Machine learning model for automated calculation of intracochlear positional index in cochlear implantation

Area of Science:

• Medical Imaging
• Machine Learning
• Computational Anatomy

Background:

• The intracochlear positional index (ICPI) is crucial for cochlear implant performance.
• Manual ICPI calculation from CT scans is time-consuming and error-prone.
• Automating ICPI calculation can enhance surgical precision.

Purpose of the Study:

• To develop and validate a convolutional neural network (CNN) for automated ICPI calculation.
• To compare the performance of a custom CNN against a pre-trained ResNet-50 model.
• To investigate the impact of training epochs on ICPI calculation accuracy.

Main Methods:

• Training and validation of custom CNN and ResNet-50 models on 34 temporal bone CT images.
• Testing models on eight independent CT images.
• Establishing ground truth ICPI values through manual measurement of electrode-to-modiolus and electrode-to-lateral wall distances.

Main Results:

• Both custom CNN and ResNet-50 models successfully automated ICPI calculation.
• Increasing training epochs from 10 to 100 significantly improved accuracy for both models.
• The pre-trained ResNet-50 model demonstrated statistically significant superior performance with lower error rates (MAE, RMSE) compared to the custom CNN.

Conclusions:

• Machine learning models, particularly ResNet-50, can effectively automate ICPI calculation.
• Optimizing training epochs to 100 iterations enhances model accuracy.
• Further validation on larger datasets is recommended to improve real-world applicability and correlate with audiological outcomes.