hqnn-blood-cell-classification

name	hqnn-blood-cell-classification
description	Hybrid Quantum-Classical Neural Network (HQNN) methodology for medical image classification, specifically blood cell classification. Combines pre-trained classical backbone (ResNet-50) with variational quantum circuit for enhanced feature representation. Use when: (1) medical image classification with limited data, (2) hybrid quantum-classical ML pipeline design, (3) comparing quantum vs classical feature transformations, (4) NISQ-era quantum advantage in medical imaging. Activation: HQNN, hybrid quantum neural network, blood cell classification, quantum medical imaging, variational quantum circuit, ResNet quantum, quantum feature transformation
license	MIT
metadata	{"arxiv_id":"2605.23324","published":"2026-05-22","authors":"Guilherme Cruz, Nouhaila Innan, Alberto Marchisio, Gabriel Falcao, Muhammad Shafique","tags":["quantum","medical","HQNN","classification","blood-cells","variational-circuit"]}

HQNN Blood Cell Classification

Core Architecture

Modular HQNN architecture combining:

1. Classical backbone: Pre-trained ResNet-50 for feature extraction
2. Latent bottleneck: Low-dimensional classical projection (reduces to match qubit count)
3. Variational quantum circuit: Parameterized quantum gates for feature transformation
4. Classical output layer: Final classification

Key Design Patterns

Three-Model Comparison Protocol

To isolate quantum contribution, always evaluate three architectures:

1. HQNN: Classical backbone → bottleneck → VQC → classifier
2. Classical Matched: Classical backbone → bottleneck → classical nonlinear layer (same capacity as VQC) → classifier
3. Baseline: Classical backbone → classifier (no intermediate transformation)

This controls for model capacity and ensures quantum improvements are genuine.

Quantum Circuit Design

• Use angle encoding for classical-to-quantum data mapping
• Strongly entangling layers (e.g., RealAmplitudes or EfficientSU2)
• Keep circuit depth shallow for NISQ compatibility (2-4 layers)
• Measure all qubits in computational basis

Training Strategy

• Freeze backbone during initial VQC training
• End-to-end fine-tuning after VQC converges
• Use parameter-shift rule for quantum gradients
• Shot noise: 1024+ shots per circuit evaluation

Hardware Evaluation

• Test on both simulator (noiseless) and real quantum hardware
• Expect modest degradation on real hardware (~1-3% F1 drop)
• Report both to demonstrate noise robustness

Pitfalls

• Barren plateaus: VQCs with too many qubits/layers suffer from vanishing gradients. Keep qubit count ≤ 10 for medical classification tasks.
• Bottleneck dimension mismatch: Latent space dimension must exactly match qubit count. Use linear projection with appropriate activation.
• Simulation vs hardware gap: Simulators are orders of magnitude faster. Train on simulator, evaluate on hardware for final validation.
• Dataset saturation: On near-saturated benchmarks (e.g., 98%+ accuracy), even small F1 improvements (0.1-0.5%) are statistically meaningful.