💻 computer science

A Unified Deep Learning Cascade for Retinal Disease Detection and Diabetic Retinopathy Severity Grading

This paper proposes a unified two-stage deep learning cascade that utilizes an EfficientNetV2-S model with asymmetric loss to detect 28 retinal diseases from the RFMiD dataset and subsequently employs a Platt-calibrated uncertainty-aware gate to route diabetic retinopathy-positive cases to a second stage for ordinal severity grading on the APTOS-2019 dataset, achieving high performance while addressing the significant degradation in recognition accuracy caused by multi-disease co-occurrence.

Original authors: Arzav Saikia, Biswajit Tamuli, Bitopan Das, Manjeet Sarmah, Nomi Baruah, Surajit Dutta

Published 2026-06-29

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Arzav Saikia, Biswajit Tamuli, Bitopan Das, Manjeet Sarmah, Nomi Baruah, Surajit Dutta

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine your eye is like a busy airport, and the retina is the control tower. Sometimes, things go wrong in the tower—maybe a light flickers, a screen cracks, or a storm rolls in. Diabetic Retinopathy (DR) is like a specific, dangerous storm that can destroy the tower if not caught early.

The problem is that there are 28 different kinds of "storms" (diseases) that can happen, and they often happen at the same time. Plus, the "storm" of DR has different levels of intensity, from a light drizzle to a hurricane.

This paper presents a new two-stage "security checkpoint" system (a deep learning cascade) designed to scan these eye images, find the diseases, and grade the severity of the DR storm. Here is how it works, broken down into simple analogies:

Stage 1: The Wide-Angle Security Scanner

Think of the first stage as a super-fast security guard standing at the main entrance.

The Job: This guard looks at the eye image and checks for any of the 28 possible diseases. It's not just looking for the DR storm; it's looking for everything.
The Tool: The researchers used a smart AI model called EfficientNetV2-S. To make it really good at spotting rare diseases (which are like rare, tricky contraband), they taught it using a special rule called Asymmetric Loss. Think of this as training the guard to be extra careful not to miss a tiny, hidden item, even if it means they might flag a few harmless items by mistake.
The Result: This guard is excellent. It correctly identifies the presence of Diabetic Retinopathy 98.3% of the time.
The Catch: The researchers found that when a patient has multiple diseases at once (like a storm plus a cracked screen), the guard gets confused. Its accuracy drops by about 42% because the diseases "interfere" with each other, like too many people shouting at once in a crowded room.

The Gatekeeper: The "Platt-Calibrated" Switch

Once the first guard spots a potential DR storm, the image doesn't just go straight to the next step. It hits a smart gate.

The Problem: The first guard sometimes gets overconfident or underconfident about its own predictions.
The Fix: The researchers added a "calibration" step (Platt scaling). Imagine this is like a translator that takes the guard's raw, shaky confidence and turns it into a precise, reliable probability score. This ensures the gate only opens for the right reasons.

Stage 2: The Expert Storm Grader

If the gate opens (meaning the image is likely DR-positive), the image moves to the second stage.

The Job: This isn't just about finding the disease anymore; it's about grading the severity. Is it a light drizzle (Mild), a heavy rain (Moderate), or a hurricane (Severe)?
The Twist: The researchers realized that grading isn't just picking a category (like "Red" or "Blue"). It's an ordinal task. Mistaking a "Moderate" storm for a "Severe" one is a bigger mistake than mistaking "Moderate" for "Mild."
The Tools: They tested two different "expert graders":
1. EfficientNetV2-S: The all-rounder. It was the best overall, getting a score of 0.93 (out of 1.0) on how well it ranked the storms correctly.
2. Swin Transformer (Swin-S): The specialist. While slightly less accurate overall, it was surprisingly good at spotting the worst storms (Severe grades), acting like a specialist who focuses only on the most dangerous situations.
The Strategy: They used a "5-fold ensemble" method. Imagine having five different experts look at the same image and averaging their opinions. This made the final decision much more stable and reliable.

Why This "Cascade" Matters

The paper argues that doing everything in one giant model is messy.

The Analogy: If you ask one person to do everything (scan for 28 diseases and grade the severity of one specific disease), they get overwhelmed, especially when multiple diseases are present.
The Solution: By splitting the work, the first stage acts as a filter. It catches the DR cases and sends them to the specialized grader. This keeps the grader focused and efficient.
The "Uncertainty" Safety Net: The system has a special "fuzzy zone" (uncertainty-aware band). If the gate is unsure whether to send an image to the grader or not, it uses a smart formula to blend the two pieces of information rather than making a hard, potentially wrong, decision.

What They Found (The Results)

Performance: The system is very good at finding DR (98% accuracy) and grading it (0.93 score).
The "Co-occurrence" Discovery: They proved that when diseases happen together, the system gets confused. This is exactly why they needed the two-stage system: to handle the complexity of real-world eyes where multiple problems often exist.
Cross-Testing: They tested the system on a different dataset (RFMiD) than the one it was trained on (APTOS). Even though the "lighting" and "cameras" were different, the system still worked well at detecting the presence of DR, proving it's robust.

In Summary

This paper built a two-step AI pipeline for eye health.

Step 1: A broad scanner finds if any disease is present, with a special focus on spotting Diabetic Retinopathy.
Step 2: A specialized grader takes those specific cases and rates how bad the DR is, using a method that understands that "bad" is worse than "okay," which is worse than "fine."

The result is a system that is more accurate and efficient than trying to do everything in one go, especially when dealing with eyes that have multiple problems at once.

Technical Summary: A Unified Deep Learning Cascade for Retinal Disease Detection and Diabetic Retinopathy Severity Grading

Problem Statement
Diabetic Retinopathy (DR) is a leading cause of preventable blindness, necessitating reliable automated screening. Current deep learning literature typically addresses retinal disease detection and DR severity grading as isolated tasks. Furthermore, DR grading is often treated as a standard categorical classification problem, ignoring its ordinal nature (where misclassifying a severe case as mild is more detrimental than a neighboring error). Additionally, the presence of multiple co-occurring retinal diseases can degrade the performance of single-stage models, a factor often overlooked in unified frameworks.

Methodology
The authors propose a two-stage deep learning cascade designed to unify multi-label retinal disease screening with ordinal DR severity grading.

Stage 1: Multi-Label Retinal Disease Detection
- Dataset: RFMiD (3,200 images, 28 disease categories).
- Architecture: EfficientNetV2-S backbone.
- Training Strategy: The model is trained using Asymmetric Loss (ASL) to handle severe class imbalance and long-tail distributions common in multi-label datasets. Preprocessing includes Ben Graham contrast normalization and extensive augmentation (MixUp, CutMix).
- Output: The model predicts probabilities for 28 retinal diseases. The DR class specifically serves as the "gate signal" to determine if an image proceeds to Stage 2.
Gate Mechanism & Calibration
- Raw sigmoid outputs from the multi-label classifier are often miscalibrated. The authors apply Platt scaling to the DR posterior probabilities to ensure reliable decision gating.
- Cascade Variants: The study evaluates four variants:
  - C0: Stage 2 alone (baseline).
  - C1/C2: Hard gating using raw or calibrated probabilities, respectively.
  - C3 (Proposed): An uncertainty-aware bidirectional cascade. It uses two thresholds ( $\tau_{low}, \tau_{high}$ ) to create three regions: definite non-DR (output 0), definite DR (pass to Stage 2), and an ambiguous band. In the ambiguous band, a fused score combining the gate probability and the Stage 2 continuous score is used.
Stage 2: Ordinal DR Severity Grading
- Dataset: APTOS-2019 (3,662 images, 5 severity grades: 0–4).
- Approach: Formulated as ordinal regression rather than categorical classification to preserve severity relationships.
- Backbones: Two architectures are compared:
  - EfficientNetV2-S: Uses Mean Squared Error (MSE) loss.
  - Swin Transformer Small (Swin-S): Uses Weighted MSE (WMSE) to address class imbalance and leverages hierarchical shifted-window self-attention.
- Ensemble: A 5-fold stratified cross-validation ensemble is used to generate the final continuous scores, which are converted to discrete grades (0–4) using an OptimizedRounder based on Quadratic Weighted Kappa (QWK) optimization.

Key Contributions

Unified Cascade Framework: A novel two-stage pipeline that integrates multi-label disease detection with ordinal DR grading, routing only DR-positive samples to the severity grader.
Systematic Backbone Comparison: A detailed evaluation of EfficientNetV2-S versus Swin-S for ordinal regression, analyzing loss function sensitivity and class-weighting strategies.
Co-occurrence Degradation Analysis: A quantitative analysis demonstrating that simultaneous retinal pathologies degrade recognition performance (a 42% drop in F1-macro for multi-disease images), providing the empirical motivation for the cascade architecture and gate calibration.
Uncertainty-Aware Routing: The introduction of a bidirectional cascade (C3) that utilizes an ambiguity band to fuse evidence, reducing errors compared to hard-gating mechanisms.

Experimental Results

Stage 1 (RFMiD): EfficientNetV2-S with ASL achieved a mean AUC of 0.910 and mAP of 0.539. The DR channel specifically reached an AUC of 0.983, validating its reliability as a gate signal.
Co-occurrence Impact: Performance degraded significantly for images with 2–3 diseases (F1-macro dropped from 0.455 to 0.266). Conditional AUC analysis showed specific disease pairs (e.g., DR with Lattice Degeneration) caused significant performance drops, justifying the need for a dedicated grader.
Stage 2 (APTOS-2019):
- The EfficientNetV2-S 5-fold ensemble achieved the highest overall performance with a QWK of 0.9339.
- Swin-S achieved an out-of-fold QWK of 0.8963 but demonstrated superior sensitivity for predicting severe grades (Grade 3 and 4).
Cascade Evaluation:
- Calibration was critical: The calibrated hard gate (C2) outperformed the raw hard gate (C1) by +0.013 QWK.
- The uncertainty-aware cascade (C3) achieved a QWK of 0.9038 on the disjoint report slice, outperforming hard-gate variants and approaching the performance of the Stage 2-only baseline (C0) on in-distribution data.
- Error attribution revealed that only 4% of errors were due to gate failures, while 76% were due to Stage 2 magnitude errors (adjacent-grade discrimination), identifying the regressor as the primary bottleneck.
Cross-Domain Validation: When applied to the RFMiD test set (trained only on APTOS), the cascade maintained a high DR detection AUC of 0.981, demonstrating robustness to domain shift, though severity predictions were unvalidated due to lack of ground truth in RFMiD.

Significance and Claims
The paper claims that this unified cascade effectively addresses the gap between general retinal disease screening and specific DR severity grading. By analyzing disease co-occurrence, the authors demonstrate that a single-stage model struggles with multi-pathology inputs, motivating the cascade design. The study highlights that Platt calibration is a prerequisite for effective hard-gating and that uncertainty-aware routing can recover performance lost to gating mechanisms.

The authors note that while EfficientNetV2-S provides superior overall accuracy, Swin-S offers complementary strengths in detecting severe cases. The framework is presented as a step toward real-world ophthalmic screening workflows where multiple abnormalities frequently coexist. The paper modestly acknowledges limitations, including the lack of ground-truth severity labels in RFMiD for quantitative validation of Stage 2 on that dataset, the absence of interpretability visualizations (e.g., Grad-CAM), and the need for external validation on hospital-acquired datasets before clinical deployment.