Investigating the influence of dimensionality reduction on force estimation in robotic-assisted surgery using recurrent and convolutional networks

Abstract

Large data computing is a research problem and a major challenge in order to successfully mine, process, and evaluate massive datasets, as they represent a useful source of knowledge across multiple and intersecting domains. This paper explores the impact of Dimensionality Reduction (DR) on estimating the force in robotic-assisted surgery using recurrent and recurrent convolutional neural networks. This work collects an extensive dataset from three ex vivo porcine samples and one ex vivo artificial skin, as well as from various sensors, surgical tools, and manipulators, to research the impact of dimensionality reduction. Three neural networks were considered to analyze and validate the results of this work: Recurrent Neural Networks (RNN-LSTM), Recurrent Convolutional Neural Networks (RCNN), and Modified Inception ResNet V2. The statistical analysis of the estimated force quality shows a 17.33% improvement in Ince. ResNet V2 networks, a 10.08% improvement in CNN+LSTM networks, and a 3.88% improvement in RNN-LSTM networks with and without dimensionality reduction in the average of all the force components in all three datasets. The analysis also shows a reduction in training time of 8.21% in LSTM-RNN, 14.91% in CNN-LSTM, and 21.01% in Ince. ResNet V2 with and without DR. Additionally, networks with DR outperform those without DR in terms of execution time per step and force prediction time, resulting in notable reductions in each aspect. Sensitivity analysis reveals that Torque, Position, Deformation, Stiffness, Tool diameter, Rotation, and Orientation features have significant impacts on the predicted force after DR. It was also observed that the predicted force quality was superior when performing feature selection and dimensionality reduction on collected features from tool, manipulator, tissue, and vision data when processed simultaneously in all three architectures. The findings have significant implications for various fields and applications beyond the specific domain of surgical robotics.