PaperlayerView exact revision on arXiv

Source-linked AI summary

LightHybridNet: A Lightweight Hybrid Deep Learning Framework for Efficient Multi-Label Ophthalmic Disease Classification in Imbalanced Fundus Images

Hem Raj Dadral Najam Ul Hasan, Naveed Ejaz, Muhammad Aslam Jarwar

arXiv:2610.20671v1eess.IV

TL;DR

Imbalanced multi-label retinal classification remains difficult, while heavyweight models may not suit resource-constrained settings. LightHybridNet combines efficient backbones with mechanisms for multi-scale features and interpretable predictions; on ODIR-5K, it reports 0.9151 weighted F1, 0.9826 macro AUC, and 0.9014 Cohen’s Kappa.

  • Problem

    Reliable multi-label retinal classification is challenging because disease features overlap and datasets such as ODIR-5K are severely imbalanced, while many high-performing models are computationally heavy.

    Source: §I · p. 2

  • Method

    LightHybridNet combines MobileNetV3-Large and EfficientNet-B1 with depthwise separable convolutions, compound scaling, and Grad-CAM visualisation.

    Source: §I · p. 2

  • Results

    0.9151 weighted F1-score, 0.9826 macro AUC, and 0.9014 Cohen’s Kappa were reported for LightHybridNet with replication-based oversampling.

    Source: D. Comparative performance and… · p. 10

  • Takeaways & Limitations

    LightHybridNet reports these ODIR-5K results with 10.74M parameters and 0.853 GFLOPs, and the conclusion reports Hypertension recall of 0.8293 despite 138 original samples.

    Source: §VI · p. 13

    Performance on ODIR-5K may not transfer fully to other populations; external validation on diverse, multi-ethnic datasets is needed before broad clinical use.

    Source: §VI · p. 14

Abstract

from arXiv · show

Eye diseases are a major global cause of blindness and vision impairment, highlighting the need for accurate and early diagnosis to prevent further deterioration. Despite advances in medical imaging, classifying retinal diseases from fundus images remains challenging due to the complexity and subtlety of visual features. This study aims to develop and validate LightHybridNet, a novel lightweight hybrid deep learning framework, to improve the classification of eye diseases from fundus images. The model incorporates architectural features, such as depthwise separable convolutions and compound scaling mechanisms, to enhance diagnostic accuracy and robustness. To validate the model, the ODIR-5K dataset, comprising 8 disease categories, was used. To enhance image quality, pre-processing approaches were applied, including histogram equalisation, colour space conversion, and contrast adjustment, resulting in high-resolution images. Furthermore, image augmentation was employed to increase the dataset size, thereby improving the model's ability to generalise during training. The hybrid architecture helps the model concentrate on relevant image regions while capturing and integrating details at different scales, significantly improving classification performance. To evaluate its effectiveness, the model was compared with transfer-learning models. The results show that LightHybridNet outperforms the baselines, achieving a weighted F1-score of 91.51%, a macro AUC of 98.26%, and a Cohen's Kappa of 90.14% on the validation set. Furthermore, model interpretability was enhanced using explainable AI methods, including Grad-CAM, to identify and emphasise features influencing the decision-making process. LightHybridNet represents a promising advancement in automated ocular disease detection, offering improved diagnostic reliability for clinical use.

I. INTRODUCTION

The introduction frames retinal disease classification as a balance between accuracy and deployability on imbalanced multi-label fundus data. It presents LightHybridNet as a lightweight hybrid approach and situates it among work on efficient models, imbalance, and interpretability.

I. INTRODUCTION

  • The paper asks whether a lightweight hybrid can classify imbalanced, multi-label retinal images reliably while keeping computational cost low for resource-constrained settings.The introduction contrasts heavyweight models with efficient approaches that may underperform on severely imbalanced data.Source: §I · p. 2
  • LightHybridNet combines MobileNetV3 and EfficientNet, using depthwise separable convolutions and compound scaling to capture diagnostic regions and features at different scales.Grad-CAM is integrated to highlight features influencing decisions.Source: §I · p. 2
  • 91.51% weighted F1-score is reported for classifying eight eye diseases with LightHybridNet.The introduction describes the result as supporting reliable real-world clinical deployment.Source: §I · p. 2

II. RELATED WORK

  • Related work identifies overlapping disease features and severe class imbalance as challenges for lightweight models, motivating hybrids that combine efficient backbones.The review also notes gaps in unified approaches to multi-disease classification, severity staging, and interpretability.Source: §II · p. 4

III. OPHTHALMIC DISEASES IDENTIFIED THROUGH RETINAL ABNORMALITIES IN FUNDUS IMAGES

The section describes the retinal abnormalities associated with the ocular conditions represented in ODIR-5K. The examples include optic-disc, vascular, macular, and lens changes, as well as other abnormalities and comorbid presentations.

III. OPHTHALMIC DISEASES IDENTIFIED THROUGH RETINAL ABNORMALITIES IN FUNDUS IMAGES

  • Glaucoma is associated with optic-disc cup-to-disc changes, nerve fibre layer thinning, and optic-disc haemorrhages; diabetic retinopathy can be asymptomatic before visual impairment develops.The section describes glaucoma as affecting the optic nerve and diabetic retinopathy as affecting retinal tissues.Source: §III · p. 4
  • AMD signs include drusen, pigmentary changes, neovascularisation, and geographic atrophy, while cataracts are associated with diffuse clouding of ocular structures.Macular fluid accumulation is also described as distorting vision.Source: §III · pp. 4–5
  • ODIR-5K examples cover Normal, Diabetic Retinopathy, Glaucoma, Cataract, AMD, Hypertension, Pathological Myopia, and Other abnormalities, including selected comorbid presentations.The section describes fundus signs associated with these disease categories.Source: §III · p. 4, Figure 1 · p. 5
  • Hypertension can involve arteriolar narrowing, arteriovenous nicking, and flame haemorrhages; pathological myopia is associated with tessellated fundus and peripapillary atrophy.The section also notes cardiovascular complications for hypertension and potential vision loss with pathological myopia.Source: §III · p. 5
  • The Other abnormalities category includes findings such as pigment epithelium proliferation, epiretinal membrane, and mixed pathologies.Source: §III · p. 5

IV. METHODOLOGY

The methodology uses paired-eye fundus images and patient-level multi-label annotations from ODIR-5K. It filters image-quality problems and assigns each patient's label vector to both eye images for per-image model training.

A. ODIR-5K Dataset

  • ODIR-5K contains around 5,000 patients’ paired fundus images, with specialist-assigned diagnostic keywords and eight disease classes, including Normal and Others.Images were captured using different cameras and resolutions; patient age and gender are also included.Source: A. ODIR-5K Dataset · p. 5
  • The dataset is highly imbalanced, and its varied cameras and illumination conditions make disease identification challenging.The Other class contains lesions associated with 12 different ophthalmic diseases.Source: A. ODIR-5K Dataset · pp. 5–6
  • 324 images, approximately 3.2% of the initial set, were removed, leaving 9,676 valid fundus images after pairing verification.Filtering excluded missing or corrupted files and images with specified non-diagnostic quality issues.Source: A. ODIR-5K Dataset · p. 6
  • Each left- and right-eye image was treated as a separate input and assigned the same patient-level eight-dimensional label vector.The passage describes this label propagation as consistent with ODIR-5K's patient-level annotations and the possibility of bilateral or unilateral disease.Source: A. ODIR-5K Dataset · p. 6

B. Pre-processing and Image Enhancement

Preprocessing standardises fundus images and applies local contrast enhancement to address variation in cameras and illumination. Training-only augmentation adds geometric, illumination, and contrast variation, while validation and test images receive deterministic preprocessing.

B. Pre-processing and Image Enhancement

  • Field-of-view-aware cropping removes redundant background, and images are resized to 224 × 224 pixels for the MobileNetV3 and EfficientNet backbones.Source: B. Pre-processing and Image Enhancement · p. 6
  • CLAHE operates on local 8 × 8 tiles with a clip limit of 2.0 to limit noise over-enhancement in homogeneous fundus regions.Clipped histogram values are redistributed before calculating the cumulative distribution function; the method is described as enhancing fine vessels and subtle lesions while maintaining anatomical fidelity.Source: B. Pre-processing and Image Enhancement · p. 6

C. Data Augmentation Strategy

  • Training-only augmentation includes flips, rotations, brightness and contrast changes, CLAHE, shift-scale-rotate, and coarse dropout.The pipeline is used to simulate acquisition variability and mitigate overfitting, particularly to replicated minority-class samples.Source: C. Data Augmentation Strategy · pp. 6–7
  • Validation and test images receive only deterministic cropping, resizing, and normalisation rather than stochastic augmentation.The passage states this is intended to ensure unbiased evaluation under realistic conditions; Figure 2 illustrates the training transformations.Source: C. Data Augmentation Strategy · p. 7, Figure 2 · p. 7

D. Class Balancing

The study treats each eye image separately, removes artefact-affected images, and addresses class imbalance by duplicating minority samples with augmentation. LightHybridNet then combines complementary features from two backbones through late fusion.

D. Class Balancing

  • Left and right fundus images are treated as separate inputs, while images with multiple diseases receive individual disease-class labels; artefact-affected images are removed.The original class distribution has substantially more normal, diabetes, and “other” images than the other disease classes.Source: D. Class Balancing · p. 7
  • Minority-class samples are duplicated to approximately match the majority-class count, with aggressive on-the-fly augmentation used to reduce overfitting to duplicates.The approach avoids synthetic samples that could be anatomically unrealistic, although duplication may introduce memorisation bias.Source: D. Class Balancing · p. 7

E. Overview of the Proposed LightHybridNet Architecture

  • LightHybridNet concatenates 1280-dimensional features from parallel MobileNetV3-Large and EfficientNet-B1 backbones, then applies dropout and an 8-output sigmoid head; the model has 10.74M parameters and 0.853 GFLOPs.Late fusion allows each backbone to develop specialised representations independently; attention mechanisms were excluded to prioritise edge deployment feasibility over marginal accuracy gains.Source: E. Overview of the Proposed… · p. 7

1. MobileNetV3-Large (mobilenetv3 large 100.ra in1k)

  • MobileNetV3-Large is used as a feature extractor that outputs 1280-dimensional embeddings; its depthwise separable convolutions, inverted residuals, and squeeze-and-excitation blocks support local texture and fine-grained detail capture.The “large” configuration uses a width multiplier of 1.0, “ra” denotes RandAugment training, and “in1k” denotes ImageNet-1k pretraining.Source: §1 · pp. 7–8

2. EfficientNet-B1 (efficientnet b1)

EfficientNet-B1 serves as a pretrained feature extractor, producing features for fundus-image classification. The evaluation section describes metrics for multi-label, imbalanced data and a severity-weighted score intended to support clinical triage.

2. EfficientNet-B1 (efficientnet b1)

  • EfficientNet-B1 uses compound scaling to balance network depth, width, and resolution, and produces 1280-dimensional features when configured as a feature extractor.Its MBConv blocks, squeeze-and-excitation optimisation, and swish activations support multi-scale feature capture.Source: §2 · p. 8

F. Evaluation Metrics

  • Binary Cross-Entropy with Logits Loss independently optimises each of the 8 labels, and macro-averaged AUC treats classes equally to emphasise rare pathologies.The reported evaluation metrics also include Hamming Score and support-weighted precision, recall, and F1-score.Source: F. Evaluation Metrics · p. 8
  • A composite risk score weights per-class probabilities by severity, with higher scores indicating greater overall ocular risk and supporting screening prioritisation for specialist referral.The weights reflect estimated vision-loss risk and intervention urgency; the scoring is described as requiring no additional training.Source: F. Evaluation Metrics · p. 9

A. Experimental Setup

Experiments use a patient-level stratified 80/20 training-validation split on preprocessed ODIR-5K, with both eyes from each patient kept in the same partition. The study compares class-imbalance strategies and reports gains for replication-based oversampling.

A. Experimental Setup

  • The patient-level stratified split assigns 80% of preprocessed ODIR-5K to training and 20% to validation, keeping fellow-eye images in the same partition to prevent data leakage.The training set contains approximately 9,968 images before replication-based oversampling expands it to approximately 24,933 samples.Source: A. Experimental Setup · p. 9
  • The study benchmarks LightHybridNet against heavyweight baselines and lightweight alternatives, and includes ablations of imbalance strategies and backbone combinations.Final reported weighted F1-score, macro AUC, and Cohen’s Kappa are computed on the held-out validation portion; the authors note that cross-validation or an external test set could provide more robust estimates.Source: A. Experimental Setup · pp. 9–10

B. Impact of Class Imbalance Mitigation

C. Training Progression

LightHybridNet converged by epochs 22–24 and outperformed comparison models while remaining compact; per-class results also show persistent false negatives in clinically critical minority diseases.

C. Training Progression

  • Training loss declined from 0.3590 at epoch 1 to 0.0043 at epoch 25, with peak validation performance around epochs 22–24.The authors report stable convergence without overfitting.Source: C. Training Progression · p. 10

D. Comparative performance and efficiency metrics (ablation and benchmarks)

  • Oversampling improved weighted F1 by 3.38% over the hybrid without oversampling, whereas hybrid fusion alone yielded modest gains.The full hybrid with replication-based oversampling achieved 10.74M parameters, 0.853 GFLOPs, weighted F1-score 0.9151, macro AUC 0.9826, and Cohen’s Kappa 0.9014.Source: D. Comparative performance and… · p. 10
  • Macro AUC was 2.18% higher than DKCNet, while parameters fell by 88% and GFLOPs by 94%.The comparison table covers ablation and benchmark performance and efficiency metrics.Source: Table IV · p. 10, D. Comparative performance and… · p. 10

E. Per-Class Performance

  • Hypertension recall was 0.8293 despite 138 original samples, while the heterogeneous Others class had the lowest reported recall, 0.8288.Cataract and myopia had near-perfect AUC (> 0.99).Source: E. Per-Class Performance · p. 11
  • Clinically critical minority classes still had nonnegligible false negatives, especially when subtle signs were masked by more salient or co-occurring disease features.The authors identify these residual misses as concerning for high-risk conditions.Source: E. Per-Class Performance · p. 11

F. Clinical Error Analysis via Confusion Matrices

The confusion-matrix analysis highlights missed cases in minority diseases, particularly when subtle signs are obscured by co-occurring or more salient pathologies.

F. Clinical Error Analysis via Confusion Matrices

  • Hypertension had a 17.1% false-negative rate, with 14 missed cases among 82 positive test samples.Misses often involved subtle or early-stage signs overshadowed by diabetic retinopathy or normal vasculature.Source: F. Clinical Error Analysis via… · p. 11
  • AMD had a 10.9% false-negative rate, with 15 missed cases among 138 positive test samples; early or intermediate cases were predominantly missed.Late-stage neovascular or geographic atrophy cases were rarely missed.Source: F. Clinical Error Analysis via… · pp. 11–12
  • The heterogeneous Others class had the highest overall false-negative rate, approximately 18.3%, while Glaucoma and Cataract had rates below 8%.The authors attribute the Others result to its mix of 12 or more pathologies.Source: F. Clinical Error Analysis via… · p. 12
  • Dominant or visually salient pathologies can overshadow subtle minority-class features, leaving false negatives in high-risk categories a clinical-safety priority.The authors note that missed detection in high-risk conditions can have serious consequences.Source: F. Clinical Error Analysis via… · p. 12

G. Efficiency and Inference Speed

LightHybridNet is presented as an efficient model for screening on constrained devices, and Grad-CAM visualisations highlight disease-relevant regions while misclassifications cluster around subtle co-morbid signs.

G. Efficiency and Inference Speed

  • Inference latency was 12 ms on a high-end GPU, 45 ms on CPU, and 25 ms in mobile emulation via TensorFlow Lite.The authors state these speeds enable real-time screening on portable devices.Source: G. Efficiency and Inference Speed · p. 12

H. Efficiency and Edge Deployment Feasibility

  • The model has 10.74 million parameters and an approximately 41.0 MB FP32 footprint; CPU inference was approximately 81 ms per image in a virtualised i7-class environment.The authors describe real-time or near-real-time screening at approximately 12–15 images per minute on CPU-only devices, without cloud connectivity or high-end GPUs.Source: H. Efficiency and Edge Deployment… · p. 12

I. Interpretability Analysis

  • Grad-CAM highlighted disease-relevant regions, while misclassifications primarily involved subtle co-morbidities obscured by dominant pathology.Reported attention regions included the optic disc for glaucoma, microaneurysms and exudates for diabetic retinopathy, macular drusen for AMD, and vascular narrowing for hypertension.Source: §I · p. 12

J. Analysis of Misclassifications

LightHybridNet shows strong ODIR-5K performance and encouraging zero-shot transfer, while its errors cluster around subtle or overlapping disease features. The analysis also identifies demographic representativeness as a boundary on broader clinical use.

J. Analysis of Misclassifications

  • Misclassifications clustered in subtle or overlapping pathology, including mild diabetic retinopathy labeled normal and borderline glaucoma confused with myopia.Grad-CAM heatmaps highlighted regions influencing these incorrect decisions; ensemble methods or clinical metadata were suggested for ambiguous cases.Source: J. Analysis of Misclassifications · p. 13

K. Generalisation to Unseen Data

  • A zero-shot MESSIDOR-2 evaluation yielded AUC 0.8564 for referable diabetic retinopathy without fine-tuning, versus 0.98 AUC for fine-tuned models.The passage describes this as transferable feature learning despite demographic, camera, and label-granularity domain shifts.Source: K. Generalisation to Unseen Data · p. 13

VI. CONCLUSION

  • LightHybridNet showed superior macro AUC and Kappa for minority-class discrimination; DKCNet achieved weighted F1 0.943 with 87.7M parameters, while DCNet reported F1 0.913 using a heavier ResNet backbone.Source: §VI · p. 13
  • ODIR-5K’s China-only, predominantly East Asian population limits confidence in generalization across other ethnic groups; diverse external validation is needed before broad clinical use.Severity grading and physical device testing were also beyond the study’s scope.Source: §VI · p. 14

Related papers

Loading 2610.20671v1…