Performance and Reliability Assessment of Deep
Learning Models for OCT -Based Retinal Disease
Classification
Baohua Li 
Zhengzhou Seventh People's Hospital
Xinfei Chen 
Henan Academy of Innovations in Medical Science
Meng Li 
Beijing Anzhen Hospital Affiliated to Capital Medical University
Kai Wang 
Tianjin Medical University
Jingnan Li 
Zhengzhou Seventh People's Hospital
Shu Li 
Zhengzhou Seventh People's Hospital
Jie Du 
Henan Academy of Innovations in Medical Science
Research Article
Keywords: retinal diseases, class imbalance, deep learning, model reliability, image feature space
Posted Date: November 11th, 2026
DOI: https://doi.org/10.21203/rs.3.rs-11439143/v1
License:   This work is licensed under a Creative Commons Attribution 4.0 International License.  
Read Full License
Additional Declarations: No competing interests reported.

1
Performance and Reliability Assessment of Deep Learning Models 1
for OCT-Based Retinal Disease Classification 2
Baohua Li2†, Xinfei Chen1†, Meng Li3, Kai Wang4, Jingnan Li2, Shu Li2, Jie Du1* 3
4
1 Institute of Biological Therapy, Henan Academy of Innovations in Medical Science, Zhengzhou, 5
China 6
2 Department of Ophthalmology, Zhengzhou Seventh People’s Hospital, Zhengzhou, 450016, 7
China 8
3 Beijing Anzhen Hospital Affiliated to Capital Medical University, Beijing, China 9
4 Department of Physiology and Pathophysiology, School of Basic Medical Sciences, Tianjin 10
Medical University, Tianjin, China 11
† Baohua Li, Xinfei Chen contributed equally to this work. 12
13

• Correspondence: Jie Du jiedubj@126.com 14

2
Performance and Reliability Assessment of Deep Learning Models 15
for OCT-Based Retinal Disease Classification 16
17
Abstract 18
Background Deep learning models have achieved high classification accuracy on 19
OCT images for discriminating among retinal diseases. However, clinical OCT 20
datasets often exhibit substantial class imbalance, and aggregate metrics may obscure 21
unreliable predictions for underrepresented classes. The reliability of such models for 22
clinical decision support therefore remains uncertain. This study aimed to 23
comprehensively assess the reliability of a deep learning-based OCT model under 24
class imbalance. 25
Methods Ten public OCT datasets were combined for a nine-class classification task 26
of normal retina and eight retinal diseases. A Swin Transformer model was developed 27
and evaluated on an internal test set, two independent public test sets (DUKE and 28
OCTID), and a prospective clinical OCT cohort. Class-specific performance and 29
bootstrap-derived 95% confidence intervals were analyzed alongside conventional 30
metrics. Expected calibration error (ECE) assessed the reliability of predicted 31
probabilities. LayerCAM was used to examine model attention to disease-related 32
structures, and UMAP to visualized the deep feature space. 33
Results On the internal test set, the model achieved a macro-AUC of 0.9992 and 34
macro-recall of 0.9886. Despite high recall for some underrepresented classes, 35
bootstrap estimates of precision and F1-score varied more widely, underscoring the 36

3
need to consider sample size and uncertainty. Macro-F1 scores were 0.9884 on DUKE 37
and 0.8944 on OCTID; recall for the underrepresented CSC class on OCTID was 38
0.7863. Pre-calibration ECE was 0.0048, indicating good agreement between 39
predicted probabilities and observed outcomes. In the clinical cohort, macro-F1 was 40
0.8813 and macro-recall was 0.8365. LayerCAM activations corresponded spatially to 41
disease-related retinal structures in many cases. UMAP showed class-specific 42
clustering, with some overlap between classes. 43
Conclusion This study assessed the reliability of deep learning-based OCT models 44
for retinal disease discrimination under class imbalance. Overall performance alone is 45
insufficient to characterize model reliability in multi-disease clinical settings. 46
Class-specific stability, uncertainty, probability calibration, and interpretability also 47
warrant assessment. These findings support a risk-aware decision-support strategy in 48
which model outputs are interpreted together with class representation and 49
uncertainty, and uncertain or underrepresented cases are flagged for specialist review. 50
Prospective human–AI studies are needed before clinical deployment. 51
Keywords retinal diseases; class imbalance; deep learning; model reliability; image 52
feature space 53
54
Introduction 55
Retinal diseases are major causes of visual impairment, and conditions such as 56
diabetic retinopathy contribute substantially to the global disease burden [1, 2]. 57
Optical coherence tomography (OCT) is widely used in ophthalmic practice because 58

4
it is noninvasive and provides high-resolution images of retinal microstructure [3, 4]. 59
In clinical settings, OCT interpretation often requires distinguishing normal-appearing 60
retinas from several diseases, rather than identifying a single condition. As 61
examination volumes grow and the range of conditions encountered becomes more 62
complex, artificial intelligence (AI)-assisted methods that can reliably distinguish 63
among retinal diseases have become an important area of medical image research [5]. 64
Deep learning models have shown strong performance in OCT-based 65
classification of common retinal diseases, including age-related macular degeneration 66
(AMD) and diabetic retinopathy (DR) [6-8]. Clinical datasets, however, are often 67
class-imbalanced. Normal images and images of common diseases may dominate, 68
while other categories have few samples because of differences in clinical frequency, 69
data availability, or study conditions. This imbalance can affect class-specific 70
performance, bias predictions, and impair probability calibration for underrepresented 71
classes—disease categories with limited representation in the datasets [9]. 72
The clinical value of an AI model depends not only on its classification 73
performance but also on whether its confidence reflects its accuracy and whether it 74
identifies uncertain or out-of-distribution cases [10, 11]. Estimates for 75
underrepresented classes are especially sensitive to sampling variability, making any 76
single metric an incomplete measure of model behavior. A fuller evaluation should 77
therefore consider class-specific performance, uncertainty [12], probability 78
calibration, and interpretability. 79
Despite advances in OCT-based AI, systematic assessments of model reliability 80

5
under class imbalance remain limited. Overall classification metrics can obscure 81
differences between disease categories and uncertainty in class-specific estimates. The 82
reliability of predicted probabilities and variation in performance across data sources 83
also warrant closer examination [13]. Differences in OCT devices, scanning 84
protocols, and patient populations can shift data distributions and reduce performance 85
in external settings. Accordingly, evaluating model performance stability and 86
predictive reliability across multiple dimensions under conditions that closely reflect 87
real-world clinical practice is important for determining the application potential of 88
AI-assisted diagnostic systems [14]. Equally important is defining the conditions 89
under which such systems can be safely relied upon in clinical practice and when their 90
outputs should be referred for human review. 91
In this study, we evaluated the performance and reliability of a deep learning 92
OCT classifier under class imbalance. We combined ten public OCT datasets into a 93
nine-class task comprising normal-appearing retinas and eight retinal diseases, then 94
developed a model based on the Swin Transformer (Swin-T). Its hierarchical feature 95
representations and shifted-window attention capture features at multiple scales while 96
preserving local structure [15]. We evaluated the model on an internal test set, two 97
independent external public test sets, and an independent clinical OCT cohort. Its 98
stability and reliability across disease categories and data sources were assessed from 99
multiple perspectives, including class-level classification performance and associated 100
uncertainty, probability calibration, model interpretability, and distributions in the 101
deep feature space. The overall study workflow is illustrated in Fig. 1. 102

6
103
Fig. 1 Overall flowchart. The flowchart illustrates the complete workflow of this study. a) 104
multi-source OCT data integration, patient-level dataset partitioning; b) Swin-T development; c) 105
Reliability evaluation, including the nine-class confusion matrix, LayerCAM interpretability 106
analysis, and UMAP feature characterization. 107
108
Materials and methods 109
We integrated multiple publicly available OCT datasets, mapped their labels to a 110
unified disease classification scheme, and performed quality control to define a 111
nine-class task comprising normal-appearing retinas and eight retinal disease 112
categories. A Swin-T-based classifier was then developed, and class-weighted 113
sampling was used to mitigate bias arising from class imbalance in the training data. 114
The model was evaluated on an internal test set, two independent external public 115
test sets, and an independent clinical OCT cohort. In addition to conventional 116
classification metrics, we assessed prediction reliability by examining class-specific 117

7
performance stability, uncertainty estimates, probability calibration, interpretability, 118
and deep feature representations. These analyses characterized the model’s 119
performance and potential limitations across disease categories and data sources. 120
121
Datasets 122
Publicly available OCT datasets were screened against predefined criteria. Eligible 123
datasets contained clinically annotated retinal OCT images with clearly defined 124
disease labels mappable to a common classification scheme and images of sufficient 125
quality for model analysis. We excluded images or records with missing labels, 126
ambiguous classifications, corrupted files, or clear duplicates. The included datasets 127
spanned different institutions, geographic regions, and OCT device platforms, 128
supporting model development and cross-source evaluation. Dataset details are 129
provided in Supplementary Table S1. 130
Because diagnostic terminology varied across datasets, we established uniform 131
label-mapping rules before model development. Based on the original diagnoses, 132
CNV and DRUSEN were mapped to AMD, and diabetic macular edema (DME) was 133
mapped to DR. Samples that could not be mapped confidently or had uncertain 134
diagnostic definitions were excluded. The resulting nine classes were 135
normal-appearing fundi (NORMAL), AMD, DR, macular hole (MH), retinal vein 136
occlusion (RVO), central serous chorioretinopathy (CSC), epiretinal membrane 137
(ERM), rhegmatogenous retinal detachment (RRD), and retinal artery occlusion 138
(RAO). 139

8
The study included 124,120 images from ten public OCT datasets. OCTID and 140
DUKE served as independent external public test sets; the other eight datasets were 141
used for model development. Where patient identifiers were available, development 142
data were split at the patient level and then divided into training, validation, and 143
internal test sets at a 7:2:1 ratio, ensuring that images from the same patient remained 144
in a single subset. 145
To assess generalizability in clinical practice, we also evaluated the model in an 146
independent OCT cohort prospectively enrolled at Zhengzhou Seventh People’s 147
Hospital (ZZSPH). Consecutive eligible patients were recruited under a pre-approved 148
protocol. These images were collected independently of model development and were 149
not used for training, validation, or parameter optimization, thereby minimizing 150
information leakage. All clinical images underwent standardized preprocessing and 151
were screened for quality using clinically interpretable criteria. Written informed 152
consent was obtained from all participants. 153
154
Development 155
All OCT images underwent standardized preprocessing, including correction of 156
retinal structural orientation, cropping of the valid image region, proportional resizing 157
followed by padding to 256 × 448 pixels, and normalization of pixel intensities to [0, 158
1]. To address class imbalance in the training set, we used class-weighted sampling to 159
adjust each class’s sampling probability during training; weights for the 160
underrepresented ERM and RVO classes were set to 2.0. Random augmentation was 161

9
applied only during training and included geometric transformations, intensity 162
perturbations, and noise simulation. Validation and test images underwent 163
preprocessing without random augmentation. 164
We used Swin-T as the primary classifier. Its hierarchical feature extraction and 165
shifted-window self-attention capture local retinal structures and broader contextual 166
information. Preprocessed images were passed through the backbone to extract 167
multiscale features, and the classification head produced probabilities for the nine 168
disease categories. The model was initialized with ImageNet-pretrained weights and 169
fine-tuned on the OCT datasets using transfer learning. 170
For architectural comparison, we included ResNet-50, DenseNet-121, 171
ConvNeXt-Tiny, and Vision Transformer-Small (ViT) as baseline models. We also 172
evaluated a Swin-T variant without the shifted-window mechanism to assess the 173
contribution of window shifting to classification performance. All models were 174
trained with multiclass cross-entropy loss. Detailed training parameters are provided 175
in Supplementary Table S2. 176
177
Assessment 178
We first evaluated overall discriminative ability using classification metrics, then 179
systematically assessed predictive reliability through class-specific stability, 180
uncertainty, probability calibration, interpretability, and feature representation 181
analyses. 182
183

10
Classification performance assessment 184
We assessed performance using conventional classification metrics, including 185
accuracy, macro-precision, macro-recall, macro-specificity, macro-F1 score, and 186
macro-AUC. Because class sizes varied substantially, we emphasized macro-averaged 187
metrics to reduce the influence of high-frequency classes. Multiclass AUC was 188
calculated using a one-vs-rest approach. Confusion matrices were used to characterize 189
misclassification patterns across disease categories. 190
191
Stability assessment for underrepresented categories 192
Performance estimates may be unstable when a class has few test samples. To 193
quantify this uncertainty, we generated 2,000 bootstrap samples by resampling the test 194
set with replacement; each sample matched the original test set in size. We calculated 195
95% confidence intervals from the bootstrap distribution. We assessed class-specific 196
stability in relation to the number of samples in each category, the range of metric 197
variation, and confidence interval width, with particular attention to underrepresented 198
classes. 199
200
Probability calibration assessment 201
Expected calibration error (ECE) was calculated from the model output probabilities 202
to assess the agreement between predicted probabilities and observed classification 203
outcomes. The raw predicted probability for class k was calculated using the model 204
logits according to Equation 1. 205

11
 =
= C
i i
k
k
z
zp
1 ) exp(
) exp(ˆ (Equation 1) 206
The ECE was calculated according to Equation 2. Here, B denotes the number of 207
equal-width probability bins, bN denotes the number of samples in bin B , ) (b conf208
denotes the mean predicted confidence in bin b ,and ) (b accdenotes the observed 209
accuracy in binb . 210
) ( ) (
1
b conf b accNnECE Bb b − == (Equation 2) 211
Three post hoc calibration methods were compared: temperature scaling, histogram 212
binning, and isotonic regression. Temperature scaling introduced a global temperature 213
parameterT where 0＞T , to rescale the logits for all classes simultaneously. The 214
calibrated probabilities were calculated according to the corresponding 215
temperature-scaling equation. 216
 =
= C
i i
kTS
k
T z
T zp
1 ) exp(
) exp(ˆ (Equation 3) 217
The optimal temperature parameter ( *T ) was estimated by minimizing the multiclass 218
negative log-likelihood on the validation set. Here, TSkPˆ denotes the 219
temperature-scaled probability for class k ; T is the temperature parameter shared by 220
all classes; valN is the number of validation samples; and ny is the true class index of 221
the n th validation sample. 222
223
 

− =
val
n
N
n
TS
y
valT
pNT 10* ˆlog1minarg ＞ (Equation 4) 224
Histogram binning divided the probability interval  1 , 0into B contiguous bins of equal 225

12
width. The raw predicted confidence in each bin was mapped to the corresponding 226
empirical event rate estimated from the validation set. The bin boundaries and 227
mapping values were estimated using the validation set and then fixed before 228
evaluation on the test set. 229
)) , [ˆII( )ˆ(ˆ 1
1
b b
B
b
b
HB a a p y p p−

  = (Equation 5) 230
where HBpˆ denotes the histogram binning calibration function; pˆ is the raw predicted 231
confidence to be calibrated; 10=B is the number of bins; Ba a a,..., ,1 0 are the bin 232
boundaries, satisfying 00 =a and 1=Ba , with all bins of equal width; by is the true 233
positive rate of validation samples in the b bin; ) (Ⅱ is the indicator function, which 234
takes the value 1 when the condition inside the parentheses holds, and 0 otherwise. 235
Once the bin boundaries and mapping values are determined on the validation set, 236
they remain fixed. 237
Isotonic regression calibrated the predicted probabilities by learning a monotonically 238
nondecreasing mapping function. A separate mapping function was fitted for each 239
class. Each mapping function was estimated on the validation set by solving a 240
monotonic constrained least-squares problem. 241
)ˆ(ˆ k kISOk p f p= (Equation 6) 242
The mapping function kf is obtained by solving the following monotonically 243
constrained least-squares problem on the validation set: 244


− =
val
k
N
i
i k k i kfk p f y f1 2, ,* ))ˆ( ( min arg (Equation 7) 245
Here, i kp ,ˆ denotes the isotonic-regression-calibrated probability for class k , kf is the 246

13
monotonically nondecreasing mapping function for classs k ; i kp ,ˆ is the raw predicted 247
probability for class k in validation sample i , and i ky , is the corresponding one-hot 248
class label. 249
All calibration methods were fitted using the validation set only, whereas the test set 250
was reserved exclusively for final evaluation. Macro-average top- k recall was also 251
calculated to evaluate the model’s ability to rank candidate classes and provide a 252
candidate disease set for AI-assisted discrimination. 253
254
Decision interpretability assessment 255
LayerCAM was used to generate class-specific activation heatmaps and assess the 256
spatial correspondence between the model’s activated regions and disease -related 257
OCT structures. For the Swin-T, the first normalization layer, norm1, of the last block 258
in Stage 4 of the backbone was selected as the target feature layer for LayerCAM. The 259
resulting feature map had a spatial size of 8 × 14. An ophthalmologist qualitatively 260
reviewed the heatmaps to determine whether activations coincided with retinal 261
abnormalities or instead focused on background regions, scale bars, or other 262
disease-unrelated structures. 263
264
Feature representation analysis 265
To examine the distribution of image features learned for different retinal diseases, 266
the activation vector immediately preceding the prediction layer was extracted as the 267
feature embedding and projected into a two-dimensional space using Uniform 268

14
Manifold Approximation and Projection (UMAP). This analysis was used to examine 269
feature clustering patterns and latent similarities among disease categories, as well as 270
the distribution of underrepresented classes in the learned feature space. It provided 271
supporting evidence for interpreting the model’s internal representations of different 272
retinal diseases. This analysis was not treated as an independent classification 273
performance metric and was used only to visualize and interpret the learned feature 274
space. 275
276
Statistical analysis 277
Model performance evaluation and uncertainty analysis were based on the final 278
predictions generated for the test set. Bootstrap resampling was performed for 2,000 279
iterations to calculate 95% confidence intervals for the primary evaluation metrics. No 280
training-phase data were used in the test performance analyses, thereby maintaining 281
the independence of the performance estimates. 282
283
Results 284
Baseline information on the datasets 285
A total of 10 publicly available OCT datasets comprising 124,120 OCT images were 286
included in this study. The datasets covered normal-appearing fundi and eight 287
categories of retinal disease (Table 1). Sample sizes differed substantially across 288
disease categories. NORMAL, AMD, and DR accounted for most of the images, 289
whereas RAO, RRD, ERM, CSC, and RVO had relatively few samples, indicating a 290

15
typical class-imbalanced distribution. 291
The internal test set comprised 10,439 OCT images obtained through 292
patient-level splitting of the eight development datasets. It was not used for model 293
training, parameter optimization, or model selection. The two independent external 294
public test sets, DUKE and OCTID, comprised 441 and 428 OCT images, 295
respectively. An independent clinical OCT test cohort comprising 312 images was 296
also included to evaluate model performance in a real-world outpatient setting. The 297
composition and class distribution of each dataset are summarized in Supplementary 298
Table S3. 299
Table 1 Sample Distribution and Sources by Disease Category 300
Disease Sample Size (Images) Proportion (%) Primary Dataset Source
NORMAL 57,552 46.37 Eye OCT, Kermany OCT2017, OCT5k, OCTA-500, OCTDL, Duke_2014_BOE, OCTID
AMD 50,649 40.81 AMD-SD, Kermany OCT2017, OCT5k, OCTA-500, OCTDL, Duke_2014_BOE, OCTID
DR 12,836 10.34 Kermany OCT2017, OCT5k, OCTA-500, OCTDL, Duke_2014_BOE, OCTID
MH 2,568 2.07 OCTDL, OIMHS, OCTID
RVO 176 0.14 OCTA-500, OCTDL
CSC 164 0.13 OCTA-500, OCTID
ERM 85 0.07 OCTDL
RRD 69 0.06 ORC SS-OCT
RAO 21 0.02 OCTDL
Total 124,120 100 -

16
301
Model classification performance and baseline comparison 302
To evaluate the baseline discriminative ability of different deep learning architectures 303
in the nine-class OCT classification task, we compared the performance of Swin-T 304
with that of ResNet-50, DenseNet-121, ConvNeXt-Tiny, and ViT (Table 2). All 305
models were trained using the same data-splitting strategy and training procedure. All 306
models demonstrated high overall classification performance on the internal test set. 307
Swin-T achieved relatively higher macro-AUC and macro-Recall values, whereas 308
some convolutional neural network models and ViT achieved slightly higher 309
macro-F1 scores than Swin-T. The training curves for all models and the 310
validation-set performance of Swin-T are shown in Supplementary Fig. S1. 311
Considering the results across the different metrics, Swin-T was selected as the 312
representative model for subsequent reliability assessment. 313
An ablation experiment was performed to further assess the effect of the 314
shifted-window mechanism on model performance. On the OCTID external test set, 315
the complete Swin-T model achieved a higher macro-Recall than the Swin-T-no-shift 316
model, in which the shifted-window mechanism had been removed (0.8646 vs. 317
0.7381). A difference of comparable magnitude was not observed on the DUKE 318
dataset (Supplementary Table S4). These findings suggest that window-based 319
interactions may contribute to feature representation under some cross-source data 320
conditions. 321
The confusion matrix for the internal test set showed high classification 322

17
accuracy, with misclassifications occurring mainly between disease categories with 323
similar OCT imaging features (Fig. 2a). For example, some AMD samples were 324
misclassified as NORMAL or DR. ROC curve analysis indicated good discrimination 325
across all disease categories (Fig. 2b). The average inference time of Swin-T was 36.9 326
ms per OCT image. Comparisons of computational efficiency are provided in 327
Supplementary Table S5 and Supplementary Fig. S2. 328
329
Table 2 Performance comparison with deep learning models on internal test dataset 330
Model Macro‑Precision Macro‑Recall Macro‑F1 Macro‑AUC
DenseNet-121 0.7155 [0.6797‑0.7561] 0.9692 [0.9651‑0.9726] 0.7791 [0.7483‑0.8102] 0.9949 [0.9942‑0.9955]
ResNet‑50 0.8075 [0.7630‑0.8585] 0.9375 [0.8999‑0.9713] 0.8527 [0.8132‑0.8909] 0.9983 [0.9979‑0.9987]
ConvNeXt 0.6640 [0.6125‑0.7429] 0.9831 [0.9799‑0.9858] 0.7474 [0.6989‑0.8091] 0.9984 [0.9980‑0.9986]
ViT 0.8031 [0.7771‑0.8346] 0.9704 [0.9446‑0.9852] 0.8509 [0.8249‑0.8759] 0.9987 [0.9984‑0.9990]
Swin‑T 0.7684 [0.7142‑0.8412] 0.9886 [0.9865‑0.9904] 0.8343 [0.7903‑0.8917] 0.9992 [0.9990‑0.9994]
331

18
332
Fig. 2 Performance and reliability of the Swin-T model on internal and independent cohorts. a) 333
Nine-class confusion matrix on the internal test set. b) One-vs-rest ROC curves (global AUC = 334
0.999). c) Reliability diagram of the raw model (ECE = 0.0048). d) Per-class ECE across the nine 335
categories. e) Macro-averaged top-k recall. f–g) External validation on DUKE: per-class 336

19
performance (f) and confusion matrix (g). h–i) External validation on OCTID: per-class 337
performance (h) and confusion matrix (i). j–k) Prospective validation on ZZSPH: per-class 338
performance (j) and confusion matrix (k). 339
340
Performance stability under class imbalance 341
Disease-Specific performance across categories 342
Although the model achieved high overall classification performance, class-level 343
analysis revealed substantial differences among disease categories (Table 3). The 344
high-sample categories, including NORMAL, AMD, and DR, showed relatively 345
stable performance, whereas the performance metrics of underrepresented classes 346
varied more substantially. In the internal test set, the numbers of samples were 347
particularly limited for RAO (n = 1), RRD (n = 4), ERM (n = 9), CSC (n = 10), and 348
RVO (n = 17). For example, the RAO category achieved a recall and AUC of 1.0000. 349
However, because only one RAO image was included in the test set, this result cannot 350
be interpreted as evidence of stable recognition performance for this category. 351
Similarly, the lower F1 scores observed for some underrepresented classes were 352
driven mainly by reduced precision, because a small number of false-positive 353
predictions can substantially alter proportion-based metrics. These findings indicate 354
that, under severe class imbalance, overall performance metrics may conceal 355
instability in underrepresented classes. Class-level results should therefore be 356
interpreted together with the corresponding sample sizes and estimation uncertainty. 357
358
359
360

20
Table 3 Per-category diagnostic performance on internal test set 361
Disease Recall Specificity Precision F1 Score AUC
NORMAL
0.9894
[0.9868‑0.9919]
0.9746
[0.9702‑0.9788]
0.9789
[0.9753‑0.9823]
0.9841
[0.9818‑0.9864]
0.9974
[0.9966‑0.9981]
AMD
0.9368
[0.9286‑0.9449]
0.9961
[0.9945‑0.9975]
0.9919
[0.9887‑0.9949]
0.9636
[0.9590‑0.9680]
0.9970
[0.9961‑0.9979]
DR
0.9779
[0.9691‑0.9865]
0.9861
[0.9836‑0.9884]
0.8859
[0.8683‑0.9032]
0.9296
[0.9189‑0.9400]
0.9982
[0.9975‑0.9988]
ERM
1.0000
[1.0000‑1.0000]
0.9997
[0.9993‑1.0000]
0.7684
[0.5625‑1.0000]
0.8650
[0.7200‑1.0000]
1.0000
[1.0000‑1.0000]
MH
0.9935
[0.9803‑1.0000]
1.0000
[1.0000‑1.0000]
1.0000
[1.0000‑1.0000]
0.9967
[0.9900‑1.0000]
1.0000
[1.0000‑1.0000]
RRD
1.0000
[1.0000‑1.0000]
0.9999
[0.9997‑1.0000]
0.8340
[0.5714‑1.0000]
0.9021
[0.7273‑1.0000]
1.0000
[1.0000‑1.0000]
RVO
1.0000
[1.0000‑1.0000]
0.9994
[0.9989‑0.9998]
0.7515
[0.6071‑0.8947]
0.8557
[0.7556‑0.9444]
1.0000
[1.0000‑1.0000]
RAO
1.0000
[1.0000‑1.0000]
0.9997
[0.9993‑1.0000]
0.3120
[0.1250‑1.0000]
0.4512
[0.2222‑1.0000]
1.0000
[1.0000‑1.0000]
CSC
1.0000
[1.0000‑1.0000]
0.9985
[0.9977‑0.9991]
0.3928
[0.2941‑0.5263]
0.5611
[0.4545‑0.6897]
1.0000
[1.0000‑1.0000]
Global
(Macro)
0.9886
[0.9865‑0.9904]
0.9949
[0.9943‑0.9955]
0.7684
[0.7142‑0.8412]
0.8343
[0.7903‑0.8917]
0.9992
[0.9990‑0.9994]
362
Uncertainty in performance estimates 363
To further quantify uncertainty in the performance estimates under class imbalance, 364
the test-set metrics were evaluated using 2,000 bootstrap resampling iterations, and 365
95% confidence intervals were calculated (Table 3). Performance estimates for 366
high-sample categories varied relatively little across bootstrap samples, whereas 367
several metrics for underrepresented classes had wider confidence intervals. For 368
example, the F1 score for RAO ranged from 0.2222 to 1.0000 within the 95% 369
confidence interval, whereas the ERM F1 score was 0.8650, with a 95% confidence 370
interval of 0.7200 to 1.0000. Some bootstrap intervals for recall or AUC were 371

21
degenerate or otherwise uninformative. For RAO, for example, the internal test set 372
contained only one image, so repeated resampling could not provide meaningful 373
statistical variation. These results should therefore not be interpreted as evidence of 374
stable or perfect recognition performance. They must be considered in the context of 375
the extremely small sample size. Taken together, these results indicate that a single 376
point estimate may be insufficient to characterize model reliability in 377
class-imbalanced OCT classification. Both confidence intervals and sample size are 378
important when interpreting performance in underrepresented classes. 379
380
Prediction probability reliability 381
In addition to classification performance, the reliability of the model’s predicted 382
probabilities is an important component of overall model reliability. We therefore 383
further evaluated the calibration of the model’s predicted probabilities. The mean 384
ECE was 0.0048 before calibration and 0.0047 after temperature scaling, indicating 385
only a small difference between the two conditions. By contrast, histogram binning 386
and isotonic regression did not further reduce the test-set ECE (Fig. 2c and 387
Supplementary Fig. S3). These findings suggest that, under the current sample size 388
and class distribution, nonparametric calibration methods may be limited by the 389
number of validation samples. Consequently, their ability to improve the reliability of 390
the predicted probabilities may be limited. 391
Class-level calibration analysis also showed differences among disease 392
categories. For high-sample categories such as NORMAL, AMD, and DR, the 393

22
probability estimates were supported by more samples, and ECE provided a more 394
stable description of the model’s probabilistic behavior. Some underrepresented 395
classes showed low ECE values, but the reliability of these estimates should still be 396
interpreted in light of the limited number of samples. Within the high-confidence 397
interval of 0.9 to 1.0, the observed accuracy for CSC was lower than the predicted 398
confidence, indicating a degree of overconfidence (Fig. 2d). The macro-average top-3 399
recall was 99.9% (Fig. 2e), indicating that the true category was included among the 400
three highest-ranked predictions for nearly all samples. This result suggests that the 401
model may provide useful candidate information for clinical discrimination. 402
403
External and clinical validation 404
To evaluate generalizability across data sources and in a real-world clinical setting, 405
the model was further evaluated on two independent external public test sets and an 406
independent clinical OCT cohort. 407
408
Validation on external public datasets 409
The DUKE dataset comprised 441 OCT images from the NORMAL, AMD, and DR 410
categories. The model maintained high and stable classification performance on this 411
dataset (Table 4 and Fig. 2f). The confusion matrix showed that only a small 412
proportion of DR cases were misclassified, accounting for 2.9% of the DR samples 413
(Fig. 2g). 414
The OCTID dataset comprised 428 OCT images from the NORMAL, AMD, DR, 415

23
MH, and CSC categories. The model showed reduced diagnostic performance for DR 416
and CSC on the OCTID dataset (Table 4 and Fig. 2h). Confusion matrix analysis 417
showed that the main errors involved DR being misclassified as RVO and CSC being 418
misclassified as RRD or AMD, which may reflect similarities in their OCT imaging 419
phenotypes (Fig. 2i). 420
Overall, the external test results indicated that the model generalized to different 421
data sources, although its performance was affected by inter-center differences in data 422
distribution. The model’s differential diagnostic performance was lower when disease 423
categories had similar imaging phenotypes. 424
Table 4 Performance on independent external validation datasets 425
Dataset Disease Precision Recall F1-score AUC
DUKE
NORMAL
0.9741
[0.9497‑0.9934]
1.0000
[1.0000‑1.0000]
0.9869
[0.9742‑0.9967]
1.0000
[1.0000‑1.0000]
AMD
1.0000
[1.0000‑1.0000]
0.9933
[0.9800‑1.0000]
0.9966
[0.9899‑1.0000]
1.0000
[1.0000‑1.0000]
DR
0.9928
[0.9779‑1.0000]
0.9712
[0.9429‑0.9929]
0.9818
[0.9635‑0.9964]
0.9999
[0.9995‑1.0000]
DUKE
Macro-average

0.9890
[0.9787‑0.9978]
0.9882
[0.9765‑0.9976]
0.9884
[0.9770‑0.9977]
0.9999
[0.9998‑1.0000]
OCTID
NORMAL
0.9598
[0.9314‑0.9845]
0.9175
[0.8786‑0.9515]
0.9380
[0.9128‑0.9606]
0.9845
[0.9720‑0.9939]
AMD
0.9262
[0.8596‑0.9815]
0.9104
[0.8364‑0.9818]
0.9176
[0.8598‑0.9643]
0.9904
[0.9773‑0.9989]
DR
0.7795
[0.7045‑0.8533]
0.8293
[0.7368‑0.9079]
0.8028
[0.7355‑0.8645]
0.9396
[0.9033‑0.9688]
MH
1.0000
[1.0000‑1.0000]
0.8797
[0.7910‑0.9552]
0.9355
[0.8833‑0.9771]
0.9970
[0.9930‑0.9995]
CSC
1.0000
[1.0000‑1.0000]
0.7863
[0.6250‑0.9167]
0.8779
[0.7692‑0.9565]
0.9747
[0.9301‑1.0000]
OCTID
Macro-average

0.9331
[0.9105‑0.9544]
0.8646
[0.8189‑0.9063]
0.8944
[0.8598‑0.9256]
0.9772
[0.9644‑0.9874]
426

24
Validation in the clinical cohort 427
To further evaluate model performance in routine clinical OCT examinations, we 428
assessed the model on an independent clinical test cohort comprising 312 images. The 429
cohort was derived from routine outpatient examinations and included NORMAL, 430
AMD, DR, and ERM images. All OCT images were independently reviewed and 431
annotated by two ophthalmologists. Cases with discrepant interpretations were 432
adjudicated by a third senior ophthalmologist, and the consensus diagnosis was used 433
as the reference standard. In the clinical cohort, the model achieved a macro-F1 score 434
of 0.8813 and a macro-Recall of 0.8365 (Table 5 and Fig. 2j). Performance was most 435
stable for the NORMAL category, whereas some misclassification occurred among 436
AMD, DR, and ERM. The confusion matrix showed that errors in the clinical cohort 437
occurred mainly between disease categories rather than as random classification errors 438
(Fig. 2k). These findings suggest that the model retained disease-related classification 439
ability in the real-world clinical setting, although its performance was affected by 440
differences in disease phenotype and data distribution. 441
442
443
444
445
446
447
448

25
Table 5 Diagnostic performance of Swin-T model on the prospective ZZSPH clinical cohort 449
Disease Precision Recall F1-score AUC Support
NORMAL 0.9892 [0.9779‑1.0000] 0.9890 [0.9742‑1.0000] 0.9891 [0.9797‑0.9963] 0.9909 [0.9754‑0.9996] 271
AMD 0.9498 [0.8500‑1.0000] 0.8592 [0.7143‑1.0000] 0.9001 [0.8000‑0.9767] 0.9987 [0.9956‑1.0000] 21
DR 0.8625 [0.6923‑1.0000] 0.7490 [0.5000‑0.9375] 0.7966 [0.6207‑0.9333] 0.9745 [0.9335‑0.9977] 16
ERM 1.0000 [1.0000‑1.0000] 0.7490 [0.2500‑1.0000] 0.8393 [0.4000‑1.0000] 0.9493 [0.8417‑1.0000] 4
Macro-average 0.9503 [0.8985‑0.9981] 0.8365 [0.6996‑0.9440] 0.8813 [0.7571‑0.9629] 0.9783 [0.9500‑0.9987] 312 (Total)
450
Decision interpretability and feature representation 451
Decision interpretability 452
LayerCAM was used to generate class-specific activation heatmaps and identify the 453
image regions associated with the model’s predictions. For representative categories, 454
including NORMAL, AMD, DR, MH, CSC, and ERM, the regions with high 455
activation were mainly located in areas associated with retinal structural 456
abnormalities. These regions included macular morphological abnormalities, 457
alterations in retinal layer structure, and areas associated with fluid accumulation (Fig. 458
3a). For the NORMAL class, activation was concentrated mainly in the foveal region. 459
For AMD, activation was concentrated in areas corresponding to choroidal 460
neovascularization or drusen. For DR, activation covered areas of retinal thickening 461
associated with macular edema. For ERM, activation followed the highly reflective 462
interface of the epiretinal membrane. For MH, activation corresponded to the central 463

26
defect of the macular hole. For CSC, activation was concentrated in areas of retinal 464
neuroepithelial detachment. These spatial correspondences suggest that, for 465
high-sample categories, the model learned image features associated with pathological 466
structures rather than relying primarily on the image background or other 467
nondiagnostic regions. 468
However, approximately 13% of samples in the underrepresented classes showed 469
a spatial shift in the activation regions (Fig. 3b). In some CSC and ERM samples, the 470
activation regions were not precisely aligned with the lesion core and instead 471
extended into surrounding normal retinal layers or toward the image margins. In a 472
small number of misclassified samples, the LayerCAM hotspots did not correspond to 473
the typical anatomical locations of the predicted disease. For example, in an AMD 474
sample misclassified as RVO, activation was concentrated within the right-sided 475
elevated lesion rather than distributed across the overall lesion pattern. These findings 476
indicate that LayerCAM can assist in identifying image regions associated with model 477
predictions but cannot be regarded as direct evidence of the model’s underlying 478
decision mechanism. When model predictions are presented to clinicians, the 479
corresponding interpretability information should also be provided. 480

27
481
Fig. 3 Model interpretability and feature-space visualization. a) Representative correctly classified 482
cases with paired OCT B-scans and LayerCAM attention maps, showing that the model attends to 483
clinically relevant retinal lesions. b) Representative misclassified cases, with LayerCAM maps 484
highlighting the discriminatory regions driving the errors. c–f) UMAP projections of Swin-T 485
embeddings on the internal test set (c), the DUKE external set (d), the OCTID external set (e), and 486
the prospective ZZSPH cohort (f). 487
488
Feature representation analysis 489
We further analyzed the distribution of image features learned by the model for 490
different retinal diseases. The activation vector immediately preceding the model’s 491

28
prediction layer was extracted as the feature embedding for each image, and UMAP 492
was used to visualize these embeddings in two dimensions. 493
The UMAP projection of the internal test set showed that different disease 494
categories formed partially distinct clusters in the feature space (Fig. 3c). This 495
suggests that the model learned disease-related image representations with some 496
degree of class separability. However, several underrepresented classes showed 497
substantial feature overlap and spatial dispersion. This finding was consistent with the 498
class-level performance and uncertainty analyses, suggesting that class imbalance 499
may hinder the formation of stable feature distributions for some disease categories. 500
For these categories, the observed feature dispersion may indicate a need for more 501
detailed subtype classification or additional data. 502
The external public test sets showed different feature distributions (Fig. 3d and 503
e). In the DUKE dataset, NORMAL, AMD, and DR formed relatively compact 504
clusters, with only a small number of DR samples extending into the NORMAL 505
cluster. In the OCTID dataset, the disease categories formed relatively distinct feature 506
clusters. A continuous feature transition was observed between NORMAL and DR, 507
whereas AMD and CSC showed localized feature overlap. These patterns suggest that 508
macular diseases share some imaging features and that substantial lesion 509
heterogeneity exists within disease categories. 510
In the clinical cohort, NORMAL samples showed a broad and continuous 511
distribution (Fig. 3f). AMD formed a relatively compact subcluster that was generally 512
separable from NORMAL, with only limited overlap at the cluster boundaries. 513

29
Because the numbers of DR and ERM samples were limited, their feature points were 514
scattered around the boundaries between categories. Some of these samples 515
overlapped with the AMD features, suggesting shared OCT imaging characteristics 516
among DR, ERM, and AMD that may contribute to classification confusion. 517
Overall, analysis of the learned image feature space complemented the 518
conventional classification metrics and helped clarify potential sources of 519
performance differences under class imbalance. In the future, such representations 520
may assist physicians in the discrimination of complex disease categories, although 521
their clinical utility requires further validation. 522
523
Discussion 524
We evaluated the reliability of a Swin-T-based OCT classifier in a nine-class task 525
with substantial class imbalance, using an internal test set, two independent external 526
public datasets, and a prospective clinical cohort. The model showed strong overall 527
discrimination and high recall for some underrepresented classes. Nevertheless, 528
class-specific results and bootstrap analyses showed that estimates for these classes 529
were limited by small sample sizes; a single metric could therefore overstate 530
reliability. Calibration, LayerCAM, and feature-space analyses offered further insight 531
into model behavior, but each provided limited evidence on its own. Together, these 532
findings show why OCT classifiers should be assessed beyond aggregate 533
performance. 534
Class imbalance is common in medical imaging AI, particularly in clinical 535

30
datasets containing multiple disease categories, where the number of samples may 536
differ substantially across classes [16]. We used class-weighted sampling and data 537
augmentation to lessen its effects during training. Although recall was high for some 538
underrepresented classes, high recall alone does not establish clinical reliability. 539
False-positive predictions may still limit precision and increase the need for clinical 540
review. Evaluation under class imbalance should therefore consider precision, recall, 541
F1 score, and uncertainty in their estimates, rather than sensitivity or overall accuracy 542
alone. 543
Performance estimates for underrepresented classes were also statistically 544
uncertain. When test samples are scarce, class-specific AUC and F1 estimates may 545
vary considerably, while confidence intervals may be wide or degenerate. High recall 546
based on only a few cases, for example, may not reflect performance in a larger 547
clinical population. Reports should present class-specific sample sizes and confidence 548
intervals so that estimates from limited data are not overinterpreted. This is especially 549
important in multi-disease OCT classification, where some diagnostically important 550
conditions have few high-quality annotated examples. 551
Generalizability is an important consideration when assessing the potential 552
clinical utility of medical AI [17]. We developed the model using multiple public 553
OCT datasets and evaluated it on independent external data. Differences in image 554
acquisition and data sources can affect performance outside the development setting 555
[18]. External performance in our study differed from internal performance and was 556
lower on OCTID, which had a more complex class composition and potentially 557

31
overlapping image features. Strong internal results alone are therefore insufficient to 558
establish clinical applicability; evaluation across independent data sources is also 559
needed. 560
In addition to classification performance, the reliability of predicted probabilities 561
is an important component of overall model reliability [19]. We used ECE to assess 562
agreement between confidence and observed outcomes and compared post hoc 563
calibration methods. Overall ECE was low, yet some high-confidence predictions for 564
underrepresented classes were incorrect. Aggregate calibration can thus conceal 565
class-specific problems, particularly when larger classes dominate the metric. ECE 566
should be considered alongside class sizes, prediction distributions, and 567
high-confidence errors. 568
Interpretability analyses provided additional, indirect evidence about model 569
behavior. LayerCAM activations corresponded spatially to retinal structures in many 570
cases, while UMAP showed clustering for some classes and overlap for others, 571
including underrepresented classes. Activation maps do not establish that a particular 572
structure caused a prediction, and a two-dimensional projection does not directly 573
measure representation quality. The cited saliency-map study supports caution in 574
interpreting activation maps but does not directly address UMAP [20]. These methods 575
can help examine model behavior, but they cannot replace clinical validation [21]. 576
Classification performance, uncertainty, calibration, interpretability, and external 577
generalizability are best considered together. 578
The findings suggest a possible approach to clinical decision support. A system 579

32
could present each predicted label with its confidence, alongside relevant class-level 580
reliability estimates, such as calibration error and bootstrap confidence intervals. 581
Predictions for well-represented classes with stable performance, such as NORMAL 582
and AMD, could potentially be handled with less routine review. Cases assigned to 583
underrepresented classes, or classes with elevated calibration error or imprecise 584
performance estimates, could instead be flagged for senior ophthalmologist review. 585
Unusual activation patterns or feature representations might be identified by a 586
dedicated detector or examined by a clinician using LayerCAM, with the physician 587
retaining responsibility for the final decision. None of these workflow options was 588
tested here. Prospective human-AI studies are needed before they can be adopted in 589
practice. 590
We selected Swin-T as the representative model because its hierarchical features 591
and shifted-window attention capture both local and broader spatial information [22]. 592
It was competitive on some macro-averaged metrics, although precision, recall, and 593
F1 score varied across architectures. Model selection should therefore reflect the 594
needs of the clinical task. In ablation experiments, shifted windows were associated 595
with better performance, most notably on OCTID. Cross-window information 596
exchange may help distinguish conditions with similar local features but different 597
broader structural patterns, although its effect varied across external datasets. 598
This study has several limitations. First, despite external and prospective 599
clinical-cohort evaluations, model development relied mainly on public data. Further 600
validation across devices, institutions, and patient populations is needed. Second, 601

33
some underrepresented classes had few test cases, leaving imprecise performance 602
estimates that require confirmation in larger clinical datasets. Third, our analyses used 603
primarily two-dimensional OCT images and did not fully use the information 604
available in volumetric scans; three-dimensional or 2.5D approaches warrant 605
investigation. Finally, we assessed model outputs rather than human-AI collaboration 606
in routine care. Physician-involved studies are needed to determine effects on 607
workflow efficiency, diagnostic agreement, and clinical decisions. 608
609
Conclusions 610
The OCT-based Swin Transformer achieved high overall discrimination in a 611
nine-class retinal disease task, but aggregate metrics did not fully characterize its 612
reliability under class imbalance. Small sample sizes limited the precision of estimates 613
for underrepresented classes, while external testing and calibration analyses revealed 614
differences in performance and probability reliability across data sources and classes. 615
Evaluation of multi-disease OCT classifiers should therefore extend beyond 616
conventional metrics to include class-specific stability and uncertainty, probability 617
calibration, interpretability, and learned feature representations. The class-level 618
reliability indicators examined here may help identify cases that warrant human 619
review. Clinical value, however, remains to be established in larger multicenter 620
cohorts and prospective studies of human–AI collaboration. 621
622
623
624

34
Abbreviations 625
Abbreviation Full name
OCT Optical Coherence Tomography
CNN Convolutional Neural Network
Swin‑T Swin Transformer‑Tiny
NORMAL Normal Retinal
AMD Age‑related Macular Degeneration
DR Diabetic Retinopathy
MH Macular Hole
ERM Epiretinal Membrane
RVO Retinal Vein Occlusion
RAO Retinal Artery Occlusion
RRD Rhegmatogenous Retinal Detachment
CSC Central Serous Chorioretinopathy
AUC Area Under the Receiver‑Operating‑Characteristic Curve
Recall (Sensitivity) True Positive Rate / Sensitivity
ECE Expected Calibration Error
626
Acknowledgements 627
Not applicable. 628
629

35
Authors’ contributions 630
Baohua Li: Conceptualization, Data curation, Investigation, Formal analysis, Writing 631
– original draft. Xinfei Chen: Conceptualization, Data curation, Methodology, 632
Software, Writing – original draft. Meng Li: Methodology, Formal analysis, 633
Investigation, Writing – review & editing. Kai Wang: Data curation, Investigation, 634
Writing. Jingnan Li: Data curation, Investigation. Shu Li: Data curation, Investigation. 635
Jie Du: Supervision, Project administration, Funding acquisition, Writing – review & 636
editing. All authors have read and approved the final manuscript. 637
638
Funding 639
This study was supported by the Zhengzhou Science and Technology Benefit-People 640
Program (Grant No. 2024KJHM0029). 641
642
Availability of data and materials 643
The datasets and code analysed during the current study are publicly available as 644
follows. The prospective clinical OCT cohort from Zhengzhou Seventh People’s 645
Hospital (ZZSPH) is accessible at 646
https://huggingface.co/spaces/dujie-lab/uncommon-retinal-disease-AI-OCT/tree/main. 647
The ten public OCT datasets used for model development and external validation can 648
be obtained from their respective original repositories, as detailed in Supplementary 649
Table S1. The source code for model training and evaluation is available at 650
https://github.com/dujie-lab/uncommon-retinal-disease-AI-OCT. 651
652
Declarations 653
Ethics approval and consent to participate 654
This prospective study was approved by the Medical Ethics Committee of Zhengzhou 655
Seventh People’s Hospital (approval number: 2026-KY-019). The study was 656
conducted in accordance with the Declaration of Helsinki. All participants provided 657
written informed consent prior to enrolment. Patient data used in this study were 658
de-identified prior to analysis. 659
660
Consent for publication 661
Not applicable. This manuscript does not contain any individual person’s data in any 662
form (e.g., identifiable clinical photographs or personal information). 663
664
665

36
Competing interests 666
The authors declare no competing interests. 667
668
References 669

1. Organization WH: World report on vision. In. Geneva: World Health Organization; 2019. 670
2. Teo ZL, Tham YC, Yu M, Chee ML, Rim TH, Cheung N, Bikbov MM, Wang YX, Tang Y, 671
   Lu Y et al: Global Prevalence of Diabetic Retinopathy and Projection of Burden through 672
   2045: Systematic Review and Meta-analysis. Ophthalmology 2021, 128(11):1580–1591. 673
3. Zeppieri M, Marsili S, Enaholo ES, Shuaibu AO, Uwagboe N, Salati C, Spadea L, Musa M: 674
   Optical Coherence Tomography (OCT): A Brief Look at the Uses and Technological 675
   Evolution of Ophthalmology. Medicina (Kaunas) 2023, 59(12):2114. 676
4. Langlo CS, Amin A, Park SS: Optical coherence tomography retinal imaging: narrative 677
   review of technological advancements and clinical applications . Ann Transl Med 2025, 678
   13(2):17. 679
5. Zhang J, Zhong J, Lin L, Li J, Ma K, Li N, Li M, Pan Y, Meng Z, Zhou M et al : Full 680
   end-to-end diagnostic workflow automation of 3D OCT via foundation model-driven AI 681
   for retinal diseases. npj Digital Medicine 2026. 682
6. De Fauw J, Ledsam JR, Romera-Paredes B, Nikolov S, Tomasev N, Blackwell S, Askh am H, 683
   Glorot X, O’Donoghue B, Visentin D et al : Clinically applicable deep learning for 684
   diagnosis and referral in retinal disease. Nat Med 2018, 24(9):1342–1350. 685
7. Kermany DS, Goldbaum M, Cai W, Valentim CCS, Liang H, Baxter SL, McKeown A, Yang 686
   G, Wu X, Yan F et al : Identifying Medical Diagnoses and Treatable Diseases by 687
   Image-Based Deep Learning. Cell 2018, 172(5):1122–1131 e1129. 688
8. Laouarem A, Kara-Mohamed C, Bourennane EB, Hamdi-Cherif A: HTC-retina: A hybrid 689
   retinal diseases classification model using transformer-Convolutional Neural Network 690
   from optical coherence tomography images. Comput Biol Med 2024, 178:108726. 691
9. Pan L, Zhang Y, Yang Q, Li T, Chen Z: Long-tailed medical diagnosis with relation-aware 692
   representation learning and iterative classifier calibration . Comput Biol Med 2025, 693
   188:109772. 694
10. Peng Y, Lin A, Wang M, Lin T, Liu L, Wu J, Zou K, Shi T, Feng L, Liang Z et al: Enhancing 695
    AI reliability: A foundation model with uncertainty estimation for optical coherence 696
    tomography-based retinal disease diagnosis. Cell Rep Med 2025, 6(1):101876. 697
11. Penso C, Frenkel L, Goldberger J: Confidence Calibration of a Medical Imaging 698
    Classification System That is Robust to Label Noise . IEEE Trans Med Imaging 2024, 699
    43(6):2050–2060. 700
12. Zhao J, Lin T, Wang M, Sun L, Chen H, Zhang D, Fu H: F2UE-NET: An 701
    Uncertainty-Driven Dynamic Expert Network for Long-Tail Retinal Disease 702
    Classification. In: 2025 IEEE 22nd International Symposium on Biomedical Imaging (ISBI ). 703
    2025: 1–5. 704
13. Leyba-Mesa MV, Barkana BD: Interpretable Deeply Supervised Networks for 705
    Class-Imbalanced OCT Classification. AI 2026, 7(5):164. 706

37
14. Zhu Y, Niu Y, Sun S, Liu W, Dou Y, Guo Y: Performance of Deep Learning in Classifying 707
Age-Related Macular Degeneration From Images: Systematic Review and 708
Meta-Analysis. J Med Internet Res 2026, 28:e97174. 709
15. Liu Z, Lin Y, Cao Y, Hu H, Wei Y, Zhang Z, Lin S, Guo B : Swin Transformer: 710
Hierarchical Vision Transformer Using Shifted Windows . In: 2021 IEEE/CVF 711
International Conference on Computer Vision (ICCV). Montreal, QC, Canada: IEEE; 2021: 712
10012–10022. 713
16. Wu Z, Guo K, Luo E, Wang T, Wang S, Yang Y, Zhu X, Ding R: Medical long-tailed 714
learning for imbalanced data: Bibliometric analysis . Comput Methods Programs Biomed 715
2024, 247:108106. 716
17. Matta S, Lamard M, Zhang P, Le Guilcher A, Borderie L, Cochener B, Quellec G: A 717
systematic review of generalization research in medical image classification . Comput Biol 718
Med 2024, 183:109256. 719
18. Su Z, Guo J, Yang X, Wang Q, Coenen F, Hussain A, Huang K: Navigating Distribution 720
Shifts in Medical Image Analysis: A Survey. arXiv preprint arXiv:2411.05824, 2024. 721
19. Vijayan M, Veerapu G, J SN: A spatial adaptive multi-scale ConvNeXt framework for 722
robust, calibrated, and statistically validated multi-class retinal OCT classification . Front 723
Artif Intell 2026, 9:1898978. 724
20. Venkatesh K, Mutasa S, Moore F, Sulam J, Yi PH: Gradient-Based Saliency Maps Are Not 725
Trustworthy Visual Explanations of Automated AI Musculoskeletal Diagnoses . J 726
Imaging Inform Med 2024, 37(5):2490–2499. 727
21. Carrick RT, Rowin EJ, Gasperetti A, Madias C, Maron M, Wu KC: Enhancing 728
explainability in clinical deep-learning models: Latent-space variable decoding is 729
superior to gradient-weighted class activation mapping . Heart Rhythm O2 2025, 730
6(9):1248–1258. 731
22. Ait Hammou B, Antaki F, Boucher MC, Duval R: MBT: Model-Based Transformer for 732
retinal optical coherence tomography image and video multi-classification . Int J Med 733
Inform 2023, 178:105178. 734
735

Supplementary Files
This is a list of supplementary files associated with this preprint. Click to download.
Supplementarymaterial.tar