Deep learning model for detecting cystoid fluid collections on optical coherence tomography in X-linked retinoschisis patients

Abstract

Purpose: To validate a deep learning (DL) framework for detecting and quantifying cystoid fluid collections (CFC) on spectral-domain optical coherence tomography (SD-OCT) in X-linked retinoschisis (XLRS) patients.

Methods: A no-new-U-Net model was trained using 112 OCT volumes from the RETOUCH challenge (70 for training and 42 for internal testing). External validation involved 37 SD-OCT scans from 20 XLRS patients, including 20 randomly sampled B-scans and 17 manually selected central B-scans. Three graders manually delineated the CFC on these B-scans in this external test set. The model's efficacy was evaluated using Dice and intraclass correlation coefficient (ICC) scores, assessed exclusively on the test set comprising B-scans from XLRS patients.

Results: For the randomly sampled B-scans, the model achieved a mean Dice score of 0.886 (±0.010), compared to 0.912 (±0.014) for the observers. For the manually selected central B-scans, the Dice scores were 0.936 (±0.012) for the model and 0.946 (±0.012) for the graders. ICC scores between the model and reference were 0.945 (±0.014) for the randomly selected and 0.964 (±0.011) for the manually selected B-scans. Among the graders, ICC scores were 0.979 (±0.008) and 0.981 (±0.011), respectively.

Conclusions: Our validated DL model accurately segments and quantifies CFC on SD-OCT in XLRS, paving the way for reliable monitoring of structural changes. However, systematic overestimation by the DL model was observed, highlighting a key limitation for future refinement.

Keywords: AI in ophthalmology; X‐linked juvenile retinoschisis; cystoid macular oedema; deep learning; deep reinforcement learning; image segmentation; optical coherence tomography.

FIGURE 1

Example of segmentation of cystoid fluid collections (CFC) from the external test set by the model and the reference. (a) The raw original horizontal B‐scan of an XLRS patient showing CFC. (b) Automated segmentation by the deep learning model. (c) Inconsistent areas of delineation between the observers. (d–f) Manual segmentation of the three graders.

FIGURE 2

Scatter plot of the model area (denoted with AI area) and reference area. Each data point represents the area segmented by the independent graders for each B‐scan, with those from the Gaussian distribution shown on the left and those manually selected on the right. The solid black line represents a correlation coefficient of 1.

FIGURE 3

Bland–Altman plots for cystoid fluid collections quantification for the randomly sampled B‐scans from the Gaussian distribution (left) and the manually selected B‐scans (right). The differences are calculated as reference minus model.

FIGURE 4

Difference in metric scores between the model and observer for cystoid fluid collections. These violin plots show the distribution of the differences (model minus observer) for Dice and intraclass correlation coefficients for the randomly sampled B‐scans (left) and the manually selected B‐scans (right). Positive values suggest that the model outperforms the observers. The vertical lines represent the 95% confidence intervals derived from the bootstrapped samples.