Abstract:
The workflow of the development of three-dimensional (3D) computational models for cochlear implants was investigated, with specific focus on applied automation techniques. No fully automated process was found in the literature that could perform the entire 3D model development from the first to the final stage, with the greatest lack of automation identified in the data interpretation and processing phase. It is proposed that the workflow of 3D model development can be automated to such an extent that an automated cochlear model generator can be developed. The aim of such a method is to reduce time spent on model development, and decrease the number of complicated manual procedures often involved in 3D model development whilst maintaining model accuracy. A knowledge-based landmark detection algorithm was used to develop a semi-automated cochlear model creation tool by using standard CT scan data. Six 3D electric volume conduction models were produced by applying the automated method. Electric potential distributions, as a result of intracochlear stimulation, were calculated and then used to predict neural activating function patterns. Predictions from models resulting from automated generation were compared to predictions from models that were created by a purely manual generation method. Automation of the model development workflow was achieved, although an initial manual calibration procedure was required for each model. For the
development of 3D models, the use of multiple geometrical landmark points (GLP) greatly affected cochlear model morphology, potential distributions, and neural excitement as opposed to the use of a singular GLP.
This work suggests that the semi-automated method developed and presented in this study is able to detect cochlear landmarks with an 84.28% similarity to the manual method. Higher intracochlear potentials were predicted with the automated method because of the reduced volume of the automatically generated models compared to that of manually created models. The higher potentials indicated a greater probability of neural excitation when compared to the manually created models, under similar stimulation conditions.
