A Computer Vision Algorithm to Classify Pneumatization of the Mastoid Process on Temporal Bone Computed Tomography Scans

Abstract

BACKGROUND: Pneumatization of the mastoid process is variable and of significance to the operative surgeon. Surgical approaches to the temporal bone require an understanding of pneumatization and its implications for surgical access. This study aims to determine the feasibility of using deep learning convolutional neural network algorithms to classify pneumatization of the mastoid process. Methods: De-identified petrous temporal bone images were acquired from a tertiary hospital radiology picture archiving and communication system. A binary classification mode in the pretrained convolutional neural network was used to investigate the utility of convolutional neural networks in temporal bone imaging. False positive and negative images were reanalyzed by the investigators and qualitatively assessed to consider reasons for inaccuracy. Results: The overall accuracy of the model was 0.954. At a probability threshold of 65%, the sensitivity of the model was 0.860 (95% CI 0.783-0.934) and the specificity was 0.989 (95% CI 0.960-0.999). The positive predictive value was 0.973 (95% CI 0.904-0.993) and the negative predictive value was 0.935 (95% CI 0.901-0.965). The false positive rate was 0.006. The F1 number was 0.926 demonstrating a high accuracy for the model. Conclusion: The temporal bone is a complex anatomical region of interest to otolaryngologists. Surgical planning requires high-resolution computed tomography scans, the interpretation of which can be augmented with machine learning. This initial study demonstrates the feasibility of utilizing machine learning algorithms to discriminate anatomical variation with a high degree of accuracy. It is hoped this will lead to further investigation regarding more complex anatomical structures in the temporal bone.

Applications of Artificial Intelligence in Temporal Bone Imaging: Advances and Future Challenges

“…TB: temporal bone; Net: network; CNN: convolutional neural networks; DC: Dice coefficient; ASSD: average symmetric surface distance; AH-Net: anisotropic hybrid network; ResNet: residual neural network; DSC: Dice similarity coefficient; PWD: patch-wise densely connected; YOLACT: You Only Look At CoefficienTs; DSD: deep supervised densely; MSSIM: mean structural similarity index; IoU: intersection over union; PSO: particle swarm optimization; BF: Bayes factors; N/S: not specifiedAI in middle ear diseaseChronic Otitis Media With or Without CholesteatomaAI's role in the management of chronic otitis media (COM) has been well-established, including image analysis, automated diagnosis, surgical planning, treatment recommendations, monitoring, and prognostication[43,44]. Multiple software programs, including CNN, VGG-16, and MobileNetV2, have been used for the detection of COM[45][46][47][48][49][50][51][52][53][54][55][56][57][58]. Studies involving AI technologies in middle ear diseases are presented in Table2.…”