A Multi-center Study on the Artificial Intelligence Enabled Diabetic Retinopathy Screening Based on Fundus Images

A Prospective, Multi-center Clinical Study on the Application of An Artificial Intelligence Enabled Disease Detection Software to Diabetic Retinopathy Screening Based on Fundus Images

Early detection and intervention of diabetic retinopathy (DR) is critical in preventing DR-related vision loss among type 1 (T1DM) and type 2 diabetic mellitus (T2DM) patients, currently estimated at over 100 million in China alone. Yet the healthcare resources, particularly retinal specialists, are in short supply and unevenly distributed. In order to help address this enormous mismatch and implement population-based screening, an artificial intelligence (AI) enabled, cloud based software is developed by training a custom-built convolutional neural network.

This study is designed to evaluate the safety and efficacy of such device in detecting referable diabetic retinopathy (moderate non-proliferative DR or worse).

This prospective, multi-center clinical study is designed to validate the performance of an AI enabled software - Shenzhen SiBright AIDRScreening - in detecting referable diabetic retinopathy (RDR, defined as more than mild NPDR) among study subjects primarily by evaluating its sensitivity and specificity.

The subjects enrolled in this trial are patients with T1DM and T2DM. For those who qualify, color fundus images of each eyes are taken and then independently graded for RDR by both the device under test and a centralized reading center, which, for the purpose of this trial, is the Image Reading Center at Zhongshan Ophthalmic Center, Sun Yat-sen University (ZIRC). The grading from ZIRC serves as the gold standard to compare the device performance against.

The trial plans to enroll 1000 subjects. With a 95% confidence interval, the sensitivity is expected to be at least 85% whereas the specificity at 90% or above.

Fundus image quality assessment is performed according to the National DR Screening Imaging and Grading Guideline jointly published by Chinese Ophthalmological Society and Chinese Medical Doctor Association in 2017.

The diagnosis of RDR is based on the National DR Clinical Diagnosis and Treatment Guideline published by Chinese Ophthalmological Society in 2014.

A brief overview of the clinical protocol is as follows:

1. Candidate screening phase: recruiting qualified trial subjects;
2. Clinical phase: imaging and grading by AI and ZIRC;
3. Statistical analysis phase: comparing two outputs;
4. Closing phase: final report and archiving

Inclusion Criteria:

1. Subject must understand the study, has chosen to participate voluntarily, and has signed informed consent
2. Age 18 or older, no limitations on gender identity
3. Type 1 Diabetes Mellitus and Type 2 Diabetes Mellitus

Exclusion Criteria:

1. As it is difficult to obtain fundus images of satisfactory quality with small pupils, mydriasis is advisable under certain circumstances except if:

   1. the subject is allergic to mydriatic drugs;
   2. the subject's intraocular pressure (IOP) ≥ 22 mmHg;
   3. the subject is prone to post-dilation angle closure, pupillary block, etc.;

2. The subject has refractive media opacity and/or pupil abnormalities that affect fundus examination and imaging;

3. The subject has severe vitreous hemorrhage;

4. The subject has received fundus laser treatment;

5. The subject has had eye surgery such as scleral buckling, vitrectomy, macular transposition, etc., BUT cataract surgery or external eye surgery are exempt from exclusion criteria;

6. The subject is participating in other ophthalmic clinical trials;

7. In cases when the researchers believe the subject is not suitable for taking fundus photograph, including but not limited to:

   1. The subject had nystagmus and could not obtain a gradable fundus photograph;
   2. other disqualifying condition deemed by the researcher.