AI Applications in Diagnostic Imaging

Advances in diagnostic imaging have transformed clinical practice in ophthalmology by providing quantitative rather than qualitative information. Now, researchers are looking to use artificial intelligence (AI) to improve diagnostic accuracy and consistency of image inter­pretation. Achieving this goal, however, requires attention to several issues, said Michael F Chiang MD. 

One key challenge for developing useful AI-enabled diag­nostic systems relates to the fact the algorithm performance improves as the task narrows. In the real world, however, diagnosis involves many parallel tasks, as clinicians need to integrate data from images with their clinical findings and account for the possibility that patients can have multiple comorbid diseases.

“AI systems are improving rapidly but have yet to reach this level of sophistication,” Dr Chiang said.

To be robust and avoid bias, AI-enabled image diagnostic systems must also be trained on large, diverse data sets to ensure their application is generalisable to the diverse popu­lations encountered in practice. However, the data being used in the research world where AI systems are being developed is often homogenous because it is extracted from a population representing a single centre or geographic region. 

Recognising the importance of building large, AI-ready data sets to avoid bias, the National Institutes of Health (NIH) has become involved and established a programme known as Bridge2AI that is providing $130 million in funding across four “data generation” groups. A team led by ophthalmolo­gists at the University of Washington in Seattle is one of the recipients of the funds from this NIH-sponsored project, Dr Chiang said.

Data sharing and collaboration for knowledge discovery are also critical for developing AI-enabled diagnostic systems, and this depends on having imaging devices that adopt imag­ing standards.

“The problem, however, is that in many existing ocular im­aging devices, the data are often locked in proprietary formats and therefore cannot be easily exchanged or retrieved,” Dr Chiang said. He provided some personal examples from his career in academia of how this issue impeded clinical produc­tivity and research.

Addressing the need for conformance to DICOM stan­dards, representatives from three US agencies—the Food and Drug Administration, Office of the National Coordinator for Health Information Technology, and the National Eye Institute—held a workshop in May 2022 focused on improv­ing interoperability among ocular imaging modalities and devices. The meeting led to the identification of knowledge gaps, barriers to progress, and potential approaches for overcoming obstacles.

“Levers that can be pulled to advance the field are described in a published position statement.1 In a nutshell, I am not enthusiastic about funding future studies using ocular imaging devices which do not conform to the DICOM standard because they would not represent the best research or clinical care.”

Noting the need for all stakeholders to work together, Dr Chiang said he was heartened by seeing the community of oc­ular imaging device vendors appears to be working towards platforms that conform to DICOM standards. Still, he encour­aged users of ocular imaging devices to insist on adherence to standards through their interactions with vendors in the device industry.