Improving Bioimpedance-based Tissue Identification with Frequency Response Similarity Metrics.

Abstract

Tissue identification is essential for surgeons to properly perform procedures and make informed decisions to minimize potential harm to patients. Minimally invasive surgery (MIS) offers enhanced patient safety and outcomes at the cost of lost information due to restricted vision and loss of touch, among other factors. This makes it more difficult to quickly and consistently identify tissues correctly. Bioimpedance spectroscopy (BIS) offers the potential to identify tissues using rapid measurements that leverage differences in electrical properties between tissues. However, using BIS to differentiate large sets of tissues in a singular anatomical area, such as the gastrointestinal (GI) tract, has remained a significant challenge because of the overlap of similar tissues' responses and variability between measurements. This work proposes the application of frequency response function (FRF) similarity metrics as a signal processing technique to extract new features from BIS measurements on porcine tissues. These features are then used as inputs to machine learning (ML) models that are trained on an ex vivo dataset for identification of eight different in vivo porcine abdominal tissues. The ML models using similarity metric inputs performed on par or better than models using raw measurement inputs, except for the support vector machine (SVM) models. A neural network (NN) model using a similarity metric input performed best by achieving a mean accuracy of 70.3% and F-measure of 0.716. More importantly, the similarity metrics enhanced the ability of the models to identify all tissues rather than considering tissues from similar anatomical areas as the same. Ultimately, the FRF similarity metrics are a novel approach for extracting features from BIS measurements that improved identification performance when considering both accuracy and capability of differentiating all tissues in the dataset.