Audio Pre-Processing and Neural Network Models for Identification of Orthopedic Reamers in Use

Abstract

In order for a successful outcome of Total Hip Arthroplasty (THA) to occur, implant stability is a key concern. A means to achieve this is ensuring that the implant has an optimum seating within the femur cavity. This is achieved during surgery by progressive reaming of the cavity interior. Both under and over reaming have undesirable effects towards implant longevity and post-operative prognosis, and so through education/experience, orthopedic surgeons have learned to anticipate when optimal reaming occurs. The work presented here is part of a larger research project which seeks to use bone resonance as an indicator of good implant seating. Here we present results of initial work using several neural network models on classification of orthopedic reamer type on the basis of sound signature. These results are discussed in the context of an interesting feature found when comparing differing audio- preprocessing methods. Despite identical audio raw data being used for both representations, the models that used the Mel Spectrograms categorically outperformed those which used the STFT Spectrogram.