Deep Learning for Synthetic CT from Bone MRI in the Head and Neck

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Bone MR imaging offers radiation-free visualization of cortical bone.
• Automated conversion of bone MR to synthetic CT is crucial for clinical applications.
• Deep learning is well-suited for complex anatomical mapping in head and neck imaging.

Purpose of the Study:

• To develop and optimize a deep learning model for converting bone MR imaging to synthetic CT.
• To evaluate the performance and generalizability of different deep learning architectures.
• To assess the impact of various training strategies on model accuracy.

Main Methods:

• Retrospective study of 39 patients with head and neck bone MR and CT scans.
• Spatial coregistration of MR and CT data, followed by slice generation for training.
• Training and validation of three encoder-decoder models (Light_U-Net, VGG-16 U-Net) with different loss functions and cross-validation.

Main Results:

• Light_U-Net architecture demonstrated superior quantitative performance compared to VGG-16 models.
• Mean absolute error loss optimized bone precision, while mean squared error improved bone recall.
• Model generalizability improved with augmented training data from diverse sources (hospitals, vendors, techniques).

Conclusions:

• A robust deep learning model was optimized for bone MR to synthetic CT conversion.
• The model exhibits good performance and generalizability across varied imaging conditions.
• This approach facilitates downstream image processing and holds potential for clinical integration.