DDLSNet: A Novel Deep Learning-Based System for Grading Funduscopic Images for Glaucomatous Damage

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Glaucoma assessment relies on manual grading of optic disc photographs (ODPs).
• Accurate estimation of optic disc size and rim width is crucial for glaucoma diagnosis.
• Automating this process can improve efficiency and consistency in clinical practice.

Purpose of the Study:

• To develop and evaluate DDLSNet, an automated image analysis pipeline for estimating the disc damage likelihood scale (DDLS).
• To assess the performance of DDLSNet's components: RimNet for rim segmentation and DiscNet for disc size classification.

Main Methods:

• DDLSNet integrates RimNet (InceptionV3/LinkNet) and DiscNet (VGG19) for ODP analysis.
• Datasets included 1208 ODPs for RimNet and 11,536 for DiscNet, with 120 ODPs for performance evaluation.
• Ground truth for rim width and disc size was established using clinician markings and OCT images, respectively.

Main Results:

• RimNet achieved low mean absolute errors for rim width estimation (mRDR: 0.04, ARW: 48.9 degrees).
• DiscNet achieved 73% classification accuracy for optic disc size.
• DDLSNet demonstrated moderate agreement (weighted kappa: 0.54) with clinician grading, comparable to interclinician agreement (0.52).

Conclusions:

• DDLSNet achieves moderate agreement with clinicians for automated DDLS grading.
• This automated pipeline shows promise for assessing glaucoma severity using ODPs.
• Future work could enhance performance by increasing sample sizes and improving clinician grading consistency.