Developing a ground truth for a convolutional neural network-based segmentation of anatomical structures in endoscopic ear surgery videos

Objective: This study aimed to develop a ground truth for a convolutional neural network-based segmentation of the stapes, chorda tympani and facial nerve in endoscopic ear surgery videos, and to evaluate the accuracy of artificial intelligence (AI) predictions on test videos.

Methods: Forty prospectively-gathered endoscopic ear tympanotomy videos were segmented to establish a ground truth. This ground truth was used to prime a convolutional neural network (CNN) to predict the stapes, chorda tympani and facial nerve. The CNN was then tested on an unlabeled series of videos, and the accuracy of predictions compared to ground truth labeling by calculating Dice scores, sensitivity and specificity.

Results: An overall Dice score of 77.94 % across all three elements was obtained, with a sensitivity of 78.42 % and specificity of 99.79 %.

Conclusion: Convolutional neural network analysis is effective at identifying key anatomical structures in endoscopic ear surgery videos. Further validation of the CNN on additional video datasets is necessary to optimize model performance.