PaperlayerView exact revision on arXiv

Source-linked AI summary

QuantumBoostNet: A Hybrid Classical-Quantum Architecture for Enhanced Accuracy in Cardiac Ultrasound View Identification

Mihai Udrescu-Milosav, Stefan-Alexandru Jura, Mihai Udrescu, Gerhard-Paul Diller

arXiv:2608.27302v1cs.LG

TL;DR

Specialized cardiac-ultrasound view identification remains challenging for standard models, motivating methods that can handle noisy C/SHD images. QuantumBoostNet combines classical and quantum branches with staged training, and it outperforms competing models across cardiac-view and standard image-classification benchmarks.

  • Problem

    Standard models perform well on curated view-identification benchmarks but can be suboptimal on specialized C/SHD ultrasound images, where accurate views matter for interpretation and measurement.

  • Method

    QuantumBoostNet uses a classical backbone with classical and quantum heads, including a parametrized 10-qubit circuit and an adaptive two-phase training protocol.

  • Results

    QuantumBoostNet consistently surpasses competing classical and hybrid models across cardiac ultrasound view identification and established image-classification benchmarks.

  • Takeaways & Limitations

    The results provide empirical support for further development of structured hybrid classical-quantum training strategies in medical image analysis.

  • Takeaways & Limitations

    Hybrid models remain constrained by limited near-term qubits, which can force compression of high-dimensional classical features and potentially discard discriminative information.

Abstract

from arXiv · show

Accurate identification of the correct view or angle in cardiac ultrasound (echocardiogram) is a critical component of cardiologic imaging. This step is essential for precise anatomical interpretation, reliable measurement, and the reduction of clinical errors. Although computer vision has advanced significantly, most state-of-the-art models perform well on standard benchmarks but often yield suboptimal results in specialized medical imaging tasks due to the high level of noise present in the data. QuantumBoostNet, a hybrid classical-quantum architecture, is introduced to address these challenges. This model integrates a classical backbone with two heads: one classical and one quantum, with the quantum head implemented as a parametrized 10-qubit quantum circuit. Training occurs in two stages, with an adaptive transition between heads governed by a mixing parameter that monitors loss dynamics. Extensive experiments indicate that, despite the limited number of qubits that can be simulated, QuantumBoostNet consistently outperforms state-of-the-art classical and hybrid classical-quantum models in cardiac ultrasound view identification, achieving a relative improvement over the best competitor. QuantumBoostNet also demonstrates superior performance on established image classification benchmarks and exhibits robustness to noise. These findings support the continued development of hybrid classical-quantum models for specialized medical imaging applications.

I. INTRODUCTION

Correct echocardiographic view identification is essential for anatomical interpretation and quantitative measurement, yet specialized C/SHD ultrasound remains difficult for standard deep-learning models. The paper investigates whether hybrid classical-quantum methods can improve view identification and broader image classification.

  • Motivation: Correct imaging planes enable specific anatomical interpretation and quantitative measurements, including LV dimensions and volumes.Off-axis or misidentified views can cause anatomical misinterpretation and measurement error.
  • Motivation: View recognition routes echocardiographic images to downstream segmentation, chamber quantification, and disease-detection tasks.
  • Motivation: Speckle, low contrast, blurred boundaries, attenuation, operator variability, and degraded motion information challenge both interpretation and feature extraction.These limitations are especially harmful for precise border localization and temporal-tracking tasks.
  • Research gap: Standard deep-learning models perform well on curated view-identification data but can be suboptimal on C/SHD ultrasound images.
  • Objective: The paper asks whether hybrid classical-quantum methods improve C/SHD view-identification accuracy and image-classification accuracy more broadly.
  • Contributions: QuantumBoostNet combines a classical backbone with classical and quantum heads, while experiments evaluate it against competing models and on standard image-classification benchmarks.The quantum head is implemented as a parametrized 10-qubit circuit.

A. Classical state-of-the-art models

Classical and hybrid image-classification research spans residual backbones, quantum circuits, fusion architectures, transfer learning, and staged optimization. The literature also identifies information compression and circuit complexity as central constraints for hybrid models.

  • Classical backbones: Residual networks use skip connections to improve gradient propagation through deep architectures.This addresses the vanishing-gradient problem affecting conventional CNNs.
  • Classical backbones: Echocardiographic pipelines commonly use CNN backbones as feature extractors, including Inception-ResNet-v1 in EchoNet.
  • Quantum components: Parameterized quantum circuits encode classical data into quantum states, apply trainable unitary gates, and produce predictions through Pauli measurements.The parameter-shift rule provides analytic gradients compatible with automatic differentiation frameworks.
  • Hybrid integration: Hybrid architectures include sequential classical-to-quantum transfer, quanvolutional preprocessing, cross-attention fusion, and reuse of discarded qubit states.
  • Optimization: Two-phase optimization trains classical and quantum components in stages to improve hybrid optimization stability.This staged strategy has been studied in applications such as hate-speech detection.
  • Limitations and alternatives: Sequential hybrids face an information bottleneck because high-dimensional classical features must be compressed to fit limited near-term qubits.Parallel Hybrid Networks address this by processing inputs through quantum and classical branches whose outputs are combined with trainable weights.

III. QUANTUMBOOSTNET

QuantumBoostNet combines a shared ResNet-18-style feature extractor with parallel classical and quantum prediction paths. Its quantum path compresses 512-dimensional features into a 10-qubit variational circuit, while predictions can be fused through α and training transitions between phases.

  • Shared architecture: QuantumBoostNet uses a shared ResNet-18-style backbone for single-channel grayscale inputs and produces a common feature representation.The backbone contains four residual stages and feeds both prediction branches.
  • Quantum path: The shared feature vector f ∈R512 is projected into a 10-dimensional circuit-input vector before quantum processing.The projection uses learnable parameters followed by tanh squashing and scaling by π.
  • Classical path: A classical two-layer multilayer perceptron independently processes f using a 256-dimensional hidden representation, batch normalization, ReLU, and dropout with probability 0.3.Its output is a class-logit vector used either independently or in fusion.
  • Prediction fusion: Quantum and classical logits are fused as ŷ = α Q(f) + (1 −α) C(f), with α controlling each branch’s relative contribution.The fusion mechanism and α behavior vary across model variants.
  • Training protocol: Training uses a nominal budget T = T1 + T2 and transitions when validation accuracy plateaus or the Phase 1 budget T1 is exhausted, whichever occurs first.Plateau detection uses a validation-accuracy improvement below ϵ = 0.002 for two consecutive epochs after a 5-epoch warm-up.
  • Quantum path: The variational quantum circuit uses four layers of trainable single-qubit rotations and template-defined CNOT entanglers, then measures Pauli-Z expectation values.The circuit architecture is shared across QuantumBoostNet variants.

B. Variants description and implementation

QuantumBoostNet variants share the same backbone and prediction heads but differ in branch order, fusion timing, gate behavior, and freezing or detachment rules. V1 is quantum-first with learnable gating, V2 uses a rigid quantum-to-classical transition, and V3 is classical-first.

  • Variant design: All variants share structural components but differ in branch order, fusion activation, α parameterization, and whether the frozen branch remains in the computation graph.These implementation choices define their distinct staged training paradigms.
  • V1: V1 begins quantum-dominant with α0 ≈0.95, trains the quantum path first, then activates the classical head after transition.The classical branch subsequently serves to improve the representation learned during the quantum-first stage.
  • Training phases: Phase 1 and Phase 2 use different active parameter sets, with one branch frozen or inactive while the other branch is optimized.The exact parameter transitions depend on the variant’s training protocol.
  • V2: V2 trains a quantum classifier independently, then uses frozen quantum logits as an auxiliary signal while gradients update the classical branch and shared backbone.Its scalar mixing coefficient is schedule-controlled rather than learned by a gate.
  • V3: V3 reverses the order by initializing α0 ≈0.05, training the classical-dominant system first, and later activating the quantum path.The quantum branch is initially frozen and the gate remains learnable during the first phase.

A. Experimental environment and tested models

The experiments used a Python-based Jupyter and Conda environment with PennyLane, scikit-learn, PyTorch, and torchvision. View Classification experiments ran on a high-end consumer CPU-GPU workstation.

  • Software environment: Experiments used Python 3.12.12 with Conda, Jupyter Notebook, PennyLane 0.44.0, scikit-learn 1.8.0, torch 2.9.1, and torchvision 0.24.1.The software environment supported both classical and quantum model experiments.
  • Hardware environment: The View Classification experiments ran on an Intel Core i9-14900K, NVIDIA GeForce RTX 4070 Ti SUPER, and 64 GB of RAM.

1) Classical models:

The study benchmarks five classical architectures spanning lightweight CNNs, optimized CNNs, ResNet-based models, and a view-specialized CNN implementation.

  • Classical model benchmark: The benchmark includes ViewCNN, ViewOptimizedCNN, ViewResNet, ViewOptimizedResNet, and MadaniCNN.The models increase in architectural complexity and include normalization, dropout, residual backbones, and a view-specialized design.

2) Hybrid classical-quantum models:

The study compares six hybrid classical-quantum architectures that combine quantum circuits with different classical feature-extraction and prediction structures.

  • 2) Hybrid classical-quantum models:: Six hybrid classical-quantum model types represent progressively distinct architectural designs, including direct hybridization and architecture-level hybrids.The comparison adapts existing classical models and introduces more structurally advanced hybrid designs.
  • 2) Hybrid classical-quantum models:: DualPathQuantumNet uses parallel classical and quantum feature-extraction paths whose outputs are concatenated before a shared classifier.It also uses focal loss and cosine-annealing scheduling.
  • 2) Hybrid classical-quantum models:: PHN processes CNN-derived features through parallel variational quantum-circuit and multilayer-perceptron branches, then fuses outputs with trainable scalar weights.PHN adapts the proposed Parallel Hybrid Network for image input by adding a lightweight CNN extractor.

B. Benchmarks and datasets

The evaluation uses an echocardiographic dataset focused on congenital or structural heart disease, alongside FashionMNIST and MNIST benchmarks under multiple experimental settings.

  • B. Benchmarks and datasets: The cardiac dataset extends prior work on view classification in altered cardiac anatomy and was selected because state-of-the-art methods found it challenging.It was collected at University Hospital Muenster with local ethics approval.
  • B. Benchmarks and datasets: The source cohort included 262 patients with congenital or structural heart disease and 62 structurally normal controls, spanning heterogeneous cardiac pathologies.The C/SHD cohort was deliberately heterogeneous.
  • B. Benchmarks and datasets: The dataset was reduced from 23 to 17 view classes because rarer views lacked enough examples for robust learning.The retained classes cover major acquisition windows in comprehensive transthoracic echocardiography.
  • B. Benchmarks and datasets: The refined cardiac dataset contains 14 labeled view classes and 210,345 grayscale frames, split at the patient level into training/validation and testing.Frames were resized to 224 × 224 pixels; 175,146 were used for training and validation and 35,199 for testing.
  • B. Benchmarks and datasets: FashionMNIST and MNIST provide standard image-classification benchmarks, while settings S1–S4 evaluate clean cardiac data, gate-level noise, and the two benchmark datasets.Except for S2, models use 10-fold cross-validation and 15 training epochs.

D. Performance comparison

On the cardiac view-classification task, QuantumBoostNet variants achieve the strongest overall accuracy, with V1 leading while maintaining a favorable computational position among hybrid models.

  • D. Performance comparison: 75.95% accuracy makes ViewResNet the strongest classical baseline, while ResNet-backed hybrids approach but do not exceed it.ViewQResNet, ViewOptimizedQResNet, and DualPathQuantumNet reach 73.23%, 72.94%, and 74.21%, respectively.
  • D. Performance comparison: 77.19 ± 0.89% accuracy is achieved by QuantumBoostNet V1, a +1.24 percentage point improvement over the best classical model, ViewResNet.All three QuantumBoostNet variants occupy the top three positions.
  • D. Performance comparison: 410.9–429.3 seconds per epoch places QuantumBoostNet at the lower end of the hybrid training-time range while delivering the highest accuracy.Classical models require 93–253.4 seconds per epoch, compared with 410.3–515.8 seconds for hybrid models.
  • D. Performance comparison: Narrow interquartile ranges for QuantumBoostNet V1 and V3 indicate consistent fold-wise accuracy, whereas ViewQNN and ViewOptimizedQNN show higher variance.The distributions are reported for 10-fold cross-validation in S1.
  • D. Performance comparison: QuantumBoostNet V1 leads accuracy at 77.19% ± 0.89%, macro recall at 70.8%, macro F1 at 71.8%, weighted precision at 76.8%, and weighted F1 at 76.7%.DualPathQuantumNet instead achieves the highest macro precision at 78.3%.

2) Results for setting S2:

Under noisy circuit conditions and across standard benchmarks, QuantumBoostNet variants remain competitive, with V1 and V3 showing the strongest reported performance. The results also indicate that architecture choice and statistical testing matter when interpreting hybrid-model gains.

  • Noise robustness: 77.75% and 77.63% are the highest average accuracies across noise channels, achieved by QuantumBoostNet V3 and V1, respectively.The noise experiments use reduced 4-qubit, 2-layer circuits, with single training runs per entry.
  • Noise robustness: QuantumBoostNet V2 recovers from a weaker quantum head through assistance from its classical head.V2’s quantum head reaches 73.47%, while the reported interpretation attributes recovery to the classical branch.
  • Benchmark transfer: 99.61 ± 0.04% is QuantumBoostNet V3’s highest MNIST accuracy, outperforming all classical models.On FashionMNIST, V3 also ranks first at 93.96 ± 0.24%.
  • Benchmark transfer: Across S1, S3, and S4, the QuantumBoostNet family consistently occupies the top positions, while simple hybrid models rank among the weakest.V3 and V1 alternate between first and second place, whereas ResNet-backed hybrids remain competitive.
  • Statistical analysis: The statistical analysis compares the best classical model with the best QuantumBoostNet variant on matched cross-validation folds using paired and rank-based tests.Effect sizes are reported with paired Cohen’s d and rank-biserial correlation r.

4) Statistical analysis:

Statistical analyses place QuantumBoostNet V3 and V1 among the strongest models across datasets, while showing that significance varies by task. The conclusions emphasize structured training and strong classical backbones, but the study remains limited to simulated quantum circuits.

  • Paired comparisons: +0.89 pp is QuantumBoostNet V3’s mean advantage over ViewOptimizedResNet on FashionMNIST, with both tests confirming significance.The paired Cohen’s d is 2.79.
  • Paired comparisons: +1.24 pp is QuantumBoostNet V1’s mean improvement over ViewResNet on View Classification, but neither test reaches significance at 0.05.The effect size is medium, d = 0.65, consistent with higher fold-to-fold variance on the challenging task.
  • Omnibus testing: p-value < 10^-18 in all three datasets rejects the null hypothesis that all 14 models perform equally under the Friedman test.The test treats models as treatments and folds as blocks.
  • Global ordering: QuantumBoostNet V3 has the best mean rank across all 30 evaluations at 1.63, followed by V1 at 1.97.The next listed models are ViewOptimizedResNet at 4.07, ViewResNet at 4.40, and V2 at 4.97.
  • Conclusions: 77.19% versus 75.95% for the leading competitor corresponds to a 1.63% relative improvement in cardiac ultrasound view identification.The reported improvement is described as consistent across folds and benchmarks.
  • Conclusions: The results suggest that strong classical feature extractors are a prerequisite for effective hybridization.The paper identifies backbone expressive power as the primary bottleneck and recommends deploying quantum heads atop strong representations.
  • Conclusions: The two-phase protocol separates branch specialization before fusion and dynamically switches phases, while the classical head provides a fallback under quantum gate errors.The paper states that this mitigates gradient interference and co-adaptation issues.
Loading 2608.27302v1…