A Dynamic Method of Bowel Lesions Characterisation Using Self-Propelled Robotic Capsule and Machine Learning

A non-invasive method of characterising bowel lesions using machine learning (ML) and a robotic capsule has been proposed and investigated. As the capsule travels and encounters lesions in the lumen, its exhibited dynamics are envisaged to vary significantly in response to any change in the biomechanical properties of encountered lesions. Resulting capsule dynamics in the form of displacement signals have been processed for features that may be indicative of variation in biomechanical property such as stiffness (E). In the first stage ML, multi-layer perceptron (MLP) and support vector regression (SVR) were trained to predict <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">E</tex> -values from the processed features. In the second stage, unsupervised K-means clustering is used to group the lesions into clusters of high intra-cluster similarities but low inter-cluster similarities using the MLP and SVR predicted E-values. The method achieved greater than 97 % clustering accuracies for both simulation and experimental data.