💻 computer science

A Hybrid Quantum Neural Network - Assisted Deep Learning Framework for Automated Centralserous Retinopathy Detection Using Oct Images

This paper presents HQNNDL, a hybrid quantum-classical deep learning framework utilizing a frozen ResNet-18 backbone and an 8-qubit variational quantum circuit to classify OCT images into four categories, achieving 79.0% accuracy on a proxy dataset derived from Kermany OCT2017 while acknowledging the need for further ablation studies and clinical validation to confirm its efficacy for Central Serous Retinopathy detection.

Original authors: Annamnedi Govardhan, P Suresh Varma

Published 2026-07-23

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Annamnedi Govardhan, P Suresh Varma

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 the human eye as a high-definition camera that never stops taking pictures, capturing a world of light and color. But sometimes, a tiny, invisible leak develops inside this camera, letting watery fluid seep under the lens's most important part, the retina. This condition, called Central Serous Retinopathy (CSR), is like a slow-motion fog rolling in; patients might just feel a slight blur or a bend in their vision, often not noticing until the damage has already spread. To catch this leak early, doctors use a special scanner called Optical Coherence Tomography (OCT), which takes incredibly detailed cross-section slices of the eye, revealing the hidden fluid like a cross-section of a cake showing a layer of jelly.

The problem isn't taking the pictures; it's reading them. Today's eye clinics are so busy that they produce more scans than any team of doctors can carefully examine one by one. This has led scientists to build computer programs, or "AI," to help sort through the images. For years, these programs have gotten smarter, moving from simple math tricks to deep learning networks that can spot patterns on their own. But now, a new frontier is opening up: quantum computing. Think of a classical computer as a super-fast librarian who reads books one by one, while a quantum computer is like a librarian with a magical ability to arrange the books in a complex, multi-dimensional way that reveals hidden connections between stories a normal librarian might miss. The big question scientists are asking is: Can this "magical" way of organizing information actually help a computer diagnose eye diseases better than the best normal computers we have today?

This paper introduces a new system called HQNNDL, which tries to answer that question by building a hybrid team: a classic "deep learning" brain paired with a tiny, experimental "quantum" brain. The researchers didn't have a massive collection of real CSR eye scans to train on, so they used a clever workaround. They took a public dataset of eye images and used two other types of fluid-filled eye problems (DME and CNV) as stand-ins, or "proxies," for the mild and severe stages of CSR. It's like trying to learn how to fix a specific type of leaky pipe by practicing on two other types of leaks that look very similar.

The system works in a relay race. First, a classical computer network (ResNet-18), which has been trained on millions of regular photos, looks at the eye scan to pull out the most important features. Then, instead of passing those features to another standard computer layer, the system hands them off to a small, 8-qubit quantum circuit. This quantum circuit is like a special filter that rearranges the information in a complex, multi-dimensional way that normal computers find hard to do. Finally, a simple classifier makes the final guess: Is this a healthy eye, a mild case, a severe case, or something else entirely?

The results of this experiment are a mix of promise and honesty. On a test set of 800 images, the system got the diagnosis right about 79% of the time. It was very good at spotting healthy eyes (about 88% recall) but struggled a bit more with the "Other Pathology" category, getting that right only about 64% of the time. However, the most important finding isn't the score itself, but what the authors admit they didn't prove. Because the quantum part of the system was kept "frozen" (its settings were not adjusted or trained during the experiment), the researchers cannot say for sure if the quantum circuit actually helped the system get those 79% correct, or if it was just the classical part doing all the heavy lifting.

In fact, the paper explicitly rules out the idea that this is a finished, ready-to-use medical tool. The authors are very clear that they haven't isolated the quantum contribution yet; they don't know if the quantum part adds magic or just adds weight. They also note that the system is currently slow, taking about 475 milliseconds to process a single image, mostly because they are simulating the quantum computer on a regular laptop rather than using a real quantum machine. The system does produce extra helpful outputs, like a map showing where the problem is and a summary report, but these haven't been checked against expert doctors' notes either.

Ultimately, this paper is a "feasibility study," which is a fancy way of saying, "We built the whole machine, and it works, but we need to do more tests to see if the quantum part is actually the secret sauce." The authors are careful not to claim a breakthrough. Instead, they present a complete, working pipeline that goes from a raw image to a full report, proving that such a hybrid system can be built. They suggest that the next steps involve training the quantum part properly, testing it on real quantum hardware, and, most importantly, finding a dataset of actual, doctor-graded CSR scans to see if the system can truly diagnose the disease it was designed to find. Until then, the system remains a fascinating prototype, a bridge between today's AI and tomorrow's quantum possibilities, waiting for more evidence to cross the gap.

Technical Summary: HQNNDL Framework for Automated Central Serous Retinopathy Detection

Problem Statement
Central Serous Retinopathy (CSR) is a retinal disorder characterized by the accumulation of serous fluid beneath the neurosensory retina. While Optical Coherence Tomography (OCT) is the standard imaging modality for detection, the volume of scans generated by modern clinics exceeds the capacity of specialists for manual review. Existing automated diagnosis systems have evolved from classical image processing to deep learning, yet the authors identify a need to explore whether quantum processing stages can offer architectural advantages over classical layers of comparable size. A significant practical constraint for this study was the lack of a large, clinician-graded CSR dataset. Consequently, the authors faced the challenge of developing a detection system without direct ground-truth CSR labels, necessitating the use of structural proxies.

Methodology: The HQNNDL Framework
The paper proposes HQNNDL, a hybrid quantum-classical pipeline designed for four-class OCT image classification. The system architecture is divided into several distinct stages:

Input Validation: Before processing, an input-validation gate rejects non-retinal images (e.g., color photographs) by analyzing four properties: color-channel variance, horizontal layer banding, dark-region proportion, and intensity bimodality. Images scoring below a threshold of 0.42 are rejected.
Classical Feature Extraction: A ResNet-18 backbone, pre-trained on ImageNet and frozen (not fine-tuned), extracts a 512-dimensional feature vector from the validated OCT image. The authors explicitly note that this feature extraction must be applied to the original image to avoid distribution mismatches.
Quantum Feature Processing: The eight feature dimensions with the largest magnitude are amplitude-encoded into an 8-qubit register. These are processed by a Variational Quantum Circuit (VQC) consisting of three layers. Each layer applies single-qubit rotations ( $R_z$ and $R_x$ ) followed by a ring of CNOT entangling gates. The circuit contains 48 parameters. Crucially, these parameters are held fixed at their initial values; the VQC acts as a fixed non-linear transform rather than a jointly trained component. The output is an 8-dimensional vector of expectation values.
Hybrid Classification: The 8-dimensional quantum output is concatenated with the 512-dimensional classical features. This combined vector is passed through a fully connected network (512→256→64→4) with ReLU activations, dropout, and L2 regularization. This classifier is the only trained component of the pipeline.
Supporting Modules: The system includes auxiliary modules for biomarker scoring (calculating six structural scores like subretinal fluid extent), affected-area detection (generating saliency maps and bounding boxes), and automated report generation. These modules operate independently of the quantum stage.

Key Contributions
The authors position this work as a feasibility study and an honest first step toward testing hybrid architectures, rather than a finished clinical tool. The specific contributions are:

End-to-End Pipeline: A working system that processes raw OCT images through validation, feature extraction, quantum processing, classification, and report generation.
Transparent Evaluation: A detailed analysis of a held-out validation set, including per-class metrics and confusion pattern analysis, rather than just a headline accuracy number.
Auxiliary Outputs: The integration of biomarker scoring, localization, and report generation to simulate a practical screening workflow.
Explicit Limitations: A clear accounting of what the study does not establish, specifically the unverified contribution of the quantum stage due to the lack of joint training.

Results
The system was trained and evaluated on a stratified subsample of 4,000 images from the public OCT2017 dataset. Due to the lack of a dedicated CSR dataset, the authors used DME (Diabetic Macular Edema) and CNV (Choroidal Neovascularization) categories as structural proxies for CSR Grade I and Grade II, respectively. The "Other Pathology" class was represented by Drusen.

Performance: On a held-out validation set of 800 images (200 per class), the system achieved 79.0% accuracy with a macro F1-score of 0.787.
Class Performance: The "Normal" class achieved the highest recall (0.885). The "Other Pathology" class (Drusen proxy) showed the weakest recall (0.645), which the authors attribute to the subtle nature of Drusen deposits compared to the clear fluid buildup in the proxy classes, combined with the frozen backbone.
Latency: The total inference time per image on a CPU was approximately 475 ms, with the quantum simulation stage (NumPy-based) accounting for the majority of the time (~320 ms).

Significance and Claims
The paper maintains a modest and rigorous stance regarding its findings:

No Claimed Quantum Advantage: The authors explicitly state that the 79.0% accuracy is a property of the entire fixed pipeline, not evidence of a quantum advantage. Because the VQC parameters were not trained, the actual contribution of the quantum stage (whether it helps, harms, or is neutral) has not been isolated.
Feasibility over Clinical Validity: The work is framed as a proof-of-concept for the hybrid architecture. The authors emphasize that the proxy labeling scheme means the reported numbers cannot be read as validated CSR detection rates.
Architectural Curiosity: The primary motivation is to test whether a small quantum circuit can represent non-linear decision boundaries more compactly than a classical layer. The paper argues that answering this requires empirical testing, which this study initiates but does not fully resolve.
Future Requirements: The authors outline that establishing clinical validity or performance claims requires scaling the dataset, fine-tuning the CNN backbone, jointly training the quantum circuit, performing ablation studies against classical baselines, and validating auxiliary outputs against expert annotations.

In conclusion, the paper presents a complete, transparently evaluated hybrid pipeline that integrates quantum processing into a medical imaging workflow, while rigorously delineating the boundaries between its current experimental results and the requirements for future clinical deployment.

Technical Summary: HQNNDL Framework for Automated Central Serous Retinopathy Detection

More like this