Impact of Class Imbalance on Deep Learning-Based Diabetic Retinopathy Classification: An Analysis of the APTOS 2019 Dataset

Abstract

Diabetic retinopathy (DR) is a major microvascular complication of diabetes and an important
cause of preventable visual impairment. The increasing use of deep learning for retinal image
analysis has demonstrated promising diagnostic performance; however, the distribution of
disease severity within publicly available datasets can substantially influence the interpretation
of classification performance. This study analyses the implications of class imbalance in the
APTOS 2019 diabetic retinopathy dataset and examines how imbalance should be considered
when interpreting transfer-learning-based classification metrics. The APTOS 2019 dataset
contains 3,662 retinal fundus images distributed across five severity categories: No Diabetic
Retinopathy, Mild, Moderate, Severe and Proliferative Diabetic Retinopathy. The dataset
exhibits considerable class imbalance, with 1,805 images in the No DR category and only 193
images in the Severe DR category, corresponding to a largest-to-smallest class ratio of 9.35:1.
Reported performance values for DenseNet121, ResNet50, VGG16, InceptionV3 and
MobileNetV2 were examined using accuracy, area under the curve (AUC), precision, recall
and F1-score. Although the reported accuracy values were generally high, substantial
differences were observed between precision and recall for some architectures. MobileNetV2,
for example, achieved 99.64% precision but only 74.57% recall and an F1-score of 85.22%,
whereas VGG16 demonstrated a more balanced precision of 94.68% and recall of 95.42%.
These findings highlight the importance of considering multiple performance indicators rather
than relying on accuracy alone when evaluating multiclass DR classification under imbalanced
class distributions. The analysis provides a dataset-level perspective on class imbalance and
its implications for responsible interpretation of deep-learning-based diabetic retinopathy
classification.