Automatic Identification of Hard and Soft Bone Tissues by Analyzing Drilling Sounds

Abstract

Our purpose is to assess if bone drilling sounds could be used to automatically distinguish between hard- and soft-bone tissues. Such a capability could be of value in a variety of surgical procedures involving drilling through bone. We acquired sound signals from six bovine tibial bones that were being drilled with a surgical drill. We investigated various classifiers including logistic regression, support vector machine (SVM), random forest (RF), and hidden Markov model. We explored different time and frequency features, and considered two training/testing scenarios: leave-one bone-out (LOBO) and bone-specific (BSP). Moreover, we conducted a survey of practicing surgeons to provide a baseline measure for assessing whether our proposed algorithm could improve the ability of surgeons to identify hard- and soft-bone tissues based on drilling sounds. The average accuracies for LOBO and BSP were 79.1% and 84.3%, respectively. In both scenarios, the feature that resulted in the highest accuracy of classification was wavelet packet transform coefficients. Moreover, RF and SVM produced the highest accuracies for LOBO and BSP, respectively. In the survey, the surgeons were unsure on 29.2% of questions, and had an average accuracy of 51.4% on their judged answers. In conclusion, the drilling in hard and soft bones could be automatically identified with good accuracy based on the drilling sounds. Several of the algorithms tested were generalizable across specimens and their accuracy significantly exceeded surgeons' performance. The significance is that these algorithms can potentially be used to reduce challenges associated with bone-drilling procedures.