PWD-3DNet: A Deep Learning-Based Fully-Automated Segmentation of Multiple Structures on Temporal Bone CT Scans

Area of Science:

• Medical Imaging
• Neurosurgery
• Artificial Intelligence

Background:

• Temporal bone surgery is complex due to intricate 3D anatomy.
• Accurate segmentation of intra-temporal structures is crucial for surgical training.
• Current segmentation methods are time-consuming and limited by image quality.

Purpose of the Study:

• To develop a fully automated deep learning pipeline for multi-class segmentation of temporal bone structures.
• To improve the accuracy and efficiency of temporal bone segmentation compared to existing methods.

Main Methods:

• A novel patch-wise densely connected (PWD) three-dimensional (3D) convolutional network, PWD-3DNet, was developed.
• The algorithm performs multi-class segmentation of key temporal bone structures from CT volumes.
• Augmentation layers were incorporated to enhance robustness against varying image acquisition protocols.

Main Results:

• PWD-3DNet achieved an average Dice similarity score of 86% and a Hausdorff distance of 0.755 mm for all segmented structures.
• The algorithm demonstrated superior accuracy and speed compared to manual and semi-automated techniques.
• The model showed robustness when tested on low-resolution CT scans from a different center.

Conclusions:

• The proposed PWD-3DNet offers a highly accurate and efficient solution for automated temporal bone segmentation.
• This deep learning approach has the potential to significantly advance surgical training and pre-operative planning.
• The algorithm's robustness suggests broad applicability across different CT scanner parameters and datasets.