AI Has a Clinical Role, But Many Challenges Exist

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.

In a presentation given during the European College of Radiology 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.

Related Content: AI Is Not the End of Radiology’s World

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.”

The Stop Sign at Clinical Care

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.”

Read more of Diagnostic Imaging's ECR 2020 coverage here.