Improved bioimpedance spectroscopy tissue classification through data augmentation from generative adversarial...

Area of Science:

• Biomedical Engineering
• Machine Learning
• Signal Processing

Background:

• Bioimpedance spectroscopy (BIS) is a valuable tissue classification method.
• Accurate BIS requires large datasets, which are difficult to obtain.
• Existing data augmentation methods for BIS lack effectiveness in preserving key features.

Purpose of the Study:

• To propose a novel data augmentation technique for bioimpedance spectral data using generative adversarial networks (GANs).
• To evaluate the performance of GAN-generated data in improving tissue classification accuracy.
• To compare different GAN architectures for their efficacy in augmenting BIS data.

Main Methods:

• Employed three GAN architectures: vanilla GAN, deep convolutional GAN (DCGAN), and Wasserstein GAN (WGAN).
• Generated augmented bioimpedance spectral datasets using the proposed GAN models.
• Trained five classification methods on both original and augmented datasets, comparing results against a baseline.

Main Results:

• DCGAN generated data statistically similar to the original dataset.
• Augmentation with DCGAN improved classification accuracy by 15% compared to using original data alone.
• WGAN architecture demonstrated significant accuracy improvements, up to 24%.

Conclusions:

• GAN-based data augmentation effectively addresses the challenge of limited datasets in bioimpedance spectroscopy.
• The proposed method generates high-fidelity spectral data, enhancing the generalizability of classification models.
• DCGAN and WGAN show significant potential for improving tissue classification accuracy in clinical applications.