Explainable Artificial Intelligence–Assisted Exploration of Clinically Significant Diabetic Retinal Neurodegeneration on OCT Images

Purpose

To explore clinically significant diabetic retinal neurodegeneration in OCT images using explainable artificial intelligence (XAI) and subsequent evaluation by retinal specialists.

Design

A single-center, retrospective, consecutive case series.

Participants

Three hundred ninety-seven eyes from 397 diabetic retinopathy patients for XAI-based screening and 244 fellow eyes for subjective human evaluation.

Methods

We acquired 30° horizontal OCT images centered on the fovea. An artificial intelligence (AI) model was developed to infer visual acuity (VA) reduction using fine-tuned RETFound-OCT. Attention maps highlighting regions contributing to VA inference were generated using layer-wise relevance propagation. Retinal specialists assessed OCT findings based on salient regions indicated by XAI. Two newly described findings, a needle-like appearance of the ganglion cell layer (GCL)/inner plexiform layer (IPL) (“ice-pick sign”) and dot-like alterations in the outer nuclear layer (ONL) (“salt-and-pepper sign”), were evaluated alongside 2 established findings: EZ disruption and choroidal hypertransmission.

Main Outcome Measures

Identification of clinically significant OCT findings associated with diabetic retinal neurodegeneration.

Results

The AI model effectively discriminated eyes with poor vision (decimal VA ≤0.5) from those with good vision (VA ≥1.0) (area under the receiver operating characteristic curve of 0.947). Explainable artificial intelligence–based attention maps highlighted salient regions in the GCL/IPL (65.2% or 70.0%), ONL (52.2% or 28.3%), EZ (39.1% or 21.7%), and choroid (26.1% or 5.00%) in eyes with poor or good vision, respectively. Subjective evaluation by retinal specialists revealed the frequencies of these 4 findings as follows: ice-pick sign (32.4%), EZ disruption (25.0%), salt-and-pepper sign (16.0%), and choroidal hypertransmission (13.5%). Eyes with decimal VA ≤0.9 had these findings more frequently than those with VA ≥1.0 (P < 0.001 for all comparisons). Salt-and-pepper sign and choroidal hypertransmission exhibited high specificity for identifying eyes with poor vision. Statistical analyses demonstrated more significant associations between EZ disruption, salt-and-pepper sign, and hypertransmission compared with their relationships with the ice-pick sign.

Conclusions

Artificial intelligence–assisted exploration of OCT findings identified 2 established lesions and 2 novel OCT biomarkers indicative of clinically significant diabetic retinal neurodegeneration.

Financial Disclosure(s)

The author(s) have no proprietary or commercial interest in any materials discussed in this article.