Abstract
Purpose :
Cataract diagnosis is performed by ophthalmologists using a slit-lamp microscope. Non-ophthalmologists is difficult to screen out cataracts because they do not have any slit-lamp instruments or experience to make a diagnosis of cataracts. In general, diagnostic imaging is compatible with artificial intelligence (AI), and AI is used as a diagnostic aid tool in radiological imaging and endoscopic imaging. Currently, in ophthalmology, diagnostic AI is being actively developed mainly for fundus diseases such as fundus photography and optical coherence tomography, but not for the cataract especially using anterior segment images. We developed a user-friendly handheld slit-lamp medical device that is able to diagnose cataracts as a conventional slit-lamp microscope. Moreover, this device is a smartphone attachment so that it can record and send anterior-segment videos. We collected the anterior-segment videos and developed cataract diagnostic AI algorithms in order to use them as a screening tool for cataracts.
Methods :
This retrospective study is conducted with the approval of the Minami Aoyama Eye Clinic Ethical Review Committee (Approval No. 202101). We gathered retrospective anterior segment datasets of 835 retrospective cases, and 141,326 frames using a hand-held slit-lamp microscope (Smart Eye Camera; OUI Inc. Tokyo Japan).
After the pre-process which eliminates insufficient images, the extracted images were evaluated for nuclear cataract severity classification by multiple ophthalmologists (Annotation by WHO classification). The annotated data were divided into training (9,618 frames) and test datasets (3,931 frames), and the training dataset was used for machine learning. The test dataset was used for evaluation.
Results :
The nuclear cataract diagnosis algorithm, developed by machine learning, was analyzed on the test dataset and resulted in an accuracy of 92.9% (95% Confidence Intervals; CI 88.2-95.6%), a sensitivity of 95.8%(95%CI 91.5-98.1%), and specificity of 89.3% (95%CI 83.9-92.3%) for the diagnosis of nuclear cataract.
Conclusions :
The AI-based, nuclear cataract diagnosis algorithm successfully diagnosed cataracts with high performance compared to ophthalmologist diagnosis in Non-mydriasis eyes. With its high sensitivity, it may use in a screening where the ophthalmologist did not exist.
This abstract was presented at the 2023 ARVO Annual Meeting, held in New Orleans, LA, April 23-27, 2023.