he was motivated by the opportunity to have a meaningful impact on the future of medicine.
“There’s hype about artificial intelligence, but most of the approaches suffer from poor data quality or not enough data,” Schäfers says. “If you want to use AI as an expert system, to train people, to support people in their decisions, you have to make sure that the data, the ground truth, is not wrong from the beginning.”
“There’s hype about artificial intelligence, but most of the approaches suffer from poor data quality or not enough data. If you want to use AI as an expert system, to train people, to support people in their decisions, you have to make sure that the data, the ground truth, is not wrong from the beginning.”
University of Münster
Sibille, Seifert, Schäfers, and several of their colleagues published their findings in the journal Radiology.1 Lymphoma and lung cancer were chosen for the study because they are relatively common metastatic cancers, but a preliminary internal evaluation that examined melanoma, colorectal cancer, breast cancer, cancers of unknown origin, and inflammatory disease found that Auto ID differentiated physiologic vs. pathologic uptake with an overall accuracy of 92%.[d] “The beauty of this is that the physiological uptake pattern is similar in patients and is, to the most degrees, not disease-specific,” von Gall says. “Because that pattern is robust, physicians may be comfortable applying that to their daily routine.”
Additional retrospective studies have continued to evaluate Auto ID’s ability to quantify MTV in lymphoma, breast cancer, and other miscellaneous cancer types. The studies reinforce the workflow of Auto ID when compared to manual segmentation and quantification efforts. The ability of Auto ID to assist in the segmentation and quantification of MTV TLG underscores the prognostic value in the ability to predict overall- and progression-free survival.2-4
Additional retrospective studies have continued to evaluate Auto ID’s ability to quantify MTV in lymphoma, breast cancer, and other miscellaneous cancer types. The studies reinforce the workflow of Auto ID when compared to manual segmentation and quantification efforts. The ability of Auto ID to assist in the segmentation and quantification of MTV TLG underscores the prognostic value in the ability to predict overall- and progression-free survival.2-4
Robert Seifert, MD (left) and Michael Schäfers, MD (right) pictured at
the University of Münster, Münster, Germany.
“The future of our field”
Robert Seifert, MD, of the University of Münster, emphasizes that PET/CT parameters such as MTV and TLG have been shown to provide prognostic value for cancer patients while also potentially providing valuable data for assessing response to treatment. Measuring these parameters has been too tedious and time consuming for routine use, but the speed and accuracy of Auto ID has the potential to enable much broader use.
“The Auto ID functionality and similar approaches are the future of our field.”
University of Münster
Seifert and Schäfers point out several additional ways in which artificial intelligence has the potential to enhance clinical care and basic research, including enabling more nuanced staging of cancer, the analysis of dynamic images, and the integration of imaging data with population-level datasets to advance research. “The Auto ID functionality and similar approaches are the future of our field,” Seifert says.
“AI, if it’s truly meaningful, needs to be almost invisible. Don’t change the reader’s method—support it, add to it, augment it, but don’t change it.”
Siemens Healthineers
From its inception in 2013, the development of Auto ID has been guided by the belief that new technologies should support and streamline the work of physicians without adding complexity or burden. Auto ID provides color-coded proposals for uptake that likely should be excluded or included—green for exclude and orange for include, for example—but the reading physician remains in control and ultimately decides the content of the final report.
“AI, if it’s truly meaningful, needs to be almost invisible,” von Gall says. “Don’t change the reader’s method—support it, add to it, augment it, but don’t change it. So, that’s why the only thing that you’re going to see from Auto ID in the configuration is one checkbox, which says ‘Enable Auto ID.’ And that’s where the magic happens.”
“AI, if it’s truly meaningful, needs to be almost invisible,” von Gall says. “Don’t change the reader’s method—support it, add to it, augment it, but don’t change it. So, that’s why the only thing that you’re going to see from Auto ID in the configuration is one checkbox, which says ‘Enable Auto ID.’ And that’s where the magic happens.”
Lobby vestibule of the Department of Nuclear Medicine at the University of Münster
About the author
Sameh Fahmy, MS, is an award-winning freelance medical and technology journalist based in Athens, Georgia, USA.