We read with great interest the study by Kumar et al. 1 on the detection of reticular pseudodrusen (RPD) using a deep learning (DL) model on optical coherence tomography (OCT). The authors' comprehensive validation on multiple independent external datasets is commendable. Their work represents a significant step toward the automatic detection of this key high-risk phenotype in age-related macular degeneration (AMD). However, we wish to highlight three points to further evaluate the robustness and clinical applicability of the model. First, the reliance on a “single observer” for the entire training dataset (9800 B-scans) raises concerns regarding potential systematic bias. In medical image analysis, interobserver variability is substantial for subtle radiographic features such as RPD (intergrader Dice similarity coefficient DSC = 0.68 in this study). Consequently, relying on a single expert's criteria could systematically introduce subjective judgements into the model training, potentially codifying grader-specific biases as demonstrated in other retinal deep learning applications 2. The finding that the model-grader agreement (DSC = 0.76) exceeded the intergrader agreement is noteworthy but warrants caution: does this reflect superior objective accuracy, or does it suggest the model has successfully “overfitted” to the personal interpretation style of that specific expert? Future iterations should prioritise consensus grading from multiple experts to mitigate annotation noise and ensure the model learns generalised pathological features. Second, while the authors suggest the model holds potential for quantifying RPD extent to monitor disease progression, the external validation was limited to qualitative detection (presence/absence). Given that the clinical significance of RPD is closely tied to its area expansion and association with geographic atrophy growth 3, validating the segmentation performance in external cohorts is essential. We suggest the authors, if data permit, perform a supplementary analysis reporting segmentation metrics (e.g., Dice scores) on the external validation sets. This would provide stronger evidence for the model's utility in longitudinal monitoring. Finally, the model was developed and validated exclusively using Heidelberg Spectralis images. In real-world clinical settings, variability across devices (e.g., Zeiss or Topcon) can lead to performance degradation due to domain shift 4, a well-established vulnerability of medical AI models trained on single-source data. Therefore, the model's generalisability to other common imaging platforms remains to be established. Future research utilising multi-vendor data would significantly enhance its applicability across diverse clinical scenarios. In summary, this study provides a foundational tool for automated RPD detection. Addressing these points regarding ground truth consensus, quantitative validation, and cross-vendor universality will further solidify its role in the clinical management of AMD. The authors have nothing to report. The authors declare no conflicts of interest. Data sharing not applicable to this article as no datasets were generated or analysed during the current study.
Kuang et al. (Mon,) studied this question.