Validation of a Rule-Based Automated Method for RPE Cell Detection Using Adaptive Optics Transscleral Flood Illumination Authors: Masakazu Hiraoka, Marika Taniguchi, Leandre Fura, Shuhei Kimura, Michiko Mandai, Masayo Takahashi, Yasuo Kurimoto, Akiko Maeda Published: 2026 ABSTRACT PurposeThe purpose of this study was to validate a rule-based automated method for detecting retinal pigment epithelium (RPE) cells in adaptive optics (AOs) transscleral flood illumination (AO-TFI) images by comparison with manual annotation and inter-grader agreement.MethodsThis cross-sectional study analyzed AO-TFI images from six eyes with retinitis pigmentosa (RP) and five healthy eyes. Regions of interest (ROIs) meeting predefined image quality criteria were selected. After background correction and contrast inversion, RPE cells were detected using a rule-based approach based on local maxima detection and distance filtering. Detection parameters were optimized using 30 ROIs and applied without modification to an independent validation set of 39 ROIs. Automated detections were compared with manual annotations (ground truth [GT]) using nearest-neighbor matching, and Precision, Recall, and F1-score were calculated. Inter-grader agreement between two graders (GT1 versus GT2) was assessed.ResultsAutomated detection showed stable performance across ROIs. F1-scores ranged from 0.55 to 0.89 in healthy eyes and from 0.70 to 0.92 in RP eyes, with a micro-averaged F1-score of 0.80. In the linear mixed-effects model (LMM), the F1-score of automated detection was significantly lower than inter-grader agreement, and this difference was consistently observed regardless of disease status. F1-scores remained stable across the analyzed signal-to-noise ratio (SNR) range.ConclusionsThe rule-based automated detection method applied to AO-TFI images showed lower performance than inter-grader agreement in both healthy and RP eyes, while demonstrating stable detection performance regardless of disease status. This approach represents a proof-of-concept for the feasibility of quantitative analysis of RPE cells using AO-TFI. Further validation using larger cohorts and independent datasets is warranted.Translational RelevanceThe automated RPE cell detection method based on AO-TFI evaluated in this study represents an initial step toward standardized quantitative assessment of RPE. With further validation, it may contribute to the development of imaging biomarkers and their application in clinical research. Introduction The advent of adaptive optics (AO) technology has enabled ophthalmic examinations at the cellular level.1 AOs is a technique that compensates in real time for wavefront aberrations generated as light passes through the atmosphere or ocular media, thereby improving optical resolution toward the diffraction limit.1 Since its application to ophthalmic imaging devices, AO technology has made it possible to visualize various retinal structures with cellular-level resolution.1 AOs transscleral flood illumination (AO-TFI) enables visualization of retinal pigment epithelium (RPE) cells by illuminating the retina through the sclera rather than through the pupil.2 The contrast of RPE in AO-TFI is considered to depend primarily on light absorption by melanin, with RPE cells appearing relatively hyporeflective.2 Compared with conventional AO-based infrared autofluorescence imaging, AO-TFI allows shorter acquisition times and provides a wider field of view, enabling coverage of larger retinal areas.2 These characteristics facilitate image acquisition even in diseased eyes, particularly in advanced stages, and highlight the potential of AO-TFI–derived RPE metrics as novel in vivo biomarkers for clinical applications.3,4 Despite the high spatial resolution of AO images, their analysis often involves substantial subjectivity and can be influenced by the examiner's experience and technical expertise.5,6 In addition, manual cell counting requires considerable time and effort, making it impractical for large-scale or longitudinal studies.5,6 Therefore, the development of objective and reproducible automated analysis methods is essential for advancing both clinical and basic research.7–10 Standardization of analysis methodologies is a critical foundation for inter-institutional comparisons and for future evaluation of therapeutic effects.7 In photoreceptor analyses using AO imaging, including AO scanning laser ophthalmoscopy (AO-SLO), reproducibility assessments and automated analysis approaches have been actively developed, and efforts toward standardized quantitative evaluation have been reported.11–13 In contrast, for RPE cell analysis using AO-TFI, systematic validation studies assessing the performance of analysis tools remain limited, to the best of our knowledge. In this study, we developed a rule-based automated detection method for RPE cells in AO-TFI images and performed quantitative validation against manual annotations in both healthy eyes and eyes with retinitis pigmentosa (RP). Methods Study Design and Ethical Approval This cross-sectional study was conducted at the Kobe Eye Center Hospital between November 15, 2024, and September 15, 2025. The study adhered to the tenets of the Declaration of Helsinki, and written informed consent was obtained from all participants. For participants who were minors, written informed consent was obtained from their legal guardians. The study protocol was approved by the Institutional Review Board of the Kobe Eye Center Hospital (approval number: E23004). Patients Six patients clinically diagnosed with RP at the Kobe Eye Center Hospital were enrolled in this study. The diagnosis of RP was based on a comprehensive evaluation of clinical symptoms, including bilateral visual impairment, night blindness, and visual field constriction, in combination with characteristic fundus findings such as attenuation of retinal vessels, coarse retinal pigmentary changes, bone spicule-like pigmentation, white lesions, optic disc pallor, and macular degeneration. In addition, reduced rod and cone responses were confirmed by full-field electroretinography (ERG; LE-4000; Tomey, Nagoya, Japan), supporting the clinical diagnosis of RP. In all cases, the diagnosis was independently confirmed by multiple board-certified ophthalmologists based on concordant clinical and functional findings. As a control group, five healthy volunteers with no history of ocular disease were enrolled, and AO images were acquired using the same imaging protocol as that applied to the RP group. Axial length (AL) was measured in all participants using the IOLMaster 700 (Carl Zeiss Meditec AG, Hennigsdorf, Germany). Exclusion criteria included poor visual acuity, unstable fixation, advanced cataract, cystoid macular edema, and vitreous opacity, as these conditions could compromise image quality. Image Acquisition AO images were acquired using a flood-illumination adaptive optics retinal camera (rtx1; Imagine Eyes, Orsay, France) equipped with a TFI system to visualize RPE cells.2 The principles of the imaging system and TFI modality have been described in detail previously. For each subject, the eye with better visual acuity was selected for analysis, considering the impact of fixation stability on image quality. AO correction was enabled during image acquisition, and multiple frames were recorded and averaged. Participants were instructed to maintain fixation using an internal fixation target to minimize eye movements. Illumination conditions and acquisition parameters were kept constant across all participants. The imaging field of view corresponded to 1200 × 1200 µm, assuming an AL of 24.0 mm,14 and pixel size was converted to micrometer units based on device calibration and individual AL correction.11,12 Quality Control of Image Analysis Before analysis, image quality was quantitatively assessed at the ROI level using square ROIs measuring 100 × 100 pixels. In AOs imaging, image quality metrics such as signal-to-noise ratio (SNR) and contrast have been reported to influence cellular detection accuracy and quantitative measurements.13,15 SNR was defined as the ratio of the mean to the standard deviation (SD) of pixel intensities within each ROI. Only ROIs with an SNR ≥3.5 were included in the analysis. ROI selection was based solely on this objective image quality criterion, without intentional selection according to cellular distribution or signal intensity patterns. The ROI coordinates defined during quality control were reproduced at the pixel level in all subsequent automated detection analyses. Image Preprocessing and Automated Detection of RPE Cells AO-TFI images were analyzed using ImageJ software. The original images were duplicated, and all subsequent processing steps were applied under identical conditions. Analysis was restricted to RPE cells within the predefined ROIs, which were reproduced at the pixel level based on the quality control output. Pixel size (µm/pixel) was calculated using device calibration data with AL correction and was used for the ROI area and cell density calculations. ROI images were converted to 8-bit format and subjected to background subtraction using a rolling ball algorithm (radius = 25 µm). Gaussian blur was then applied to reduce local noise, with σ defined as 7.5 µm × (px/µm) × 0.22, where 7.5 µm was used as an empirical spatial scale parameter rather than the true RPE cell diameter, with a minimum value of 1.2 pixels.16 To detect RPE cells as local maxima, the images were contrast-inverted, and candidate points were extracted using the Find Maxima function in ImageJ software. This approach represents a rule-based detection method relying on image quality and contrast rather than learning-based algorithms. The prominence threshold for local maxima detection was set proportional to the SD of pixel intensities within each ROI, with an initial value of SD × 0.04 and a lower bound of SD × 0.016. To reduce over-detection, non-maximum suppression based on nearest-neighbor distance was applied, and when the distance between candidate points was less than 4.1 µm, one of the points was removed. This threshold corresponds to approximately ⅓ of the RPE cell diameter (approximately 12–14 µm) and was selected as a practical criterion to allow minor localization errors of cell centers while preventing duplicate detection of adjacent cells. Cell density (cells/mm²) was calculated from the final number of detected points and the ROI area. To ensure independence between parameter optimization and performance evaluation, ROIs were divided into separate sets for threshold determination and validation. Detection parameters were optimized using a tuning set of 30 ROIs with representative image quality. The finalized parameters were then applied without modification to a spatially distinct set of 39 ROIs for validation, where agreement between automated detection and manual annotation was assessed (Supplementary Table S1). Although no identical ROI was included in both datasets, some ROIs were derived from the same eye, which was considered a limitation of this study. During validation, identical detection parameters were applied to all ROIs. Manual Annotation and Ground Truth Definition To evaluate the performance of automated RPE cell detection, manual annotations were used to establish the ground truth (GT). Manual counting was performed in ImageJ software using the multipoint tool within the same 100 × 100-pixel ROIs used for automated analysis. Annotations were conducted on background-corrected images while referencing both contrast-inverted and non-inverted images. Inverted images facilitated identification of cell centers by enhancing RPE cells as high-intensity points, whereas non-inverted images aided visualization of intracellular pigment boundaries; both representations were used complementarily for annotation. GT annotations were generated independently by two graders. For the primary analysis, annotations from the first grader (GT1) were used as the reference standard. Annotations from the second grader (GT2) were obtained for 20 ROIs to assess inter-grader reproducibility. Validation Metrics and Performance Evaluation Automated detection performance was evaluated by comparing detection results with manual GT on an ROI-by-ROI basis. Evaluation was based on point-wise matching between detected RPE cells and GT annotations. Correspondence between automated detections (AUTO) and GT points was determined using a nearest-neighbor approach. For each AUTO point, the Euclidean distance to the closest GT point was calculated. AUTO points located within a distance of 4 µm from a GT point were classified as true positives (TPs), whereas those exceeding this threshold were classified as false positives (FPs). This threshold corresponds to approximately ⅓ of the RPE cell diameter (approximately 12–14 µm) and was defined as a tolerance range that accounts for image noise, contrast variability, and localization errors associated with peak detection. In previous studies of photoreceptor detection, relative distance thresholds based on cell spacing (e.g. 0.75 × spacing) have been used; the fixed threshold of 4 µm used in this study is comparable in scale to these relative criteria.17 Accordingly, this threshold was considered a practical criterion that suppresses excessive matching while allowing for realistic positional deviations. False negatives (FNs) were defined as GT points without a corresponding AUTO point within the predefined distance threshold. Nearest-neighbor searches (k = 1) and distance calculations were performed using the FNN package in R software. Detection performance was quantified using Precision, Recall, and F1-score,18 calculated as follows: Precision=TP/(TP+FP)Recall=TP/(TP+FN)F1-score=2×Precision×Recall/(Precision+Recall) These metrics were computed for each ROI. Validation analyses were conducted at the ROI level, and Precision, Recall, and F1-score were summarized by subject and stratified by condition (healthy and RP), with distributions and mean values reported. To account for the correlation arising from multiple ROIs obtained from the same eye, bootstrap resampling was performed at the eye level to estimate 95% confidence intervals for Precision, Recall, and F1-score. Statistical Analysis Performance metrics (Precision, Recall, and F1-score) were calculated on an ROI-by-ROI basis and summarized using descriptive statistics across all ROIs. Metrics were additionally aggregated by condition (healthy versus RP) to descriptively compare performance trends between groups. Micro-averaged performance metrics were calculated by summing TP, FP, and FN across all ROIs prior to metric computation. To evaluate cell density derived from automated detection and manual annotation, cell density (cells/mm²) was calculated for each ROI and compared between methods. Density was computed based on ROI area after scale correction using AL. The relationship between automated detection and manual measurement was assessed using Pearson's correlation coefficient, and Bland–Altman analysis was additionally performed to evaluate the distribution of differences and the limits of agreement. A linear mixed-effects model (LMM) was used to compare F1-scores between groups. To account for the non-independence of multiple ROIs measured within the same subject, the eye was included as a random effect. Disease condition (healthy versus RP) and comparison type (AUTO versus GT1 and GT1 versus GT2) were included as fixed effects. Interaction terms were also evaluated as appropriate. All statistical analyses and metric calculations were performed using R software (RStudio version 2024.04.2+764), and image analysis and manual annotation were conducted in ImageJ software. A significance level of P < 0.05 was used for all statistical tests. Results Patient Characteristics Baseline demographic and clinical characteristics of the study participants are summarized in Table 1. A total of 11 eyes were included in the analysis, comprising 6 eyes from patients with RP and 5 eyes from healthy controls. All analyzed eyes met the predefined image quality criteria and were therefore included in the subsequent analyses. The overall automated detection and validation workflow is shown in Figure 1. Representative preprocessing steps and annotation examples are shown in Figure 2. Table 1.Patient CharacteristicsPatientAgeSexConditionAxial LengthLogMARLaterality137MRP25.120.05OD218MRP23.04−0.08OD341FRP24.86−0.08OS430FRP22.98−0.08OD546FRP24.25−0.18OD649FRP23.30OD728MHealthy24.95−0.08OD827MHealthy23.85−0.08OD934MHealthy24.46−0.18OD1030FHealthy25.44−0.08OD1130FHealthy25.86−0.08ODRP, retinitis pigmentosa.Visual acuity is presented as logarithm of the minimum angle of resolution (LogMAR); Laterality indicates the eye analyzed (OD = right eye and OS = left eye). Patient Characteristics RP, retinitis pigmentosa. Visual acuity is presented as logarithm of the minimum angle of resolution (LogMAR); Laterality indicates the eye analyzed (OD = right eye and OS = left eye). Inter-Grader Agreement Between Two Human Examiners Inter-grader agreement between two independent human examiners was evaluated by comparing manual annotations from GT1 and GT2. F1-scores ranged from approximately 0.80 to 0.94 across ROIs, indicating a high level of agreement between graders (Fig. 3). The 95% confidence intervals estimated by bootstrap resampling were 0.782 to 0.849 for Precision, 0.922 to 0.959 for Recall, and 0.858 to 0.891 for F1-score. Recall values were consistently high, reaching approximately 0.90 in most ROIs, suggesting minimal FN detections in the identification of RPE cell locations. In contrast, Precision showed mild variability across ROIs, which was primarily attributable to differences in annotation density between graders, reflecting more conservative or more inclusive marking strategies. However, these differences did not substantially affect overall F1-scores, suggesting that manual annotation is an appropriate and reproducible reference standard for validation. Validation of Automated RPE Cell Detection Against Manual Ground Truth Performance of the automated detection algorithm was evaluated by comparing automated detections with manual GT annotations on an ROI-by-ROI basis (1Fig. 2). Using micro-averaged metrics across all ROIs, the automated method achieved a Precision of 0.787, Recall of 0.821, and an F1-score of 0.804, demonstrating stable detection performance (Table 2; Fig. 3). The 95% confidence intervals estimated by bootstrap resampling were 0.760 to 0.816 for Precision, 0.765 to 0.860 for Recall, and 0.769 to 0.831 for F1-score. Overview of the automated detection and validation pipeline for AO-TFI images. AO-TFI images underwent region-of-interest (ROI) selection and quality control (100 × 100 pixels; SNR ≥3.5), followed by preprocessing including background correction and contrast inversion. RPE cells were automatically detected using a rule-based approach. Detection parameters were optimized using 30 ROIs and subsequently applied without modification to an independent validation set of 39 ROIs. Manual ground truth (GT) annotations were independently generated by two graders, and agreement between automated detection and manual annotation was evaluated using Precision, Recall, and F1-score. Representative preprocessing steps and annotation examples for automated detection of RPE cells. Representative AO-TFI images showing (A) the raw image, (B) background-corrected image, and (C) contrast-inverted image used for cell detection. RPE cells were identified as local maxima in the inverted image. Bottom panels show the same ROI with (D) manual ground truth (GT) annotations and (E) automated detection results overlaid. Manual annotations were performed with reference to both inverted and non-inverted images to facilitate identification of cell centers, whereas automated detection was applied using the same preprocessing pipeline. Table 2.Micro-Averaged Validation PerformanceTPFPFNPrecisionRecallF111113012420.7870.8210.804FN, false negative; FP, false positive; TP, true positive.Micro-averaged metrics were calculated by pooling all ROIs. Micro-Averaged Validation Performance FN, false negative; FP, false positive; TP, true positive. Micro-averaged metrics were calculated by pooling all ROIs. Agreement between automated detection and manual annotation. F1-scores for automated detection versus manual ground truth (AUTO versus GT) and inter-grader agreement between two independent graders (GT1 versus GT2) are shown for healthy and RP eyes. Each dot represents an individual ROI. Boxplots indicate the median and interquartile range. At the ROI level, F1-scores ranged from 0.55 to 0.89 in healthy eyes, with most ROIs exceeding 0.75. Similarly, in RP eyes, F1-scores ranged from 0.70 to 0.92, indicating no substantial difference in detection performance. Condition-stratified micro-averaged analysis showed similar Precision between healthy and RP eyes, whereas Recall was slightly higher in RP eyes (Table 3). Table 3.Micro-Averaged Validation Performance Stratified by ConditionConditionTPFPFNPrecisionRecallF1Healthy4501211190.7880.7910.789RP6611801230.7860.8430.814Metrics were calculated using micro-averaging within each condition. Micro-Averaged Validation Performance Stratified by Condition Metrics were calculated using micro-averaging within each condition. Recall remained relatively high in both groups, exceeding 0.80 in most ROIs, suggesting that missed detections of RPE cells were limited. In contrast, Precision exhibited greater variability across ROIs, which was primarily driven by increases in FP detections rather than reductions in TP counts. Comparison of F1 Score Between Automated Detection and Inter-Grader Agreement Using Linear Mixed-Effects Model To account for intra-subject correlation among ROIs, LMM with the eye as a random effect was used to compare F1-scores (Table 4). No significant difference in F1-score was observed between healthy and RP eyes (β = 0.034, P = 0.337). Inter-grader agreement (GT1 versus GT2) showed significantly higher F1-scores than automated detection (AUTO versus GT1; β = 0.076, 95% confidence interval 0.045–0.106, P < 0.001). The interaction between comparison type and disease condition was not significant (P = 0.597), indicating that this difference was consistently observed across both healthy and RP eyes. Table 4.Linear Mixed-Effects Model Analysis of F1 Score Comparing Automated Detection and Manual AnnotationVariablesEstimate (β)95% CI P ValueCondition (RP vs. healthy)0.034–0.337Comparison (GT vs. AUTO)0.0760.045–0.106< 0.001Interaction––0.597AUTO, automated detection; CI, confidence interval; GT, manual annotation.Estimates (β) were derived from a linear mixed-effects model with subject as a random effect. Linear Mixed-Effects Model Analysis of F1 Score Comparing Automated Detection and Manual Annotation AUTO, automated detection; CI, confidence interval; GT, manual annotation. Estimates (β) were derived from a linear mixed-effects model with subject as a random effect. Agreement of RPE Cell Density Between Automated Detection and Manual Annotation To further evaluate the quantitative validity of automated detection, RPE cell density (cells/mm²) was calculated for each ROI based on AUTO and GT1, and the two methods were compared. In the validation ROIs (n = 39), the mean density was 5246.3 ± 748.6 cells/mm² for GT1 and 5507.9 ± 816.7 cells/mm² for AUTO. A moderate but significant correlation was observed between the two methods (r = 0.456, P = 0.0035; Fig. 4). Bland–Altman analysis showed a bias of 261.6 cells/mm², indicating that AUTO tended to yield slightly higher values than GT1 (Fig. 5). The limits of agreement ranged from −1342.4 to 1865.7 cells/mm². Although some variability was observed across ROIs, no substantial systematic bias was detected. Comparison of RPE cell density between automated detection and manual annotation. Each point represents the RPE cell density for an individual ROI. The x-axis indicates the density based on manual annotation (GT1), and the y-axis indicates the density based on automated detection (AUTO; cells/mm²). The solid line represents the linear regression line, and the dashed line indicates the line of identity (y = x). Filled circles represent healthy eyes, and open circles represent RP eyes. The Pearson correlation coefficient (r = 0.456) and P value (P = 0.0035) for all ROIs are shown in the figure. Agreement of RPE cell density between automated detection and manual annotation assessed by Bland–Altman analysis. Agreement between RPE cell densities obtained by automated detection (AUTO) and manual annotation (GT1) was evaluated using a Bland–Altman plot. The x-axis represents the mean density of the two methods (cells/mm²), and the y-axis represents the difference in density (AUTO − GT1). The solid line indicates the mean difference (bias), and the dashed lines indicate the 95% limits of agreement. Filled circles represent healthy eyes, and open circles represent RP eyes. The mean difference was 261.6 cells/mm², with 95% limits of agreement ranging from −1342.4 to 1865.7 cells/mm². These results suggest that RPE cell density estimated by AUTO showed a similar trend to GT1, whereas variability was present at the ROI level. Relationship Between Image Quality and Detection Performance SNR, calculated as an image quality metric for each ROI, ranged from approximately 3.5 to 5.2 across all analyzed ROIs, satisfying the predefined inclusion criteria. Lower SNR values were associated with a tendency toward reduced Precision, whereas Recall remained relatively preserved. As a result, overall F1-scores remained stable across the analyzed SNR range (Fig. 6). Relationship between signal-to-noise ratio (SNR) and F1-score. Scatter plot showing ROI-level F1-scores as a function of SNR. Filled circles indicate healthy eyes and open circles indicate RP eyes. The dashed vertical line represents the predefined SNR threshold (SNR = 3.5) for inclusion in the analysis. F1-scores remained generally stable across the analyzed SNR range in both conditions. Discussion Principal Findings In this study, we quantitatively evaluated the performance of a rule-based automated detection algorithm for RPE cells on an ROI basis. The automated method demonstrated stable performance across healthy and RP eyes, with consistent Precision, Recall, and F1 scores that were largely independent of disease status and image conditions. In contrast, F1-scores were significantly lower than inter-grader agreement, indicating that the performance of automated detection did not reach the same level of accuracy as manual annotation. These findings suggest that, although the algorithm has limitations in achieving absolute agreement, it demonstrated consistent performance across different conditions and may serve as a practical approach for quantitative analysis. In this study, a rule-based method was adopted; however, cell detection approaches using machine learning and deep learning have also been reported.19,20 Although these approaches can accommodate complex image patterns, they require large annotated datasets and are often associated with limited interpretability due to their black-box nature. In contrast, the rule-based method used in this study offers advantages in reproducibility and interpretability, as the detection process and parameter settings are explicitly defined. In situations where annotated data are limited, such as in AO-TFI imaging, a rule-based approach may represent a practical option. Therefore, it is important to select between rule-based and learning-based methods according to image quality and study objectives. Human Inter-Grader Agreement as a Reference Standard Manual annotation is often treated as a GT; however, it inherently involves inter-grader variability.17 In the present study, we quantified inter-grader agreement by comparing annotations generated independently by two graders. The resulting F1 scores were consistently high, with particularly robust Recall values, indicating minimal omission of cell locations by either grader. In contrast, mild variability was observed in Precision, which likely reflects differences in annotation density between graders. Given that human annotations themselves are subject to variability, it is important to interpret the performance of automated algorithms relative to human inter-grader agreement rather than against an assumed absolute GT. In this study, although the F1-score of automated detection was significantly lower than inter-grader agreement, it demonstrated stable performance and may represent a useful semi-automated approach for quantitative analysis. Interpretation of Precision–Recall Characteristics A detailed examination of performance metrics revealed that Recall remained relatively high in both healthy and RP eyes, exceeding 0.80 in the majority of ROIs. This reflects the design philosophy of the algorithm, which prioritizes comprehensive candidate detection and minimizes missed cells. In contrast, Precision exhibited greater variability across ROIs, primarily due to an increase in FP detections. These FPs were more frequently observed in regions with ambiguous pigment boundaries or localized contrast variations. Nevertheless, their impact on the overall F1 score was limited, and detection performance remained stable across ROIs. Agreement of RPE Cell Density Estimates In this study, RPE cell density estimated by automated detection showed generally good agreement with manual annotation. Bland–Altman analysis indicated that automated detection tended to yield slightly higher values on average; however, no substantial systematic bias was observed. This tendency toward overestimation is likely attributable to an increase in FPs, particularly in regions with low contrast or indistinct cell boundaries. Despite these discrepancies, density estimates obtained by automated detection were of a similar order of magnitude to those obtained by manual measurement, suggesting its potential utility as a practical quantitative metric. The extent to which the quantified structures in this study accurately represent RPE cells should be considered in relation to previously reported histological data. In vivo studies using AO modalities, including auto-fluorescence (AF)-based approaches, TFI, and AO-optical coherence tomography (OCT), have reported RPE cell densities in the range of approximately 5000 to 7000 cells/mm² near the fovea, with a gradual decrease toward the periphery.15,21–28 However, substantial variability exists across studies, with differences of up to approximately 3000 cells/mm² depending on the imaging modality used. Histological studies have reported RPE cell densities of approximately 4200 to 6200 cells/mm² in the foveal to macular regions, with some studies exceeding 7000 cells/mm².29–36 In the peripheral retina, densities decrease to approximately 3500 to 4500 cells/mm², and a consistent central-to-peripheral gradient has been observed. Although these histological findings are broadly consistent with in vivo AO-based measurements, direct comparison between modalities should be approached with caution due to differences in imaging principles and analysis methods. Such discrepancies are likely attributable to variations in contrast formation mechanisms and cell detection algorithms across modalities. Therefore, in the evaluation of RPE cell density, emphasis should be placed on relative changes within the same modality rather than direct comparison of absolute values across different methods. Robustness Against Image Quality Variations SNR, used as an indicator of image quality, ranged from approximately 3.5 to 5.2 across all analyzed ROIs, satisfying the predefined inclusion criteria. Although lower SNR values were associated with reduced Precision, Recall was largely preserved, resulting in overall stable F1-scores. These findings suggest that the proposed automated detection algorithm is robust to moderate variations in image quality and further indicate that SNR may serve as a practical quality control metric for assessing both analyzability and detection performance in AO-TFI images. Performance in Diseased Retina In RP eyes, disruption of the RPE mosaic architecture and localized reductions in image contrast are commonly observed.37 In this study, detection performance in RP eyes was maintained at a level similar to that in healthy eyes. Statistical comparison of F1-scores using LMM, accounting for intra-subject correlation among ROIs, showed no significant difference between healthy and RP eyes. These findings suggest that the algorithm is applicable to diseased retinas and exhibits a degree of robustness to disease-related structural changes. Role of Image Preprocessing and Manual Annotation Strategy In this study, manual annotations were performed using both background-corrected images and contrast-inverted images. In particular, the inverted images facilitated clearer visualization of RPE cell centers, which likely contributed to improved accuracy in cell localization. By applying the same preprocessing steps to the images used for automated detection, we ensured consistent visual conditions between manual and automated analyses. This strategy enabled a fair performance comparison and represents an important methodological feature of the present study. Limitations of the Present Study Several limitations should be acknowledged. First, the number of eyes and ROIs analyzed in this study was relatively limited. Second, severely degraded images that did not meet the predefined SNR threshold were excluded from the analysis. Third, in this study, different ROIs derived from the same eye were included in both the tuning and validation datasets during parameter optimization and evaluation. Although LMM was used to account for intra-subject correlation among ROIs, this approach does not fully eliminate the potential information sharing between tuning and validation sets. Future Directions and Clinical Implications Future studies using larger datasets are warranted to further assess the generalizability of the proposed algorithm. The automatically detected cell coordinates may enable downstream analyses, including cell density estimation, spatial distribution analysis, longitudinal assessment of disease progression, and evaluation of therapeutic effects. In addition, application of this approach to other AOs imaging modalities and to different inherited retinal diseases represents an important area for future investigation. Conclusions In conclusion, we developed and validated a rule-based automated method for detecting RPE cells in AO-TFI images by comparison with manual annotations. Automated detection showed stable performance in both healthy and RP eyes. Its level of agreement was slightly lower than inter-grader agreement and did not reach the same level of accuracy as manual annotation. In contrast, the overall trend of RPE cell density was consistent with manual measurements, suggesting that this method may be applicable as a supportive approach for quantitative evaluation. Supplementary Material Supplement 1