⚡ electrical engineering

FQPDR: Federated Quantum Neural Network for Privacy-preserving Early Detection of Diabetic Retinopathy

This paper proposes FQPDR, a federated quantum neural network framework that leverages privacy-preserving collaborative learning to achieve robust, lightweight early detection of diabetic retinopathy by identifying subtle microaneurysms without sharing sensitive patient data.

Original authors: Debashis De, Mahua Nandy Pal, Dipankar Hazra

Published 2026-05-12

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Debashis De, Mahua Nandy Pal, Dipankar Hazra

Original paper licensed under CC BY 4.0 (http://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

The Big Picture: Finding Tiny Spots Before They Grow

Imagine your eyes are like a high-resolution camera. Sometimes, people with diabetes get tiny, almost invisible red dots (called microaneurysms) on the back of their eye. These are the very first warning signs of a condition called Diabetic Retinopathy (DR). If you catch these dots early, you can save your sight. If you wait too long, the damage becomes permanent.

The problem is that these dots are tiny and faint. Finding them is like looking for a single grain of sand on a beach. Traditional computer programs (Classical AI) are good at this, but they are like giant, hungry monsters: they need massive amounts of data, huge computer power, and lots of time to learn. They also require hospitals to send all their private patient photos to a central server, which raises privacy concerns.

This paper proposes a new solution called FQPDR. It combines three advanced ideas to solve these problems: Quantum Computing, Federated Learning, and Neural Networks.

1. The "Quantum" Brain: A Super-Small, Super-Fast Detective

The Analogy: Imagine a traditional computer is like a librarian who reads books one by one to find a specific word. A Quantum Computer is like a librarian who can read every book in the library simultaneously because they exist in multiple states at once (a concept called superposition).

What the paper does:
Instead of using a massive, heavy computer brain, the authors built a Quantum Neural Network (QNN).

Lightweight: While a traditional AI might need millions of "settings" (parameters) to learn, this Quantum AI only needs 15 settings. It's like a detective who only needs a few clues to solve a complex mystery, whereas the old detective needs a whole file cabinet.
Efficient: Because it uses quantum mechanics, it can spot those tiny, faint red dots (microaneurysms) using very few training examples. It's like teaching a child to recognize a cat by showing them three pictures, whereas a traditional computer might need three thousand.

2. The "Federated" Team: Learning Without Sharing Secrets

The Analogy: Imagine a group of doctors in different hospitals who want to learn how to spot these eye dots better.

The Old Way (Centralized): They would all mail their patient files to one central office. The central office studies them and sends back the rules. The risk? If the mail truck gets hijacked, all patient secrets are stolen.
The New Way (Federated Learning): The doctors stay in their own hospitals. They train their own local "detective" on their own patients. Then, they only send the rules (the math settings) they learned to a central coordinator. The coordinator mixes these rules to create a "Super-Rule" and sends it back. No patient photos ever leave the hospital.

What the paper does:
The authors set up a system where three different "virtual" computers (clients) train their own Quantum models on local data. They share only the mathematical "weights" (the brain's settings) with a central server. This ensures that patient privacy is protected because the actual images never travel across the internet.

3. The Process: How It Works Step-by-Step

The paper describes a specific workflow to make this happen:

Cutting the Cake (Pre-processing): They take large retinal images and cut them into tiny slices (patches) to focus on the specific areas where the dots might be.
Translating to Quantum Language (Encoding): Computers speak in 0s and 1s. Quantum computers speak in "amplitudes." The authors used a method called Amplitude Encoding to translate the image pixels into quantum states. Think of this as translating a book from English into a secret quantum code that only the quantum computer can read.
The Training Game:
- The "Quantum Detective" looks at the data.
- It makes a guess.
- If it's wrong, it adjusts its 15 settings (using a smart math tool called the Adam Optimizer).
- This happens over and over (100 rounds or "epochs") until it gets really good.
The Team Huddle (Aggregation): The three local detectives send their 15 settings to the central server. The server averages them out (giving more weight to the detectives who did better) to create a smarter global detective. This global version is sent back to the local doctors to use again.

4. The Results: Did It Work?

The authors tested their system on three different sets of eye images (E-ophtha, Retina MNIST, and Kaggle).

Accuracy: The Quantum Federated system was very accurate, often reaching around 89-90% accuracy in detecting the disease.
Comparison: When compared to a traditional (Classical) AI, the Quantum version was much more efficient. The Classical AI needed 33 settings and 62 rounds of training to get a similar score. The Quantum version only needed 15 settings and 24 rounds.
Privacy: The system successfully kept the data local, proving that you can train a smart model without stealing patient privacy.
Doctor Check: A real ophthalmologist (eye doctor) reviewed the system's decisions and found them to be "inspiring" and consistent with the ground truth.

5. The Catch (Limitations)

The paper is honest about the hurdles:

Hardware: Real quantum computers are currently expensive, rare, and a bit "noisy" (like a radio with static). They don't have many "qubits" (quantum bits) yet.
Hybrid Nature: The system currently runs partly on classical computers and partly on quantum ones. It's not a fully standalone quantum machine yet.
Overhead: Translating images into quantum code takes some extra time and effort.

Summary

The paper presents a FQPDR system: a Federated Quantum Neural Network for Privacy-preserving Detection of Retinopathy.

Think of it as a team of spies (Federated Learning) who solve a mystery using superpowers (Quantum Computing). They solve the case faster and with fewer clues than the old detectives, and they do it without ever revealing their secret files to anyone else. It's a promising step toward a future where we can detect eye diseases early, securely, and efficiently.

Technical Summary: FQPDR – Federated Quantum Neural Network for Privacy-preserving Early Detection of Diabetic Retinopathy

1. Problem Statement

Diabetic Retinopathy (DR) is a leading cause of blindness among diabetic patients, with early detection being critical to prevent irreversible eye damage. The earliest sign of DR is the presence of microaneurysms (MA), which are tiny, low-contrast dots, making their detection a challenging task.

Current deep learning approaches, particularly Convolutional Neural Networks (CNNs), face significant limitations:

Resource Intensity: They require large numbers of learnable parameters and high memory, leading to high time complexity.
Data Privacy: Centralized training requires aggregating sensitive patient medical data on a central server, posing risks of data breaches and violating privacy regulations.
Data Scarcity: Acquiring sufficient annotated medical data is tedious, yet effective training often demands large datasets.

2. Methodology

The authors propose FQPDR, a Federated Quantum Neural Network (QNN) system designed for privacy-preserving early DR detection. The methodology integrates three core components:

2.1 Quantum Neural Network (QNN) Architecture

Instead of classical neurons, the model utilizes a Parameterized Quantum Circuit (PQC).

Data Encoding: The system employs Amplitude Encoding to convert classical image patches (size 7x6x3, totaling 126 pixels) into quantum states. Two zero bits are padded to reach 128 ( $2^7$ ) bits, allowing representation by a 7-qubit system.
Circuit Design: The authors propose a lightweight, multilayer Nearest Neighbor (NN) circuit topology. The circuit consists of alternating layers of $RY$ (rotation) gates and Controlled-X (CNOT) entanglement gates.
Parameters: The model is extremely lightweight, utilizing only 15 learnable parameters (14 rotation angles and 1 bias).
Optimization: The system evaluates Gradient Descent, Nesterov Momentum, and Adam optimizers. The Adam optimizer is selected as the most effective for training the QNN parameters.

2.2 Federated Learning (FL) Framework

To address privacy concerns, the QNN is trained within a Federated Learning architecture:

Decentralized Training: Local models are trained on edge devices (clients) using local patient data. Raw data never leaves the local device.
Aggregation: Only model parameters (weights) are shared with a central federated server.
Global Model: The server aggregates local models using a weighted average approach (weights assigned based on local model performance) to create a global model, which is then redistributed to clients for further training.
Privacy: This ensures that sensitive medical images remain on local nodes while benefiting from a diverse, global model.

2.3 Implementation Details

Datasets: The system was implemented and validated using three datasets with varying resolutions:
- E-ophtha: High-resolution fundus images (MA-affected and MA-free).
- Retina MNIST: Low-resolution (28x28) images.
- Kaggle: Large-scale dataset used for cross-validation.
Environment: Experiments were executed using PennyLane on Google Colaboratory and AWS Braket (simulating quantum hardware).
Task: The system focuses on detecting mild DR (microaneurysms) by classifying image patches as "affected" or "healthy."

3. Key Contributions

The paper outlines the following specific contributions:

Novel Architecture: Development of a Federated QNN system (FQPDR) for early DR detection using retinal images from E-ophtha, Kaggle, and MNIST datasets.
Privacy Preservation: Implementation of a federated learning framework to ensure medical data privacy, allowing model training without transferring raw patient data to a central repository.
Parameter Efficiency: Demonstration that a QNN with few parameters (15) can learn feature correlations effectively from limited samples, addressing the challenge of scarce annotated medical data.
Performance Analysis: A comprehensive comparison of the proposed FQPDR against existing non-federated and federated classical methods, as well as limited-parameter classical CNNs.

4. Experimental Results

The study evaluated the system using Accuracy, Precision, Recall, F1-Score, and Specificity.

Optimizer Performance: Among the tested optimizers, Adam yielded the highest performance. For the E-ophtha dataset using the Adam optimizer, the Federated QNN achieved a validation accuracy of 89.74%, outperforming the Non-FL QNN (88.46%) and the limited-parameter classical CNN (85.62%).
Convergence: The QNN model converged faster than the classical CNN, requiring fewer epochs (24 epochs for QNN vs. 62 for CNN) to achieve comparable or better accuracy.
Cross-Validation: The models trained on E-ophtha and Retina MNIST were cross-validated on the Kaggle dataset. The system demonstrated robustness in detecting mild DR, correctly identifying 17 out of 19 mild DR images in the Kaggle set using the E-ophtha trained model.
Comparison with Classical Methods:
- The proposed FQPDR uses only 15 parameters, whereas a comparable classical FL method (Jagan Mohan et al., 2023) required 968,005 parameters.
- Despite the drastic reduction in parameters, FQPDR achieved competitive accuracy (89.74% vs. 98.6% in the classical method, though the classical method used a much larger dataset and different preprocessing).
Qualitative Assessment: A registered ophthalmologist evaluated the system's predictions on test images. The results were described as "inspiring," with the system showing high capability in learning annotations from the dataset.

5. Significance and Claims

The authors claim that the FQPDR system offers a lightweight, secure, and intelligent solution for the Internet of Medical Things (IoMT).

Efficiency: By leveraging quantum superposition and entanglement, the QNN captures intricate patterns with significantly fewer parameters and less computational overhead compared to classical CNNs.
Scalability and Privacy: The federated approach allows for scalable training across distributed edge devices without compromising patient data privacy, making it suitable for remote healthcare applications.
Foundational Step: The paper positions this work as a foundational approach to retinal image classification using federated QNNs. It acknowledges current limitations, such as the high cost and limited accessibility of quantum hardware, noise in qubits, and the complexity of hybrid classical-quantum training. However, it asserts that as quantum hardware improves, this architecture holds significant potential for revolutionizing medical diagnostics.

The paper concludes that the integration of Federated Learning with Quantum Neural Networks provides a promising pathway for early, privacy-preserving detection of Diabetic Retinopathy, particularly in resource-constrained or privacy-sensitive environments.