Photo: Kajetan Sumila/Unsplash
Researchers from The Chinese University of Hong Kong and Beijing Tongren Hospital have developed a foundational AI model for automated eye disease diagnosis.
WHAT IT'S ABOUT
The generative AI model, VisionFM, was pre-trained on a large ophthalmic data cohort containing 3.4 million images from over 500,000 individuals worldwide. The pre-training dataset covers eight eye imaging modalities, various diseases, and clinical scenarios.
VisionFM has been tested on an ophthalmic database comprising 53 public and 12 private datasets for multiple applications. These include eye disease diagnosis, progression prediction, systemic biomarker prediction through ocular imaging, intracranial tumour prediction, and lesion, vessel, and layer segmentation.
Findings from this test have been published in NEJM AI of the New England Journal of Medicine.
FINDINGS
The study found that the genAI model demonstrated accuracy in diagnosing eye diseases comparable to an opthalmologist with between four and eight years of clinical experience. "In a comparative study of diagnostic accuracy of 12 ocular diseases from fundus photographs, VisionFM shows diagnostic accuracy close to that of intermediate-level ophthalmologists."
It even accurately graded diabetic retinopathy (over 90%) using an imaging modality it was never exposed to during pre-training.
Another interesting finding was VisionFM's ability to predict the presence of intracranial tumours from fundus images, which the researchers noted is a breakthrough. "While it is relatively hard for ophthalmologists and radiologists to infer the presence of an intracranial tumour from a fundus image, VisionFM can make accurate predictions directly from fundus images, and it is the first AI model that has been validated for this task." Additionally, the model can accurately predict glaucoma progression from fundus photographs.
WHY IT MATTERS
CUHK and Beijing Tongren Hospital claim their foundational model is a first of its kind that goes beyond the limits of similar existing AI models for automated ophthalmic diagnosis. These models, according to the researchers, rely on huge amounts of labelled data, which could be costly and time-consuming to collect. Many existing models also target a single or a few number of eye diseases and utilise only one imaging modality – particularly fundus photographs.
Their significant finding of the model's ability to predict tumours directly from low-cost retinal images – a first among similar models – "holds great potential for the early detection in community and primary care," they emphasised.
According to CUHK, VisionFM has been deployed in China's Henan Province to support the screening of common eye conditions.
THE LARGER TREND
Indeed, most AI tools for screening ophthalmic diseases utilise fundus images, or images of the back of the eye, captured through either professional or smartphone cameras. Such innovation has been reported in Taiwan, South Korea, Singapore and India. About four years ago in China, a multi-institutional research team also built a fundus-based AI system.
Meanwhile last year, Google's AI model specifically intended for screening diabetic retinopathy was licensed to its partners in Thailand and India.