In a paper published in the journal Scientific Reports, researchers proposed a hybrid quantum-classical 3D convolutional neural network (CNN) framework to improve blood flow velocity prediction using laser speckle contrast imaging (LCSI).
Study: Quantum machine learning enhanced laser speckle analysis for precise speed prediction. Image Credit: Aleksandra Budnik/Shutterstock.com
By integrating variational quantum circuits in place of traditional pooling layers, the model preserved 3D spatiotemporal features, enhancing prediction accuracy. Cross-validation on experimental data demonstrated improvements in prediction accuracy and error metrics compared to classical models. The results showed the hybrid model's superior learning stability and generalization capabilities for low and high blood flow velocities.
Background
Past work has demonstrated the potential of integrating machine learning with LSCI to enhance blood flow assessments. Still, classical 3D CNNs often face challenges due to information loss during feature pooling. This loss arises because global pooling compresses feature maps into single values, discarding crucial spatiotemporal correlations. Additionally, the high dimensionality of extracted features often leads to overfitting, especially with limited training data.
Quantum-Enhanced Blood Flow Prediction
The study employs a tissue phantom and a full-field LSCI system to simulate and measure blood flow. The phantom mimics the light scattering of human tissue, while a servo-controlled scattering plate emulates red blood cell motion at various speeds.
The LSCI system captures speckle patterns generated by this motion using a 532 nm green laser, a complementary metal-oxide-semiconductor (CMOS) sensor, and a field-programmable gate array (FPGA) platform for processing. Data augmentation was performed on the captured frames, increasing training data for subsequent analysis.
The research introduces a quantum-classical hybrid neural network framework to improve predictive performance. Variational quantum algorithms (VQAs) replace global pooling layers in traditional 3D CNNs, preserving spatial and temporal information. The quantum layer employs amplitude encoding to handle high-dimensional feature vectors efficiently, reducing data loss compared to classical methods.
This approach is applied to CNN-LSCI and ResNet-LSCI models, resulting in hybrid versions (quantum CNN (QCNN)-LSCI and quantum residual network (QResNet)-LSCI). The integration of quantum layers addresses the limitations of global pooling, offering potential enhancements in tasks requiring detailed spatial-temporal analysis.
The hybrid quantum-classical QCNN-LSCI and QResNet-LSCI frameworks address the limitations of traditional 3D CNNs by replacing global pooling with a variational quantum circuit, preserving spatial and temporal features. Amplitude encoding optimizes the quantum framework, handling high-dimensional data with fewer qubits. This quantum integration boosts predictive performance, especially for complex spatial and temporal pattern tasks.
Moreover, the proposed quantum-enhanced frameworks exhibit adaptability across a range of computational tasks, as the structure of the parameterized layer in VQCs can be tailored to specific data requirements. This flexibility enables QCNN-LSCI and QResNet-LSCI models to outperform their classical counterparts regarding predictive accuracy and data efficiency.
Incorporating quantum computing techniques in this domain signifies a pioneering step toward leveraging quantum advantages for practical applications, providing a robust platform for future research in medical imaging, dynamic tissue analysis, and other areas reliant on precise data-driven predictions.
Performance Comparison of Models
The experiments were conducted with classical and quantum models, specifically CNN-LSCI, ResNet-LSCI, QCNN-LSCI, and QResNet-LSCI. The models were evaluated on their ability to predict blood flow velocity. Quantum layers were implemented using PennyLane, while classical layers were built with PyTorch. PennyLane's default qubit simulator served as the backend for executing quantum circuits. Both backpropagation and adjoint differentiation techniques were utilized to compute quantum gradients in the VQC.
To evaluate model performance, k-fold cross-validation was employed, where the dataset was split into training, validation, and test sets. The dataset contained 1440 samples, with 922 for training, 230 for validation, and 288 for testing. The models were trained for 100 epochs with a batch size of 16, using the Adam optimizer and a learning rate of 0.001.
Mean squared error (MSE) and mean absolute percentage error (MAPE) were used as evaluation metrics. The results from the 5-fold cross-validation demonstrated that the quantum models consistently outperformed their classical counterparts in both training and validation phases.
The quantum models, QCNN-LSCI and QResNet-LSCI, exhibited lower and more stable losses than CNN-LSCI and ResNet-LSCI. Furthermore, the quantum models showed significant improvements in generalization, reducing test loss and MAPE, with QCNN-LSCI achieving 14.8% and 26.1% improvement over CNN-LSCI and QResNet-LSCI showing 5.7% and 18.5% improvement over ResNet-LSCI.
The quantum models outperformed classical models, especially at high blood flow velocities, where the classical models underestimated speeds. Despite data scarcity, the quantum models accurately predicted velocities due to superior learning and generalization. Additionally, quantum models excelled in detecting subtle speckle patterns, particularly for low-speed blood flow, overcoming noise and capturing fine details, while classical models struggled with low-speed predictions.
Conclusion
To sum up, this study proposed a hybrid quantum-classical 3D CNN framework for predicting phantom velocity, which improved prediction accuracy by avoiding global pooling information loss. The quantum models demonstrated superior prediction accuracy, learning stability, and generalization, particularly at extreme blood flow velocities.
Despite computational constraints, the framework showed significant promise and can be extended to other LSCI applications and beyond. Future work will focus on validating the framework with in vivo experiments, increasing data frames, and exploring other quantum circuit architectures for improved performance.
Journal Reference
Chen, Y., et al. (2024). Quantum machine learning enhanced laser speckle analysis for precise speed prediction. Scientific Reports, 14:1, 1-17. DOI: 10.1038/s41598-024-78884-4, https://www.nature.com/articles/s41598-024-78884-4
Disclaimer: The views expressed here are those of the author expressed in their private capacity and do not necessarily represent the views of AZoM.com Limited T/A AZoNetwork the owner and operator of this website. This disclaimer forms part of the Terms and conditions of use of this website.