eess.IVAug 15, 2026
SaveMulti-Channel Feature Fusion and Monte Carlo Dropout for Uncertainty-Aware Diabetic Retinopathy Grading
Organizations: Department of Computer Science & Engineering, Indian Institute of Technology Patna, Patna, Bihar, 801106, India
Abstract
Automated five-stage diabetic retinopathy (DR) grading requires more than high accuracy alone. Medical-grade deployment calls for lesion-aware preprocessing, ordinal predictions, calibrated uncertainty, and explainability to support reliable diagnostic systems. We present a unified pipeline that addresses these requirements using a Ben-Graham-green-channel CLAHE feature representation, an EfficientNetV2-L ordinal regressor, and Monte Carlo dropout for uncertainty-driven referral. Grad-CAM provides visual explanations aligned with clinically relevant lesions. The proposed method achieves a QWK of 91.31% on the APTOS-2019 official test split, placing it within the near-perfect agreement band (>80%). At a 20% referral rate, 293 of 366 images are automatically graded with a QWK of 90.40%. More complex cases are referred for specialist assessment, demonstrating a practical trade-off among grading quality, automation, and patient safety in robust, reliable, and deployment-ready medical diagnostic systems.
Figures & tables
Figure 1: Representative fundus images for each of the five ICDRSS diabetic retinopathy severity stages (No DR to Proliferative DR).
| Method | Year | QWK (%) |
|---|---|---|
| DenseNet-169 + CBAM [ 7 ] | 2022 | 88.80 |
| ResNet-50 + EfficientNet-B0 [ 25 ] | 2024 | 91.90 |
| MHSA-CNN [ 1 ] | 2026 | 91.85 |
| RadFuse [ 19 ] | 2025 | 93.24 |
| Triple-DRNet [ 12 ] | 2023 | 93.62 |
Table 1: Selected QWK values reported for five-stage DR grading on APTOS-2019. Values are not directly comparable unless the dataset split and model-selection protocol match.
| Paper | Methodology | Preprocessing | Classification Type | Evaluation Metric |
|---|---|---|---|---|
| [ 6 ] | EfficientNet-B3 + SE blocks | Image enhancement, augmentation | 5-class classification | 88.44% Acc, 85.39% QWK |
| [ 15 ] | OptiDex (NASNet-Mobile + enhanced Cat Swarm + XAI) | Morphological processing, NAS optimisation | Detection and severity classification | 97.65% Acc |
| [ 27 ] | EfficientNet-B0 with preprocessing strategies | Extensive preprocessing | DR detection | 83% Acc |
| [ 3 ] | Self-stacking ensemble | Lesion isolation with attention | 5-class classification | 88.58% QWK |
| [ 25 ] | Dual-branch transfer learning (ResNet-50 + EfficientNet-B0) | Augmentation, quality enhancement | Binary + 5-class grading | 89.6% Acc, 91.90% QWK |
| [ 22 ] | Ensemble of deep CNNs | Standard preprocessing | Multi-class classification | 82.74% Acc |
Table 2: Survey of recent works in diabetic retinopathy detection – Part 1.
| Paper | Methodology | Preprocessing | Classification Type | Evaluation Metric |
|---|---|---|---|---|
| [ 23 ] | 10-layer shallow ResNet | Lightweight preprocessing | Multi-class classification | 98.65% Acc |
| [ 14 ] | Ensemble with optimised transfer learning | Feature extraction and selection | Early-stage diagnosis | 88.95% Acc |
| [ 10 ] | Supervised contrastive learning + Xception | CLAHE enhancement | Binary + 5-class | 98.36% Acc (binary), 84.36% Acc (5-class) |
| [ 7 ] | DenseNet-169 + CBAM attention | Standard preprocessing | Binary + 5-class severity grading | 97% Acc (binary), 82% Acc (5-class), 88.80% QWK |
| [ 19 ] | RadFuse: non-linear RadEx (Radon) + fundus fusion | RadEx sinogram + fundus preprocessing | Binary + 5-class grading | 99.09% Acc (binary), 87.07% Acc, 93.24% QWK (5-class) |
| [ 21 ] | Deformable registration + multi-CNN voting ensemble | Deformable registration | 5-class grading | 85.28% Acc, 75% QWK |
Table 3: Survey of recent works in diabetic retinopathy detection – Part 2.
Figure 2: End-to-end analysis pipeline. Ordinal grade thresholds and the referral threshold must be fixed on validation data before locked evaluation.
Figure 3: Feature-enhanced images produced by the three-channel preprocessing pipeline. From left: original fundus, grayscale Ben Graham channel, Green-CLAHE channel, implemented dark-structure proxy, and composite input.
Figure 4: Feature-enhanced images produced by the three-channel preprocessing pipeline. From left: original fundus, grayscale Ben Graham channel, Green-CLAHE channel, implemented dark-structure proxy, and composite input.
| Step | Rationale | Parameter value |
|---|---|---|
| Ben Graham normalisation | Corrects global illumination to improve vessel visibility and cross-device consistency. | |
| Green-channel CLAHE | Enhances local contrast for micro-lesions and vessels in the most informative channel. | clipLimit : 2.0; tileGridSize : |
| Dark-structure proxy | Produces the implemented black-hat-equivalent response; no separate bright-lesion term is present. | weights: ; kernel: ; blur: |
| Disc mask + resize | Masks non-retinal background; standardises spatial dimensions. | Radius scale: 0.46; size: |
Table 4: Preprocessing pipeline parameters and rationale.
| Hyperparameter | Value |
|---|---|
| Backbone | EfficientNetV2-L ( tf_efficientnetv2_l.in21k_ft_in1k ) |
| Input size | (from preprocessed features) |
| Devices | Two NVIDIA Tesla T4 GPUs with DataParallel |
| Batch size | 16 per GPU (32 global) |
| Gradient accumulation | 4 steps (effective batch: 128) |
| Epochs | 20 |
Table 5: Training hyperparameters for the EfficientNetV2-L ordinal regressor.
| QWK score range | Interpretation |
|---|---|
| Worse than chance / systematic disagreement | |
| Slight agreement | |
| Fair agreement | |
| Moderate agreement | |
| Substantial agreement | |
| Almost-perfect agreement |
Table 6: Landis–Koch heuristic labels for kappa values. These descriptive bands are not clinical acceptance criteria.
| True grade | Support | Correct | Recall |
|---|---|---|---|
| No DR | 199 | 194 | 0.975 |
| Mild NPDR | 30 | 18 | 0.600 |
| Moderate NPDR | 87 | 50 | 0.575 |
| Severe NPDR | 17 | 12 | 0.706 |
| PDR | 33 | 12 | 0.364 |
Table 7: Exploratory per-grade results reconstructed from the current confusion matrix. Thresholds were selected using labels from the same analysis subset, so the values are not held-out estimates.
Figure 6: Exploratory confusion matrix on the 366-image analysis subset using analysis-selected thresholds. The associated QWK of 0.9131 is not a held-out estimate.
Figure 7: Exploratory raw regression-output distributions by true grade. The dashed thresholds were selected using the same analysis- subset labels and must not be interpreted as validation-fixed decision boundaries.
Figure 8: Exploratory grade composition at the post-hoc nominal 20% referral point. The threshold was derived from the same analysis subset and is not a locked deployment threshold.
Figure 9: Qualitative Grad-CAM maps. Within each pair, the left image is the saved preprocessed feature image and the right image is the overlay. Green labels denote a correct exploratory grade and red labels an incorrect grade. These maps are not lesion-level validation.