A Fair Performance Comparison between Complex-Valued and Real-Valued Neural Networks for Disease Detection

Area of Science:

• Biomedical Engineering
• Artificial Intelligence
• Signal Processing

Background:

• Anomalous patterns in biosignals pose challenges for early disease detection.
• Developing portable systems for real-time analysis requires robust classification methods.

Purpose of the Study:

• To compare the efficacy of complex-valued versus real-valued convolutional neural networks for biosignal anomaly classification.
• To evaluate the potential of complex-valued networks for early disease detection of melanoma and heart murmurs.

Main Methods:

• A comparative study using two similar convolutional neural network architectures, one operating in the complex domain and the other in the real domain.
• Experiments conducted on three clinical datasets: ISIC2017, PH2, and Pascal.
• Statistical analysis using mean comparison hypothesis tests to validate findings.

Main Results:

• Complex-valued networks demonstrated superior performance across Precision, Recall, F1 Score, Accuracy, and Specificity metrics.
• In melanoma detection, the best complex-valued classifier achieved a Euclidean distance of 0.26127 in ROC space, compared to 0.36022 for the best real-valued classifier.
• Complex-valued networks showed a 27.46% superiority in discriminating features for anomaly detection.

Conclusions:

• Complex-valued neural networks offer enhanced feature extraction capabilities for biosignal analysis.
• The findings suggest complex-valued networks are a powerful approach for developing portable early disease detection systems.
• This study highlights the potential of complex-valued approaches for improving diagnostic accuracy in medical applications.