Development and validation of a novel approach for socket detection in running prostheses using an AI-based video analysis software.

This thesis focuses on the study of a software application that uses artificial intelligence techniques for shape-based key points recognition, specifically designed for the kinematic analysis of two kinds of running prosthetic feet (RPF) used by elite Paralympic athletes: J-shaped and C-shaped RPF. The software application is based on AI-models designed with convolutional neural networks (CNN) and object-detection algorithms. The annotation procedure of several images of RPF allows the training of the models so that they will be capable of detecting the prostheses by analyzing high-resolution videos of running athletes. In this work, a new annotation protocol is introduced to substitute the pre-existing one and to improve the results obtained with the first protocol, following the requirements of a higher number of key points, clearer and non-ambiguous definition of the locations of the key points and greater precision in the detection of the socket. This new protocol tries to better identify the socket by means of a post-processing calibration that allows to determine the axis of the socket, that will be useful, in future research, for further mechanical analysis of the prostheses. The error analysis is first performed on the pre-existing models trained with the annotation protocol designed prior to this work, and on the models trained with the new annotation protocol. Both the analyses are performed with respect to the gold standard represented by the manual annotation performed by the operator. The new protocol produces a maximum error of 1.8 cm on the key points placement performed by the model. The kinematic analysis performed by the software produces an error of 3.6° and 1.9 cm for the angles and the distances computed.