Mizukami and colleagues from Japan have developed an artificial intelligence (AI) model for quantitative trabecular–iris angle (TIA) estimation from anterior-segment images acquired with a smartphone-attached to the slit-lamp microscope.
Narrowing of the TIA is a key anatomical feature of primary angle-closure disease and contributes to impaired aqueous humor outflow.
Eyes with crowded anterior segments characterized by shallow anterior chamber depth (ACD), thicker crystalline lenses, or an anteriorly positioned lens have reduced angular width between the iris and trabecular meshwork, predisposing them to angle closure.
In Primary Angle Closure Suspect (PACS), an occludable narrow angle, defined by at least 180° of iridotrabecular contact observed on gonioscopy, is present without elevation of intraocular pressure or evidence of optic nerve damage. Later, primary angle closure glaucoma (PACG) can develop, with optic nerve head changes, such as cupping.
Traditional methods for detecting narrow angles and angle-closure disease have significant limitations in accessibility and scalability.
Gonioscopy is inherently subjective, requiring specialized skills and proper technique, which results in notoriously low adoption rates in general clinical practice and large-scale screening programs.
Other methods such as Van Herick, and anterior segment OCT (AS-OCT) have issues of sensitivity and cost factors with specialized equipment, with respect to the latter.
AI, particularly deep learning, has emerged as a highly effective tool in medical image analysis, including ophthalmology.
Videos were acquired with a slitlamp attached smartphone using direct focal illumination, a 0.1~0.3 mm full-height slit beam angled at 40°, and were required to show both the cornea and crystalline lens in focus for at least five seconds.
Corresponding TIA measurements were obtained by AS-OCT under standardized lighting. The SEC and AS-OCT examinations were performed during the same clinical session.
After the general diagnosability filtering and patient-level dataset partitioning, frame eligibility was assessed separately for each anatomical direction.
This was followed by training a deep learning model on AS-OCT-derived parameters obtained from smartphone images, thereby facilitating the collection of large volumes of standardized data.
A paired dataset comprising slit-lamp videos and corresponding AS-OCT measurements from 2654 eyes and a ResNet50-based deep-learning model to estimate TIA500 was developed.
The predicted values showed moderate correlations and good agreement with the AS-OCT measurements.
Given the moderate correlations and observed prediction errors, the current model should be regarded as a proof-of-concept rather than a substitute for gonioscopy or AS-OCT, particularly because prediction errors may affect the classification of eyes near clinically relevant narrow-angle thresholds. Further model improvement and external validation across different populations, institutions, and imaging devices are required before its utility in clinical decision-making or population screening can be established.
REFERENCE:
Shimizu E, Tanaka K, Nishimura H, Agata N, Tanji M, Nakayama S, Khemlani RJ, Yokoiwa R, Sato S, Shiba D, Sato Y. The Use of Artificial Intelligence for Estimating Anterior Chamber Depth from Slit-Lamp Images Developed Using Anterior-Segment Optical Coherence Tomography. Bioengineering (Basel). 2024 Oct 9;11(10):1005. doi: 10.3390/bioengineering11101005. PMID: 39451381; PMCID: PMC11505230.
No comments:
Post a Comment