Enhancing Blood Cells Classification using Hybrid Quantum Neural Networks

This paper proposes a Hybrid Quantum-Classical Neural Network (HQNN) architecture that integrates a pre-trained ResNet-50 backbone with a variational quantum circuit, demonstrating superior blood cell classification performance and robustness to noise compared to classical baselines on public datasets and IBM quantum hardware.

The Gist

Imagine you are a doctor trying to identify different types of tiny blood cells under a microscope. Some cells look almost identical—like twins wearing slightly different hats. Traditionally, computers (using "classical" deep learning) have been very good at this, but they sometimes get confused when the differences are extremely subtle.

This paper asks a big question: What if we gave the computer a "quantum superpower" to help it see these tiny differences better?

Here is the story of their experiment, explained simply:

1. The Setup: A Three-Legged Race

To make sure the test was fair, the researchers didn't just compare a "Quantum Computer" against a "Normal Computer." That wouldn't be fair because the quantum computer might just be bigger or have more parts.

Instead, they built three identical teams that were exactly the same in every way, except for one specific step in their thinking process:

• Team A (The Baseline): The computer looks at the cell, simplifies the picture, and immediately guesses the type.
• Team B (The Classical Matched): The computer looks at the cell, simplifies the picture, and then passes it through an extra "thinking layer" (a standard math layer) before guessing. This ensures that if Team B does better, it's just because it had more math, not because of magic.
• Team C (The Hybrid Quantum Team): The computer looks at the cell, simplifies the picture, and then passes it through a Quantum "Thinking Layer."

The Analogy: Imagine three students taking a test.

• Student A reads the question and writes the answer.
• Student B reads the question, thinks about it for 5 seconds using a standard calculator, and writes the answer.
• Student C reads the question, thinks about it for 5 seconds using a Quantum Calculator, and writes the answer.

The researchers wanted to see if the Quantum Calculator (Student C) actually helped solve the tricky parts better than the standard calculator (Student B).

2. The "Quantum Layer": A New Kind of Lens

How does the quantum part work?
Think of the computer's "brain" as a room where it organizes information.

• Classical computers organize data like books on a shelf: one book next to another.
• Quantum computers can organize data like a kaleidoscope. They can look at all the pieces of the image at once from many different angles simultaneously, thanks to a phenomenon called "entanglement."

In this study, the "Quantum Layer" acts like a special lens that takes the simplified picture of the blood cell and twists it into this kaleidoscope view. The hope is that this view makes the differences between "twin" cells (like Monocytes and Neutrophils) much clearer.

3. The Results: Who Won?

The researchers tested these teams on two different sets of blood cell images:

1. The "Easy" Set (4 types of cells): This is like distinguishing between a cat, a dog, a bird, and a fish.
2. The "Hard" Set (8 types of cells): This is like distinguishing between 8 different breeds of dogs that all look very similar.

The Findings:

• In the "Easy" Set: The Quantum Team (Team C) won clearly. It got about 3.7% more correct answers than the other teams. It was especially good at telling the tricky "twin" cells apart.
• In the "Hard" Set: Everyone was already doing a great job (almost perfect scores). However, the Quantum Team still managed to squeeze out a tiny bit more accuracy. It was the only team that didn't get stuck in a "tie" with the others; it kept improving slightly even when things were already near perfect.
• The "Real World" Test: The researchers also ran the Quantum Team on an actual physical quantum computer (made by IBM) instead of just a simulation.
  • The Catch: Real quantum computers are currently a bit "noisy" (like trying to hear a whisper in a windy room).
  • The Result: The performance dropped a little bit because of the noise, but the model was still robust. It didn't crash; it just got slightly less accurate. This proves the idea works even on real, imperfect hardware.

4. The Big Takeaway

The paper concludes that Quantum Machine Learning isn't just hype.

When the researchers compared the Quantum Team to the "Extra Thinking Layer" team (Team B), they found that the Quantum Team did better. This proves that the improvement wasn't just because they added more math; it was because the quantum math itself was better at spotting those tiny, subtle differences in blood cells.

In short: By using a quantum "lens" to look at blood cells, the computer became a better detective, especially when the suspects (the cells) looked almost exactly the same. This suggests that in the future, these hybrid systems could help doctors diagnose diseases faster and more accurately, particularly in the tricky cases where human eyes or standard computers might get confused.

Technical Summary

Technical Summary: Enhancing Blood Cells Classification using Hybrid Quantum Neural Networks

Problem Statement

Accurate classification of microscopic blood cells remains a critical challenge in medical image analysis. While Convolutional Neural Networks (CNNs) have largely replaced manual review, they face distinct limitations in fine-grained classification tasks. Specifically, CNNs struggle with subtle morphological differences, overlapping visual traits (e.g., distinguishing monocytes from neutrophils), and high intra-class variability. Furthermore, standard deep learning models are constrained by the mathematical limits of classical vector spaces, which may restrict their ability to map complex, non-linear correlations inherent in high-dimensional biomedical imagery. The authors posit that Hybrid Quantum-Classical Neural Networks (HQNNs) could leverage the high-dimensional Hilbert space and quantum entanglement to enhance feature representation and resolve these ambiguous cases.

Methodology

The study proposes a modular HQNN architecture designed to isolate the specific contribution of quantum feature transformations. The system follows a three-stage pipeline:

1. Feature Extraction: A pre-trained ResNet-50 backbone (initialized with ImageNet weights) extracts high-level visual features from input images (256x256x3). The final classification layer is removed, outputting a 2048-dimensional feature vector.
2. Dimensionality Reduction: A fully connected layer reduces the 2048-dimensional vector to a 10-dimensional latent space. A hyperbolic tangent ($tanh$) activation is applied to bound the values, facilitating angle-based quantum encoding.
3. Intermediate Transformation & Classification: This is the variable stage where three distinct architectures are compared:
   • HQNN Model: The 10-dimensional latent vector is encoded into a quantum state using Angle Embedding (rotations around the Y-axis). It is processed by a Variational Quantum Circuit (VQC) consisting of 10 qubits and 4 trainable layers. Each layer applies trainable $RY$ rotations followed by a ring-structured pattern of Controlled-NOT (CNOT) gates for entanglement. The output is measured via Pauli-Z expectation values, producing a 10-dimensional quantum feature vector.
   • Classical Matched Model: The quantum circuit is replaced by a classical nonlinear block (linear transformation + $tanh$) with a comparable parameter count (110 parameters) to the quantum layer (40 parameters).
   • Classical Baseline: The intermediate transformation stage is removed entirely, passing the latent vector directly to the classifier.

All models share an identical lightweight classification head (32 hidden units, ReLU, Dropout) and are trained using Focal Loss with label smoothing, Adam optimizer, and cosine annealing learning rate scheduling.

Key Contributions

The paper makes four primary contributions:

1. Controlled Comparison Framework: A modular architecture that strictly controls variables (backbone, bottleneck size, classifier head) to isolate the effect of the intermediate transformation stage (quantum vs. classical).
2. Fair Classical Baselines: Introduction of a "Classical Matched Model" with an additional nonlinear layer of comparable capacity to the quantum circuit, ensuring that performance gains are not merely attributed to increased parameter counts.
3. Real-World Hardware Validation: Execution of the HQNN model on IBM quantum hardware (specifically the ibm_fez backend) to assess robustness under realistic noise and hardware constraints, moving beyond software simulation.
4. Empirical Analysis of Quantum Advantage: Quantification of performance gains specifically in challenging classification scenarios characterized by high intra-class variability and subtle inter-class differences.

Experimental Results

The models were evaluated on two public datasets: the Blood Cell Images Dataset (4 classes) and the Peripheral Blood Cell (PBC) Dataset (8 classes).

• Blood Cell Images Dataset (4 Classes):

  • The HQNN model achieved a Macro F1-score of 0.9153, outperforming the Classical Matched Model (0.8862) and the Classical Baseline (0.8784).
  • This represents an improvement of approximately 3.7% over the baseline.
  • The HQNN showed the most significant gains in distinguishing difficult classes like neutrophils and monocytes, reducing confusion between visually similar cell types.

• PBC Dataset (8 Classes):

  • Performance was near-saturated for all models due to the dataset's characteristics.
  • The HQNN still achieved the highest metrics: Accuracy of 0.9967 and Macro F1-score of 0.9869, slightly outperforming the Classical Matched Model (0.9854) and Baseline (0.9851).
  • Improvements were observed in challenging classes such as immature granulocytes and neutrophils.

• Data Efficiency:

  • When trained on only 25% of the Blood Cell Images dataset, the HQNN maintained a clear advantage (94.55% accuracy, 89.33% F1) over classical counterparts, suggesting robustness in data-scarce scenarios.

• Hardware Validation:

  • On real IBM quantum hardware, the model experienced a modest performance degradation (~2.5–3%) compared to the noise-free simulator (Accuracy: 0.9000 vs. 0.9250).
  • Despite noise, the model remained robust, particularly for well-separated classes like lymphocytes, confirming the feasibility of HQNNs on current Noisy Intermediate-Scale Quantum (NISQ) devices.

Significance and Claims

The paper claims that its findings validate the tangible benefits of quantum feature transformations in medical imaging, specifically for resolving complex classification tasks where classical models struggle. By employing a rigorous, controlled experimental design, the authors demonstrate that the observed performance improvements stem from the distinct transformation properties of the quantum layer rather than simply the addition of more parameters.

The study concludes that while classical models approach performance ceilings in well-curated datasets, HQNNs offer consistent, measurable gains in class-wise balance and discriminative power, particularly in fine-grained scenarios. The successful deployment on real hardware further supports the practical potential of integrating quantum-enhanced learning into future clinical diagnostic tools, provided that hardware noise and scalability challenges are addressed.