Exploring Deep Clustering Methods in Vibro-Acoustic Sensing for Enhancing Biological Tissue Characterization

Nonlinear dimensionality reduction techniques, often referred to as manifold learning, are increasingly valuable for data visualization and unsupervised clustering. In the context of surgery and medicine, these methods facilitate the analysis of complex datasets, enabling pattern recognition in surgical data. This study explores the characterization of six tissue types through manifold learning and unsupervised clustering, utilizing vibro-acoustic (VA) signals collected from manual palpation experiments. A wireless sensor mounted at the tip of a surgical instrument was used to acquire 1,680 VA signals, which were processed using Fourier transform and cepstral analysis for feature extraction. We assessed the performance of two dimensionality reduction techniques: uniform manifold approximation and projection (UMAP) and variational autoencoder (VAE). Results indicate that cepstral features combined with UMAP yield superior clustering performance compared to VAE, achieving higher classification accuracy ( $92~\%$ vs. $87~\%$ ) and better-defined clusters with greater compactness. The observed differences in performance are linked to the intrinsic properties of the tissues, particularly surface characteristics such as friction and moisture, which affect signal consistency. Additionally, we compared our approach with previous works, including a study utilizing the same dataset, where our methodology demonstrated improved accuracy. Future research will focus on refining the VAE model, increasing the diversity of tissue samples, and validating the proposed approach in real surgical settings to enhance its applicability in minimally invasive surgery.