💻 computer science

Feature‑Adaptive Fusion in Hybrid Quantum–Classical Neural Networks for Robust Biomedical Image Classification

This paper proposes a Feature-Adaptive Fusion Hybrid Quantum-Classical Neural Network (FAF-HQNN) that dynamically weights classical and quantum predictions to achieve superior accuracy and robustness against noise in biomedical image classification tasks.

Original authors: Yan-Yan Hou, Jian Li, Chongqiang Ye, Hengji Li, Zhuo Wang, Qinghui Liu

Published 2026-07-09

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Yan-Yan Hou, Jian Li, Chongqiang Ye, Hengji Li, Zhuo Wang, Qinghui Liu

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

The Big Picture: A Team of Two Doctors

Imagine you are trying to diagnose a patient based on a medical image (like a picture of tissue or blood cells). In the world of Artificial Intelligence (AI), we usually use "Classical" computers to do this. They are like highly trained, experienced doctors who have seen millions of images. They are great, but sometimes they get confused if the image is blurry, noisy, or has strange artifacts (like a speck of dust on the lens).

Recently, scientists have started looking at "Quantum" computers. Think of these as a new, mysterious type of doctor who sees the world in a completely different way. They can spot patterns that the classical doctor might miss. However, quantum computers right now are like a medical student: they are powerful but still a bit unstable, prone to making mistakes, and can't handle a whole patient file on their own yet.

The Problem:
Most current attempts to combine these two doctors (a "Hybrid" model) are rigid. They say, "We will always listen to the classical doctor 50% of the time and the quantum doctor 50% of the time," or "We will just let the classical doctor make the final call." This doesn't work well when the image is messy. Sometimes the classical doctor is right, and sometimes the quantum student sees something the expert missed. A fixed rule can't adapt to the situation.

The Solution: The "Feature-Adaptive Fusion" (FAF-HQNN)
The authors of this paper propose a new system called FAF-HQNN. Imagine a smart Triage Nurse standing between the two doctors.

The Setup: The medical image is first looked at by the "Classical Doctor" (a deep neural network) to get a general understanding.
The Quantum Twist: That understanding is then passed to the "Quantum Student" (a Variational Quantum Circuit) to see if they can find any hidden, weird patterns.
The Smart Nurse (The Innovation): Instead of a fixed rule, the system has a special "Fusion Mechanism." This is like a smart nurse who looks at the specific image and asks: "Is this image clear? Is it noisy? Does the quantum student see something useful here?"
- If the image is clear, the nurse might trust the Classical Doctor more.
- If the image is noisy or weird, the nurse might lean more on the Quantum Student's unique perspective.
- The nurse dynamically adjusts the "voting weight" for every single image and every single disease category.

How They Tested It

The researchers tested this new "Team" on two famous medical image datasets:

PathMNIST: Pictures of tissue samples (like looking at a slide under a microscope).
BloodMNIST: Pictures of blood cells.

They didn't just test them on perfect, clean pictures. They also tested them on "corrupted" pictures—images with added Gaussian noise (like static on an old TV), Salt-and-Pepper noise (random black and white specks), and Poisson noise (like graininess in low light).

What They Found

The Quantum Student Alone is Weak: If you let the Quantum Student work alone, they performed poorly and were very unstable. They need the Classical Doctor to guide them.
The Rigid Team is Okay: Combining them with a fixed rule (like a 50/50 split) was better than using just the Classical Doctor, but not perfect.
The Smart Nurse Wins: The FAF-HQNN (with the adaptive nurse) was the best performer.
- On Clean Images: It was the most accurate.
- On Noisy Images: It was the most robust. When the images got messy, the system knew to shift its trust to whichever "doctor" was handling that specific type of noise better.

The "Shallow Circuit" Surprise

The researchers also tried making the Quantum Student's brain more complex (deeper circuits). Surprisingly, simpler was better. A very shallow, simple quantum circuit worked just as well as, or better than, a complex one. This suggests that for these tasks, the quantum part doesn't need to be a super-complex machine; it just needs to provide a little bit of "extra flavor" to the mix.

The Bottom Line

This paper claims that for classifying biomedical images, the best approach isn't to replace our current AI with quantum computers. Instead, it's to build a smart partnership where a classical AI and a quantum AI work together, but with a flexible manager that decides who to listen to based on the specific image. This makes the system more accurate and much harder to fool by bad image quality.

Important Note: The authors explicitly state that these results were achieved using simulations on classical computers, not on real quantum hardware. They also tested on standardized, small-scale datasets (MedMNIST), not on massive, real-world hospital databases yet. The paper claims this method works well for these specific benchmarks, but it does not claim it is ready for immediate use in a real hospital.

Technical Summary: Feature-Adaptive Fusion in Hybrid Quantum–Classical Neural Networks for Robust Biomedical Image Classification

Problem Statement
Hybrid quantum-classical neural networks (HQNNs) represent a practical paradigm for quantum machine learning within the noisy intermediate-scale quantum (NISQ) era, where hardware constraints limit qubit counts and circuit depth. While existing hybrid models often combine classical deep learning with variational quantum circuits (VQCs), they typically rely on fixed or globally shared fusion strategies (e.g., sequential processing or fixed-weight averaging). These static integration methods may fail to fully exploit the complementary information provided by quantum branches, particularly when input data undergoes distribution shifts or corruptions. In biomedical image classification, where images are frequently affected by sensor noise (e.g., Gaussian, Poisson) and acquisition variability, the robustness of these hybrid models under realistic corruption conditions remains insufficiently understood. Furthermore, existing architectures often assign predetermined roles to classical and quantum branches, limiting their adaptability to varying image content and corruption severities.

Methodology
The authors propose the Feature-Adaptive Fusion Hybrid Quantum-Classical Neural Network (FAF-HQNN), an end-to-end trainable framework designed for biomedical image classification. The architecture consists of four main components:

Feature Extraction: A classical deep feature encoder based on ResNet-18 (initialized with ImageNet weights) extracts high-level visual representations from input images. The final fully connected layer is removed, and global average pooling is applied to generate a 512-dimensional feature vector.
Feature Compression: To address the dimensional mismatch between deep features and NISQ constraints, a learnable compression module projects the 512-dimensional vector into a compact 64-dimensional shared embedding. This embedding is rescaled to the interval $[-\pi, \pi]$ via a hyperbolic tangent activation to ensure compatibility with parameterized quantum gates.
Parallel Hybrid Prediction: The shared embedding is fed simultaneously into two parallel branches:
- Quantum Branch: A variational quantum circuit (VQC) performs classical-to-quantum encoding (using Hadamard, $R_y$ , and $R_z$ gates), variational transformation (trainable single-qubit rotations and nearest-neighbor CNOT entanglement), and measurement-based readout (expectation values of Pauli-X observables). The output is mapped to class logits via a linear layer.
- Classical Branch: A lightweight multilayer perceptron (MLP) processes the same shared embedding to produce classical class logits.
Feature-Adaptive Class-Wise Fusion: A key innovation is a lightweight fusion coefficient predictor (a two-layer MLP with sigmoid activation) that generates sample-dependent, class-wise fusion weights ( $\alpha_i \in [0, 1]^C$ ) from the shared embedding. The final prediction is a dynamic weighted sum of the classical and quantum logits: $\hat{y}_i = \alpha_i \odot y^c_i + (1 - \alpha_i) \odot y^q_i$ . This allows the model to dynamically balance branch contributions based on the specific input sample and class.

The model is trained using standard cross-entropy loss on the fused logits.

Key Contributions

Architecture Proposal: Introduction of FAF-HQNN, which integrates a deep classical encoder with a VQC branch via a feature-adaptive fusion mechanism, moving beyond fixed integration schemes.
Adaptive Fusion Mechanism: Development of a class-wise fusion strategy that predicts sample-dependent weights, enabling the model to flexibly exploit complementary classical and quantum information under varying input conditions.
Robustness Evaluation: Systematic assessment of hybrid models under realistic image corruptions (Gaussian, salt-and-pepper, and Poisson noise) on biomedical benchmarks, addressing a gap in current literature where robustness is often overlooked.
Empirical Analysis: Comprehensive investigation into the impact of VQC design choices (circuit layout, measurement basis, and depth) on hybrid performance.

Experimental Results
The framework was evaluated on two MedMNIST benchmarks: PathMNIST (histopathology) and BloodMNIST (blood cell recognition). Experiments included clean data evaluation and robustness testing under varying corruption intensities.

Clean Data Performance: FAF-HQNN achieved the highest overall performance among all compared methods (including purely classical, purely quantum, and fixed-fusion hybrid baselines) on both datasets. On PathMNIST, it achieved an accuracy of $0.8910 \pm 0.0149$ , outperforming the classical baseline by 2.63 percentage points.
Robustness: Under Gaussian and Poisson noise, FAF-HQNN consistently demonstrated superior robustness, maintaining higher accuracy and AUC than single-branch baselines and fixed-fusion hybrids across most corruption levels. Under salt-and-pepper noise, while FAF-HQNN performed best at low-to-moderate corruption, a scalar-weight fusion baseline (H-Scalar) showed competitive or superior performance at severe corruption levels, indicating that the optimal fusion strategy may vary with noise severity.
Quantum Module Analysis:
- Circuit Depth: Shallow circuits (single-layer) yielded the best performance; increasing depth to 2 or 3 layers degraded results, suggesting that deeper circuits introduce optimization challenges without adding necessary expressivity for these tasks.
- Measurement Basis: No single measurement basis (Pauli-X, Y, Z) or combined basis (XYZ) consistently outperformed others across all circuit layouts, though Circuit 1 with Pauli-X measurement achieved the best overall result.
- Stability: The purely quantum branch (HQNN) exhibited the lowest performance and highest variance, confirming that the quantum component serves best as a complementary predictor rather than a standalone classifier.

Significance and Claims
The paper claims that feature-adaptive fusion is an effective strategy for improving both accuracy and robustness in hybrid quantum-classical models for biomedical image classification. The results suggest that while purely quantum models are currently insufficient for standalone high-dimensional tasks, quantum branches can provide useful complementary information when integrated with classical encoders. The adaptive fusion mechanism allows the model to dynamically leverage these complementary strengths, particularly under corrupted input conditions.

The authors maintain a modest stance, noting that the findings are based on simulation (PennyLane with default.qubit) and controlled MedMNIST benchmarks. They acknowledge that the results do not yet reflect hardware-specific effects (e.g., gate noise, decoherence) and that the observed advantages need validation on larger, higher-resolution clinical datasets and real NISQ hardware. The study concludes that the practical value of hybrid learning in the NISQ era may lie less in replacing classical deep learning and more in enriching representation diversity and improving robustness through adaptive integration.