Balanced Few-Shot Learning Accurately Diagnoses Retinal Disease With Limited Data, Addressing Imbalance In 1,000 Test Images
Understand the challenge: automatically diagnosing retinal diseases like diabetic retinopathy and macular degeneration is increasingly crucial as these conditions become more common. Yet many deep learning approaches falter when training data is scarce or imbalanced. This work from Jasmaine Khale (Northeastern University–Silicon Valley) and Ravi Prakash Srivastava (Indian Institute of Information Technology, Ranchi) tackles that gap by crafting a few-shot learning strategy that works reliably from only a handful of labelled images per disease. The result is a balanced framework that ensures all disease categories contribute equally during learning, mitigating bias toward more prevalent conditions. By pairing balanced sampling with targeted image augmentation and a strong image analysis backbone, the researchers report notable gains in diagnostic accuracy and a move toward fairer, more robust clinical tools for detecting a broad spectrum of retinal diseases.
Why this matters: collecting large, annotated medical image datasets is often costly and time-consuming. The team therefore developed a method that can generalize from very few examples per disease. At the core is a Prototypical Network, which learns to map images into a space where similar diseases cluster together while dissimilar ones are separated. New images are classified by comparing them to disease prototypes—representative examples for each condition.
Training that mirrors real life: the training process simulates real-world scenarios where only a few instances exist per disease. Episodes—mini-training tasks—teach the model to generalize from limited data and avoid skew toward common conditions. A key innovation is balanced episodic training, which enforces equal representation of each disease within every episode to prevent bias toward frequent conditions. The team also uses data augmentation, including Contrast Limited Adaptive Histogram Equalization (CLAHE), to boost image contrast and emphasize subtle features that signal disease.
What the results show: the approach improves retinal disease classification performance in settings with limited data compared with standard methods. Balanced episodic training helps prevent bias and enhances recognition of rarer diseases, while CLAHE improves image quality and feature extraction. While differentiating visually similar conditions remains challenging, this work marks a meaningful advance toward clinically reliable diagnostic tools. It also highlights the value of applying Prototypical Networks to retinal disease classification and demonstrates how balanced episodic training can reduce bias toward common diseases. CLAHE as a targeted augmentation further strengthens the system’s ability to reveal fine retinal details.
Balanced Episodic Learning for Retinal Diagnosis
The study introduces a balanced few-shot episodic learning framework aimed at boosting both accuracy and fairness in automated retinal disease diagnosis. Acknowledging the limitations of traditional deep learning—especially the need for large labeled datasets—the researchers designed a system that can generalize from a small number of examples per disease category. The framework combines balanced episodic sampling with targeted data augmentation, specifically CLAHE, to increase the diversity and visibility of minority classes. The Retinal Fundus Multi-Disease Image Dataset (RFMiD) serves as the evaluation ground, addressing inherent class imbalances.
How the method works: during training, episodes present a scenario where classification relies on only a few examples per disease. Balanced episodic sampling ensures equal participation of all diseases within each episode, curbing the tendency to favor frequent conditions. To amplify minority-class representation, targeted augmentations—such as CLAHE and geometric transformations—are applied, expanding training image variety and clarifying subtle retinal features. A ResNet-50 encoder, pre-trained on ImageNet, captures these fine-grained details.
What the experiments reveal: substantial improvements in classification performance emerge from balanced episodic sampling, which preserves equal class representation during training and reduces bias toward prevalent diseases like Diabetic Retinopathy. Targeted augmentation, including CLAHE, further enhances the visibility of subtle retinal features, improving the model’s discrimination between closely related presentations. This line of work underscores the potential of dataset-aware, few-shot learning approaches for building robust, clinically fair diagnostic systems in ophthalmology.
Few-Shot Learning for Retinal Disease Diagnosis
A novel few-shot learning framework is presented to raise accuracy and fairness in automated retinal disease diagnosis. Recognizing the data demands of conventional deep learning, the team designs a system that can generalize from a small number of examples per disease category. The framework fuses balanced episodic sampling with targeted data augmentation—in particular CLAHE—to enrich minority-class diversity. The focus covers the ten most represented disease categories within RFMiD.
Core approach: training constructs episodes where the model learns from limited examples per disease. Balanced episodic sampling guarantees equal disease participation within each episode, preventing bias toward more frequent conditions. To bolster minority-class representation, CLAHE and geometric transformations augment the training data, highlighting subtle retinal features. A ResNet-50 encoder pre-trained on ImageNet captures the nuanced details necessary for fine-grained differentiation.
Results and implications: the experiments show meaningful gains in diagnostic accuracy, especially for underrepresented retinal diseases, and a reduction in bias toward common conditions. Coupled with a ResNet-50 backbone, the method yields a robust, clinically relevant system capable of performing well even with limited data.
For more details
👉 More information
🗞 Balanced Few-Shot Episodic Learning for Accurate Retinal Disease Diagnosis
🧠 ArXiv: https://arxiv.org/abs/2512.04967
