Table 1.
Distribution of the 330 patients with IRD (full information in S1Table).
Table 2.
Distribution of the 79 patients in the non-IRD group, including 24 patients with a diagnosis other than IRD and 55 patients with a non-retinal ocular disease.
Fig 1.
Representative fundus autofluorescence (FAF, left) and color fundus (pCF, right) images illustrating variability in the dataset, including differences in intensity, magnification, text annotations, and partial occlusion by eyelashes or eyelids.
These images are intended to show the diversity of imaging conditions rather than the difficulty of classification or the underlying diagnosis.
Fig 2.
The modified EfficientNet-V2 architecture used in this study, in which the last fully connected layer (FC 1280) was replaced by two fully connected layers (FC 256 and FC 128).
The number of output neurons in the last fully connected layer was determined by the number of classes. Layers FC 256 and FC 128 are the combination of a fully connected layer, batch normalization, and ReLU activation [28].
Table 3.
Results of the three studies for the three imaging modalities. Values in bold indicate the best result; while underlined values indicate statistical significance in Studies 2 and 3 (not applicable for Study 1). Data are presented as the mean ± SD values from the five-fold cross-validation.
Fig 3.
Normalized confusion matrices for Study 1 (IRD vs. non-IRD classifiers) based on the combined FAF and pseudocolor model (a), the FAF-only model (b), and the pseudocolor-only model (c).
For each model, the columns indicate the ground truth, and the rows indicate the model’s prediction.
Fig 4.
Normalized confusion matrices for Study 2 using Group 1 (genes ABCA4, NR2E3, RDH12, TRPM1, USH2A) classifiers based on the combined FAF and pseudocolor model (a), the FAF-only model (b), and the pseudocolor-only model (c).
For each model, the columns indicate the ground truth, and the rows indicate the model’s prediction.
Fig 5.
Normalized confusion matrices for Study 3 using Group 2 (genes BEST1, CHM, CNGA3, CRB1, FAM161A) classifiers based on the combined FAF and pseudocolor model (a), the FAF-only model (b), and the pseudocolor-only model.
For each model, the columns indicate the active gene for the patient scan, and the rows indicate the model’s prediction.