At every turn, artificial intelligence (AI) continues to blossom as a coveted tool used in radiology research as the industry strives to identify ways to can improve diagnostic and productivity performance. That is largely where its utility has stalled, however. The vast horizon of AI clinical implementation awaits, but the specialty has work to do before it reaches that goal, one industry expert says.

2020 virtual meeting, Nickolas Papanikolaou, Ph.D., head of Computational Clinical Imaging Group, Center for the Unknown, at Champalimand Foundation in Lisbon, Portugal, discussed the current state of AI in medical imaging, as well as some of the challenges this technology faces when approaching clinical integration.

“The main driver behind artificial intelligence applications in radiology has been the necessity for greater efficacy and efficiency in clinical care,” Papanikolaou said. “Diagnostic demand continues to grow at a disproportionate rate when compared to the number of available board-certified radiologists.”

Already, the radiologist industry has made some headway in using AI models with prostate, lung, and breast cancer, he said. But, these applications have fallen largely within areas of research, and there is still much room for improvement.

The need for the augmented use of artificial intelligence with active patient care exists, though. And, as the specialty continues to move in this direction, there are still many things to consider before these tools can become commonplace, he stressed. One area with particular promise is radiomics – the combination of computer science, biostatistics, and imaging that can be used to identify patterns that can make predictions to inform clinical decision-making.

“Among the most exciting expectations we have for this technology is to be able to stratify patients into sub-groups non-invasively and at an early stage in a way that we can optimize treatment strategy and clinical outcomes,” he said.

He noted that the time is rapidly approaching, however, for radiology to no longer blindly rely on data when it comes to assessing and analyzing certain disease states. Instead, the industry could benefit from more human-centric AI that builds on human experience and knowledge to reduce bias and improve explainability and transparency. Doing so would make AI tools more reliable, he said.

What Radiologists Must Consider and Remember About AI

To date, however, the data to support the clinical use of AI models has been lackluster. In a query of existing literature, Papanikolaou said he found 20,500 articles that touched on radiomics, but less than 1 percent were sufficiently robust. Only 25 studies employed external validation for their AI models, and just 14 compared the performance of AI directly to that of health professionals using the same test sample.

From a purely data-focused perspective, said co-presenter Ben Glocker, Ph.D., a reader in machine learning for imaging from Imperial College in London, radiologists must consider four characteristics of the data used in creating, training, or validating models used in studies. Pay attention to data quality, data variety, data volume, and data readiness, he cautioned.

Given these parameters, Papanikolaou stressed, it is critical for radiologists to approach any existing literature armed with a cadre of questions. Examine the study design – how many patients were involved, was it a single or multi-institutional study, is it a retrospective or prospective design, and does it use standard imaging protocols? In addition, look at the way segmentation was used – does it concentrate on single-slice or whole-lesion segmentation – and how many radiologists were involved? And, lastly, investigate whether the study team address pre-processing steps, such as bias field or motion correction. Be sure to give attention to other technical aspects, as well, including feature reduction, model training, validation, and performance assessment.

“These are technicalities that radiologists will claim,” Papanikolaou said. “But, in the same way that MR physics has been incorporated into the radiological lexicon a couple of decades ago, something similar must take place with data science and radiomics.”

Even with the potential use of radiomics in clinical settings, there are still many stumbling blocks that must be considered, he said. Several spring from the models themselves, but there are multiple industry-based obstacles that will also make it difficult.

For example, he said, these models are data-hungry and could require millions of patients to be properly trained, and they are also very resource heavy. They do not interpret uncertainty well, have low interpretability, and are vulnerable to spurious associations. Additionally, they are difficult to optimize, are prone to over-fitting, and they can be easily fooled by adversarial images.

And, even if these features are fixed, Papanikolauo said, larger system barriers remain. The regulatory process is complex and challenging, and there are few approved products currently on the market. More studies are also needed to highlight the effectiveness of deep learning algorithms in bolstering radiologist performance in real-world clinical settings. And, there is a significant need for greater collaboration between algorithm developers and image IT vendors to ensure that any solutions can be truly integrated into the radiology workflow.

Despite these roadblocks standing in the way, he said, the future for the use of AI in patient care looks bright. There is much room for greater development in the clinical applications of artificial intelligence in radiology.

“We are only in the very first steps of a symbiotic relationship with machines,” Papanikolaou said. “They are here to improve the radiologist, not to replace them.”

Articles

Artificial intelligence tools have many drawbacks and face many barriers to clinical implementation, and radiologists must pay close attention to how – and what kind of – data is used in models. 

AI Has a Clinical Role, But Many Challenges Exist

Using PSMA PET/CT can change the disease treatment and management in two-thirds of patients who have prostate cancer, new research has confirmed.

In a presentation at the Society of Nuclear Medicine and Molecular Imaging 2020 virtual meeting, investigators from the University of Toronto in Ontario, Canada, discussed their assessment of the impact 18F-DCFPyL prostate-specific membrane antigen (PSMA) PET/CT has on how a patient’s cancer management is directed.

Additional PET/CT Tracer to Detect Prostate Cancer

As the second most-common cancer among American men, prostate cancer strikes approximately one in nine patients during their lifetimes. Among that group, 30 percent-to-40 percent have a biochemical recurrence when their PSA levels rise after initial treatment.

DCFPyL PSMA PET/CT is known to effectively pinpoint prostate cancer. But, to determine how impactful it is in managing patients with a suspected limited recurrence, the team, led by Ur Metser, M.D., professor of radiology at the University of Toronto, conducted a prospective, large-scale multi-center trial with 410 men who had biochemical failure post-primary therapy. These men either had no or limited disease on CT scans and bone scintigraphy, and they had already received several prostate cancer treatments.

“The identification of extent of recurrent and specific sites of recurrence is crucial in determining the most appropriate mode of therapy for these men,” Metser said. “Findings from this study add to the body of evidence on the utility of PSMA PET in the management of prostate cancer patients.”

PSMA-PET/CT Outpaces Conventional Imaging with High-Risk Prostate Cancer Detection

According to the study results, he said, PSMA PET/CT found disease in more than half of the men who had negative CT and bone scan scintigraphy. The PET/CT also identified additional disease sites in roughly two-thirds of men who had limited metastases detected prior to the scan. These findings led to PSMA PET-directed management changes in 66 percent of patients, including conversion from observation or systemic therapy to surgery or radiation or adding nodal-directed therapy to salvage surgery or radiation.

While PSMA PET is still investigational in North America, he said, this study’s evidence could support efforts for regulatory approval to make molecular imaging with 18F-DCFPyL more widely available. It could also open the door for clinical trials to assess whether it can effectively be incorporated into patient care as a treatment planning tool.

SNMMI

Using 18F-DCPyL PSMA PET/CT can pinpoint more cancers in men who have biochemical failure with initial treatment.

PSMA PET/CT Prompts Prostate Cancer Management Changes in 66 Percent of Patients With Recurrent Disease

Patients with knee arthritis have abnormally high bone metabolism, according to images captured with new functional bone imaging tactic.

In a presentation at the Society for Nuclear Medicine and Molecular Imaging 2020 virtual annual meeting, investigators from Stanford University discussed their use of PET/MRI with 18F NaF to assess the degeneration of the knee joint in patients, and marker for osteoarthritis.

“Osteoarthritis is not very well understood, in part because we lack the tools to objectively evaluate early and reversible changes in key tissues,” said Lauren Watkins, MS, a researcher in the Stanford University Imaging of Musculoskeletal Function Group. “While many MRI methods have been developed for assessment of early degenerative changes in cartilage, functional imaging of bone in the joint remains a major challenge.”

To assess the interplay between structural and physiological changes in knee osteoarthritis, the team used PET/MRI 18F NaF to image 12 subjects’ knees. They also tested the repeatability of the technique by imaging five patients twice, five days apart.

A musculoskeletal radiologist provided additional evaluation with an MRI Osteoarthritis Knee Score (MOAKS) for each knee. In addition, data collected from dynamic PET was used to determine the rates of bone perfusion, tissue clearance, and mineralization. That information was also used for tracer extraction fraction and total bone uptake rate. Additionally, the investigators  performed kinetic modeling for regions representing the subchondral bone of the patella, medial and lateral tibia, and anterior, central, and posterior regions of the medial and lateral femur. 

Based on their MOAKS findings, the team divided knees into regions of cartilage loss, bone marrow lesions, and osteophytes and analyzed them along with the PET-data derived kinetic parameters. They determined there is a strong association between abnormal bone metabolism in regions with bone marrow lesions, osteophytes, and adjacent cartilage lesions and greater bone perfusion rates when compared to bone that appears normal on MRI.

These results highlight the usefulness and potential of PET imaging in studying the role bone physiology plays in degenerative bone disease, Wakins said.

“This knowledge may help us understand the order of events leading to structural and functional degeneration of the knee,” she said. “Further, this will help us to develop and quickly evaluate new interventions that target specific metabolic pathways to give us the best chance to slow or arrest the onset and progression of osteoarthritis.”

Using PET/MRI 18F NaF offers new functional measure for assessing degeneration of knee joints.

New MRI Approach Helps Assess Knee Osteoarthritis Progression

F-FDG PET/MRI can reveal where discomfort is coming from in patients who have chronic pain, potentially changing their treatment plans, new research has found.

In a presentation at the Society of Nuclear Medicine and Molecular Imaging 2020 annual meeting, researchers from Stanford University School of Medicine discussed the results of their investigation, highlighting the clinical implications for being able to address chronic pain, a condition that affects nearly 50 million adults nationwide and contributes to $635 billion in healthcare expenses, including imaging and treatment costs.

“In the past few decades, we have confirmed that anatomic-based imaging approaches, such as conventional MRI, are unhelpful in identifying chronic pain generators,” said Sandip Biswal, M.D., musculoskeletal radiologist and associate professor of radiology at Stanford University School of Medicine.

F-FDG PET is known to accurately evaluate the rise in glucose metabolism that comes from acute or chronic pain generators, he said. Consequently, his team looked at PET/MRI as a possible solution for determining the precise molecular basis for a patient’s pain.

To make this determination, his team conducted head-to-foot 

F-FDG PET/MRI scans on 65 patients with chronic pain. They used image analysis software to measure maximum standardized uptake values and target-to-background ratios, and two radiologists examined the images to identify whether an uptick in 18F-FDG uptake occurred at symptoms sites or elsewhere in the body.

The radiologists discussed their examinations with referring providers to determine if any changes in treatment were needed.

According to the study analysis, 58 out of 65 patients experienced an increased uptake of 18F-FDG in the nerves and muscles that patients pointed to as pain sites. As a result, 16 patients received mild recommended changes to their management plan, including adding another diagnostic test. However, 36 patients had significant recommended changes, such as the ordering of new invasive procedures. All total, 40 patients received new treatment plans that referring providers had not anticipated.

Overall, Biswal said, these results show that it is possible to secure better outcomes for patients who live with chronic pain. Maximizing the potential, though, could require collaboration between several areas of radiology.

“The clinical molecular imaging approach is addressing a tremendous unmet clinical need, and I am hopeful that this work will lay the groundwork for the birth of a new sub-specialty in nuclear medicine and radiology,” he said. “Using this approach will require knowledge and expertise not only in nuclear medicine, but also in musculoskeletal imaging, neuroradiology and potentially other fields, such as body imaging and pediatric radiology, where pain syndromes are important clinical problems.”

18F-FDG PET/MRI can identify pain generator locations, leading to changes in pain management.


PET/MRI Can Specifically Pinpoint Chronic Pain Source

Working remotely can be a challenge for any radiologist under the best of circumstances, but the current environment controlled by the COVID-19 pandemic has highlighted the need for information technology (IT) solutions that can help providers and their staff maximize their productivity in this situation.

During SIIM2020, company leaders from vendor Konica Minolta Healthcare Americas, Inc., highlighted the challenges radiologists and their IT colleagues face when trying to streamline remote workstations and workflow, as well as some solutions to overcome these obstacles.

According to Chris Van Houten, Konica Minolta’s vice president of sales, radiologist and IT professionals face a myriad of problems when trying to create a seamless workflow that allows providers to download, access, and send out image files with ease. Every facility should be aware of these stumbling blocks so they can ensure any solution they choose addresses the pain points.

In addition to the cost of updated technology, radiologists could find themselves facing these obstacles:

 Without sufficient bandwidth, radiologists can find themselves waiting for images – particularly advanced images – to load into their system. This can bring productivity to a screeching halt, Van Houten said, as some files, such as 3D mammograms, can take up to 20 minutes to download. 

 It is commonly said in the industry that it is imperative for the right radiologist to read the right study at the right time. Without a way to filter through studies in the worklist, meeting that goal can be difficult. For a streamlined workflow system, a worklist must include a filter so providers can sift through the waiting studies for the most urgent ones within their specialty area, he said.

 Even the most up-to-date, tech savvy office can be stymied when working with referring clinicians or other collaborators that are relying on older technologies. In these situations, he said, it is harder to share all types of files. 

IT professionals also encounter their own stumbling blocks, Van Houten explained. In most situations, upgrading a workstation’s software means physically working with the machine. With many providers currently working remotely, IT departments have had to find ways to access and upgrade each system, and the process can be time consuming. Also, if an office purchases any new software, they must make sure that the workstations can be updated to leverage it appropriately.

Fortunately, said Kevin Borden, Konica Minolta’s vice president of product, the industry does offer two solutions to help radiology offices overcome these challenges. Pairing server-side rendering with zero-footprint viewers can pave the way for rapid, smooth workflow.

 This solution eliminates the need for traditional workstations to process all the calculations needed associated with reading images. Instead of downloading images to the radiologist’s browser, SSR renders the images in the server and send it to the workstation, alleviating the stress from the workstation itself and allowing for faster downloading and processing.

“Server-side rendering does all of the heavy lifting that a traditional workstation had to do,” he said. “It performs all of the calculations, so your expensive powerful server-based workstations don’t really have to exist.”

Not only does this solution allow for large files and prior studies to be downloaded and accessed quickly, but it also side-steps the need to push whole sets of DICOM data to a workstation.

 Implementing this technology requires no installation on workstations, and it functions on any operating system or browser. The high diagnostic quality and full tool set is available for use on any modality at any workstation, Borden explained, even for radiologists working remotely.

As with the challenges, radiologists and IT professionals experience their own unique benefits, especially when SSR and ZFP are combined, Borden said, in a product, such as Konica Minolta’s Exa® Enterprise Imaging.

For the radiologist, a combined platform makes accessing larger files, including 3D mammography or echocardiograms, much faster without any pre-fetching or moving data from workstation to workstation. It also frees the provider up – he or she is no longer strapped to a specific workstation because they can now view any modality from any location.

“The radiologist can pick up a laptop with an approved monitor, and as with any workstation before, hit a URL, log into the application, and begin reading remotely,” he said.

IT professionals also enjoy several benefits with an SSR+ZFP platform, Borden explained. By not having to physically touch workstation, their workload is significantly lighter. And, without installing applications or moving data around frequently, there is less strain on a department or office’s network, reducing utilization.

The benefit that has been most helpful during the pandemic, however, he pointed out, is that with this type of platform – with no installations or downloads required – once the appropriate equipment is in place, radiologists can begin to immediately read remotely.

“In today’s environment, that’s very important,” he said. “Before the pandemic and post-pandemic, the future needs to be now downloads or installations necessary.”

SIIM

Konica Minolta leaders outline the challenges to – and discuss tech solutions for -- provider and IT productivity during SIIM2020.

Technology for Improving Productivity and Efficiency

A custom-designed app made specifically for radiologists can successfully encourage providers to submit reports of incidents that could negatively impact patient safety, workflow, and quality, according to research presented at 

Historically, incident reports have been largely submitted by radiology technologists. To broaden participation in submitting these reports, researchers from Thomas Jefferson University Sidney Kimmel Medical College created an internal incident and issue reporting app to not only capture more episodes, but to also increase radiologist involvement with the hopes of acting upon issues within six months.

“Customized reporting tools are a useful supplement to existing systems to capture issues that may impact quality and workflow, may have downstream effects on patient safety, and may otherwise go unreported,” investigators said. “A custom application built for radiologists can increase their engagement in the reporting process, by providing a tool for discussing and resolving latent but important issues.”

Their six-month intervention proved the new tool could successfully reach that goal. According to their analysis, radiologist involvement spiked from 5 percent prior to the tool launch to 97 percent of all submissions. They, consequently, accounted for 46 percent of all submissions across both the enterprise and internal tools.

The findings also showed several other positive outcomes. While the existing enterprise tool collected roughly the same number of incident reports (215 pre-launch and 203 post-launch), the internal tool received an additional 168 submissions – a 72-percent increase. Additionally, submissions via the new tool spanned seven categories with a 62-percent rise in the types of issues captured. Workflow, protocol, IT, and communications issues were the most frequently reported.

By the end of the six-month intervention, 37 percent of reports submitted via the internal app had been either triaged or resolved.

The research team collected their data from March 1, 2019, to Sept. 1, 2019 via the new mobile-friendly app. Radiology staff voluntarily submitted confidential and optionally anonymous reports. The team compared pre- and post-launch data for the number of submissions, incident categories, severity scores, and staff roles. In addition, they took steps to ensure the new tool did not replace the existing enterprise system, and they sent out reminders for the staff and providers to use the app, encouraged user engagement through a feedback loop, created a volunteer committee to resolve submissions.

Internal reporting app, designed for radiologists, increased provider participation and captured incidents in a greater number of categories.


Improving Incident Reporting: There’s an App for That

Turnaround times are slower for radiologists who use more mouse clicks to work through radiographs, according to new research presented at 

, highlighting the need for PACS efficiency training.

In recent years, there has been a simultaneous push for providers to process more studies while also decreasing the amount of time it takes for them to successfully complete their radiology reports. Consequently, said investigators from Cincinnati Children’s Hospital, it is becoming increasingly important for all providers to use their PACS system effectively and efficiently.

Dashboard Reduces Turnaround Time, Fosters Competition Between Radiologists

To see whether radiologists need a different number of mouse clicks to process a study – and whether that number correlates to turnaround time – the team observed six providers over a four-week period. They determined their hypothesis was correct.

“There is a wide variation in the number of mouse-clicks radiologists used to read studies,” the researchers said. “The average number of clicks correlates with the average time required to read the study.”

Radiology Turnaround Times: Are They Ever Fast Enough?

During the study, they observed workflow when participating radiologists dictated five-to-10 radiographs in succession. They collected data either through direct observations or with a mouse-click counting application on several factors:

Overall number of clicks specifically occurring within the PACS application

Number of distinct images for each imaging study

Their analysis confirmed their hypothesis – radiologists who used more clicks took longer to read radiographs. For each radiologist’s turnaround time, the one-way ANOVA to determine significance produced a 0.095 p-value.

These result will be used to create improvement projects that will help radiologists learn to use PACS more efficiently.

Radiologists who use more mouse clicks to process radiographs have higher turnaround times.

Mouse Clicks: Higher Number Means Slower Turnaround Time

Patients have an easier time understanding mammogram reports that include text and picture notes than radiology reports that do not include explanations, according to research presented at 

As an increasing number of patients now have access to their breast imaging reports via patient portals, investigators from the University of Pennsylvania hypothesized they would be better able to comprehend their test results if the reports also included text and illustrations rather than only definitions or only illustrations.

To test their theory, they surveyed 300 women with an average age of 43.6 about the complexity of mammography reports. First, the participants were asked to rate, on a scale of 1-to-5, how difficult an original unannotated mammogram report was to read. They were, then, asked to do the same thing for three different versions of the report: text-only, pictures-only, and text-and-pictures.

Evolution of Radiology Reports: Getting Information To the Patient

According to the results, the women, 70 percent of whom had at least a college degree and 64 percent of whom had undergone at least one mammogram, rated the difficulty level of the original report at 2.9. They rated the text-only and pictures-only versions as 2.1 and 2.6, respectively. The text-and-pictures report received a 2.3 rating.

“Understanding was significantly better for the ‘text only’ reported compared to the ‘pictures only’ version, although there was no difference between the ‘text and picture’ version and either of the other two annotated versions,” the researchers said.

Additionally, when the team isolated responses from women at least 50-years-old – 95 percent of whom had had a mammogram previously – there was still a statistically significant difference in their ease of understanding between the original report and the annotated versions.

Patients rate mammogram reports with annotations as easier to comprehend.


Text and Picture Notes Improve Patient’s Understanding of Mammography Reports

An artificial intelligence model based on MRI images can accurately identify which brain lesions cause mass effect and which do not for a wide variety of brain diseases, according to findings presented at 

Brain lesions have a number of diagnostically and prognostically relevant features – a key one being whether it exerts a “mass effect” – a distortion or compression of ventricles and sulci. To make this determination, investigators from the University of Pennsylvania, University of California San Francisco, and University of Texas Austin developed a convolutional neural network (CNN) that can be used across many underlying pathologies.

To create the CNN and achieve a high level of accuracy, they extracted T1 and T2-FLAIR images from 384 MRI studies (298 negative, 88 positive) of patients who had 60 different disease entities. Of those images 189 were used for training, 54 for validation, and 142 for a held-out test, ensuring that two-to-three samples were in the test set. In addition, they extracted cerebral spinal fluid masks from the T1 images, and the brain was extracted from the FLAIR image.

Overall, researchers determined their mass effect detection model was 84.5 percent accurate. The network, they said, performed well across 53 disease entities.

As next steps, the team plans to: classify negative mass effect, use multiple channels to improve tissue segmentation, and produce saliency maps to determine why the classifier is making its decisions.

Convolutional neural network accurately identifies “mass effect” lesions in more than 50 disease entities.

AI Model Uses MRI To Identify Brain Lesions Causing Mass Effect

Radiology has always been on the cutting edge of technology, and the latest tech tool craze is still artificial intelligence (AI). But, as the industry continues to push forward, finding new ways to make use of these algorithms, it is critical for radiology leaders to also pay attention to how AI is impacting the provider, said one industry expert.

Using AI is about more than data results in a lab, said Elizabeth Krupinski, Ph.D., professor and vice chair of research in the radiology and imaging sciences department at Emory University School of Medicine. It is about how the technology touches the radiologist’s thinking, perceptions, workflow, and other human factors.

“We really need to understand and appreciate how technology – no matter how cool it is to develop and implement – impacts the human user at the end,” she said, during the 2020 Dwyer Lecture at the 

 virtual meeting. “And, after that, how it impacts patient care.”

Despite being widely used throughout the industry for several years now, there has not been much investigation into the presentation of AI results during clinical use or how these tools affect providers in cognitive, perceptual, or ergonomic ways. There is also a lack of data on how AI touches patient outcomes, radiologist training, and decision-making, she added.

“Basically, what’s missing in a lot of the AI, deep learning, and some of the technology development – even hardware – is a consideration of the users,” she explained. “And, I think that ends up being critically important.”

Throughout the industry, she said, a growing population of radiologist are also turning to AI tools to help alleviate the stress, fatigue, and burnout that has become increasingly prevalent in the specialty, particularly when it comes to complex research efforts. But, even when AI is used in investigations, sufficient context about clinical significance or relevance is rarely included.

In many cases, Krupinksi said, investigators will identify a particular type of image or finding, apply an AI scheme, and, then, point to the strength of their outcomes.

“But, then, when you read it, there’s nothing in the significance or the background section that says why they’re doing this,” she said. “Is it a problem right now? Do radiologists have a problem with that particular type of diagnosis or type of image?”

Collecting a true assessment of any new method, initiative, or innovation in an actual clinical setting can be complicated and costly. But, quasi-experimental designs could be options for reach this goal, she said.

Overall, she recommended three types of quasi-experimental designs:

 In a direct comparison, a group with an AI intervention is analyzed – both before and after adoption – with a non-randomized, similar control group that does not receive AI.

 Investigators measure outcomes at consecutive time points (three-to-eight) prior to and after an intervention is conducted. All measurements occur within the same group, as it functions as its own control.

 The intervention is launched at various sites over a certain time period with control groups eventually receiving the intervention and functioning as their own controls. This design allows investigators to compare the intervention by different sites or groups, enabling assessment across institutions, practices, and patients groups.

Implementing these designs offers several advantages, she said. Although they can have biases and errors, they provide for faster intervention uptake, enhanced acceptability, and reduced cost.

Videos

The industry should pay more attention to the interplay between providers and AI tools, says Elizabeth Krupinski, Ph.D., during SIIM 2020.

Latest Conference Articles