A Deep Learning Algorithm to Identify Anatomical Landmarks on Computed Tomography of the Temporal Bone

Area of Science:

• Medical Imaging
• Artificial Intelligence in Surgery
• Neurosurgery

Background:

• Petrous temporal bone cone-beam computed tomography (CBCT) is crucial for diagnosing and identifying surgical landmarks in temporal bone and mastoid procedures.
• Accurate identification of anatomical structures on CBCT scans is essential for successful surgical outcomes.
• Deep learning, specifically convolutional neural networks (CNNs), offers potential for automating and augmenting image analysis in medical diagnostics.

Purpose of the Study:

• To evaluate the accuracy of a deep learning CNN algorithm in identifying key structures on petrous temporal bone CBCT scans.
• To compare the performance of the CNN when trained by clinicians with varying levels of experience (senior vs. junior).

Main Methods:

• 129 petrous temporal bone CBCT scans were analyzed.
• Key intraoperative landmarks were manually labeled on 68 scans by an otolaryngology registrar and a board-certified otolaryngologist.
• A CNN (Microsoft Custom Vision) was trained on labeled data and used for automated structure identification on the remaining 61 scans, with results verified by an otolaryngologist.

Main Results:

• The CNN achieved high accuracy in automated structure identification on both axial (0.958) and coronal (0.924) CBCT slices (P < .001).
• CNN accuracy correlated positively with the seniority of the clinician who provided the training data.
• More complex structures like the cochlea, vestibule, and carotid canal showed greater accuracy improvement with senior clinician training.

Conclusions:

• CNNs demonstrate high accuracy for automated structure identification in petrous temporal bone CBCT scans.
• The expertise of the training clinician significantly impacts CNN performance, with senior clinicians yielding superior results.
• Training CNNs with data from the most experienced clinicians is recommended to maximize identification accuracy for surgical applications.