AI segment is transforming the practice of radiology in two specific areas: clinical decision support for imaging and increasing revenue.
The landscape at the Radiological Society of North America’s 2020 (RSNA20) conference was different for everyone this year — attendees and exhibitors alike — as we all grappled with the third wave of COVID-19. Through its virtual presence, Konica Minolta highlighted how their advancements in the artificial intelligence (AI) segment are transforming the practice of radiology in two specific areas: clinical decision support for imaging and increasing revenue through natural language understanding (NLU).
Creating Partnerships to Achieve Goals
To further these efforts, Konica Minolta has partnered with 1QBit, an advanced and quantum computing software company with a strong focus on AI, based in Canada. 1QBit uses sophisticated techniques to tackle computationally intractable problems, and both companies share the same goal. “It’s really important to improve patient care with one goal, and that’s with quality systems,” said Kevin Borden, Vice President of HCIT Product at Konica Minolta Healthcare. “To reach this goal, we partnered with 1QBit to provide an unmatched suite of tools to assist radiologists and radiology practices.”
1QBit has a team of machine learning scientists, researchers and software developers who look at various industries with computationally difficult problems and identify where computing could help solve them. “We have spent a fair amount of time and investment building our healthcare vertical. We have learned from working in different computing paradigms and distilled that learning to the classical world of computation where we build responsible solutions based on solid science,” explained Deepak Kaura, M.D., who is the Chief Medical Officer of 1QBit, and also a practicing radiologist. The partnership’s primary focus is on two specific areas: X-ray clinical decision support and NLU.
A Complete Platform with Specialized Viewing
Konica Minolta’s Exa Platform is unique in that it has a single database across all modules — RIS, PACS and Billing. Within this integrated platform, the Exa system can incorporate various applications that provide tools and resources.
“It’s truly one application, and not separate databases, integrated together,” Borden said. “It’s a complete platform with specialized viewing that includes 3-D mammo, echo stress and orthopedic toolsets.”
This tight integration provides many partnership opportunities, the first being the technology that 1QBit has spent nearly three years developing called XrAI. XrAI is a clinical decision support tool that uses machine learning to identify lung abnormalities and its algorithm works within the Konica Minolta Exa platform.
“We see very few solutions out there that attempt to provide a comprehensive review of abnormality detection in the lungs and pleural spaces. These tools have not received regulatory approval yet,” said Kaura. “We’ve trained our algorithm on about 500,000 different images and I think it’s important to recognize that the training dataset was really diverse. It was demographically diverse — all these radiographs were taken using different X-ray machines across a variety of institutions in Canada. We then conducted a clinical trial and the first randomized control trial on the use of AI in the clinical space. The results of that trial were enlightening in that physicians who used XrAI demonstrated significantly improved detection of abnormalities on chest radiographs. We’ve submitted that for peer reviewed publication.”
Kaura stressed the strength of the data gathered and the strength of the performance of the algorithm. As a result, they received Health Canada Class III Medical Device approval. “It’s the first time in Canadian history that any AI tool has received Class III Medical Device clearance, and now we’ve fully integrated into the workflows with Konica Minolta,” he stated. “At present, we’re deploying across Canada and it’s very exciting to see a significant rollout of an imaging-based AI like this. As our partnership with Konica Minolta grows, we anticipate expansion of the solution within and beyond Canada.”
Control in the Hands of Radiologists
One of the synergies that both companies agreed on from the first meeting was giving the radiologist the chance to actually make the diagnosis, while still allowing for adjustments. “XrAI allows the radiologist to adjust the threshold of what he or she might deem a finding versus AI actually telling the clinician that this is a positive finding with no threshold,” explained Borden. “I think that’s very important, and what we’ve done here in terms of the algorithm is one of the first we’ve seen — rightly so, we put the decision in the hands of the radiologists.”
Kaura agreed, adding that most physicians who use the technology would recommend it to their colleagues and institutions, based on the above and the performance of the algorithm. “Providing user-controlled transparency is a significant step towards engendering trust in clinical AI. We are confident that this empowerment of the radiologist will create the standard for displaying AI outputs in the future. Konica Minolta’s single, tight interface of RIS/PACS lends itself perfectly to this unique approach,” said Kaura.
From Language Processing to Language Understanding
Another area Konica Minolta and 1QBit are advancing is Natural Language Understanding (NLU) and how it relates to managing patient follow-up care. NLU represents the next evolution of Natural Language Processing (NLP). Historically, and before NLU, NLP worked by identifying words that were close together to determine conclusions about their meaning. NLP’s negation algorithms relied on words that were in close proximity to the diagnosis term to accurately determine a positive or negative result. For example, in the phrase “no pleural effusions,” the NLP negation algorithm recognized the proximity of the word “no” to “pleural effusions” and interpreted accordingly.
This is a challenge, however, when analyzing long or compound sentences, where the negation terms are not in close proximity to the diagnosis terms. For example, in the phrase “No consolidation, pulmonary edema or pleural effusions,” NLP may not accurately interpret this to mean there are no pleural effusions, because the negating term “no” is more than three words away from “pleural effusions”.
In comparison, NLU uses a neural net-based model to actually understand the context of the entire conversation. The 1QBit NLU engine works to understand the full report, produce an overall result, and generate action items based on all information.
“We are able to take a look at a radiology report and ask the question, ‘Does this patient display any pulmonary abnormalities?’ and because the natural language understanding model actually understands the entire report, it then produces a result that says ‘no’,” Kaura said.
This NLU technology can already be used to help identify and automate patient follow-ups that may be recommended in a radiology report. As a radiologist dictates a report, the NLU engine can take the plain text, process it, identify any necessary patient follow-ups, and then generate a structured schedule order that is sent to the Exa Platform for processing. The order can include the type of appointment, the body part, the time frame and the modality.
“The natural language understanding model understands what things might be relevant to follow-up,” Kaura said. “Even if someone says a CT scan is recommended, it recognizes that that represents follow-up. There are a number of variants on how physicians and clinicians express what follow-up means, and our understanding models actually figure out what that is.”
Originally Published at ITN