Interpretable Diabetic Retinopathy Classification Using Hybrid ConvNeXt–MLP-Mixer and ConvNeXt–PoolFormer Architectures
Abstract
Diabetic retinopathy (DR) severity grading from retinal fundus images is challenging because retinal abnormalities vary in size, distribution, and severity. This study evaluates two hybrid architectures, ConvNeXt–MLP-Mixer and ConvNeXt–PoolFormer, for five-class DR classification by combining convolutional feature extraction with global feature mixing. Experiments used the APTOS 2019 Blindness Detection dataset from Kaggle. After exact-duplicate filtering, 3,596 images remained: 2,502 for training, 544 for validation, and 550 for held-out testing. The imbalanced test set comprised five DR severity classes. The proposed models were compared with EfficientNet, Vision Transformer, ConvNeXt, MLP-Mixer, and PoolFormer using accuracy, macro- and weighted F1-scores, and macro- and weighted-average one-versus-rest (OvR) ROC-AUC. ConvNeXt–PoolFormer achieved 82.7% test accuracy and 82.7% weighted F1-score, while ConvNeXt–MLP-Mixer achieved 82.4% accuracy and 82.5% weighted F1-score, with the latter obtaining the highest macro F1-score of 67.5%. Moderate DR showed the clearest class-specific improvement, whereas Severe and Proliferative DR remained difficult to classify. Layer-wise Relevance Propagation (LRP) and Grad-CAM generated post-hoc visual explanation maps for representative predictions. These visualizations were treated as exploratory because they were not independently validated using lesion-level annotations or by clinical experts, including ophthalmologists.
Related
- CBAM-enhanced convolutional networks for diabetic retinopathy classification in the ODIR dataset
- Diffusion-augmented contrastive learning for diabetic retinopathy severity classification
- Successful management of high-risk proliferative diabetic retinopathy during pregnancy with intravitreal faricimab
- ISG15+ Müller glial subtype contributes to early blood–retinal barrier breakdown in diabetic retinopathy
- A unified deep learning framework for multi-lesion detection in diabetic retinopathy using fundus images
BibTeX
@article{jadmika2026interpretable,
title = {{Interpretable Diabetic Retinopathy Classification Using Hybrid ConvNeXt–MLP-Mixer and ConvNeXt–PoolFormer Architectures}},
author = {Mochammad Arya Jadmika and Aris Sugiharto and Dinar Mutiara Kusumo Nugraheni},
journal = {Journal of Information Systems and Informatics},
year = {2026},
month = oct,
doi = {10.63158/journalisi.v8i5.1863},
url = {https://doi.org/10.63158/journalisi.v8i5.1863},
note = {CC BY},
}
RIS
TY - JOUR
TI - Interpretable Diabetic Retinopathy Classification Using Hybrid ConvNeXt–MLP-Mixer and ConvNeXt–PoolFormer Architectures
AU - Mochammad Arya Jadmika
AU - Aris Sugiharto
AU - Dinar Mutiara Kusumo Nugraheni
PY - 2026
DA - 2026/10/24
JO - Journal of Information Systems and Informatics
DO - 10.63158/journalisi.v8i5.1863
UR - https://doi.org/10.63158/journalisi.v8i5.1863
AB - Diabetic retinopathy (DR) severity grading from retinal fundus images is challenging because retinal abnormalities vary in size, distribution, and severity. This study evaluates two hybrid architectures, ConvNeXt–MLP-Mixer and ConvNeXt–PoolFormer, for five-class DR classification by combining convolutional feature extraction with global feature mixing. Experiments used the APTOS 2019 Blindness Detection dataset from Kaggle. After exact-duplicate filtering, 3,596 images remained: 2,502 for training, 544 for validation, and 550 for held-out testing. The imbalanced test set comprised five DR severity classes. The proposed models were compared with EfficientNet, Vision Transformer, ConvNeXt, MLP-Mixer, and PoolFormer using accuracy, macro- and weighted F1-scores, and macro- and weighted-average one-versus-rest (OvR) ROC-AUC. ConvNeXt–PoolFormer achieved 82.7% test accuracy and 82.7% weighted F1-score, while ConvNeXt–MLP-Mixer achieved 82.4% accuracy and 82.5% weighted F1-score, with the latter obtaining the highest macro F1-score of 67.5%. Moderate DR showed the clearest class-specific improvement, whereas Severe and Proliferative DR remained difficult to classify. Layer-wise Relevance Propagation (LRP) and Grad-CAM generated post-hoc visual explanation maps for representative predictions. These visualizations were treated as exploratory because they were not independently validated using lesion-level annotations or by clinical experts, including ophthalmologists.
SN - 2656-5935
SN - 2656-4882
ER -