Segmentation of Craniomaxillofacial Bony Structures from MRI with a 3D Deep-Learning Based Cascade Framework

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Craniomaxillofacial Surgery

Background:

• Computed tomography (CT) is standard for craniomaxillofacial (CMF) surgery planning but involves ionizing radiation.
• Magnetic resonance imaging (MRI) is a safer, noninvasive alternative, ideal for soft tissue analysis, but bone segmentation is challenging due to low signal and artifacts.
• Accurate segmentation of CMF bony structures from MRI is difficult due to bone and air appearing similarly (black) and technical imaging limitations.

Purpose of the Study:

• To develop and validate a 3D deep learning cascade framework for accurate segmentation of CMF bony structures using MRI.
• To overcome the limitations of MRI for bone imaging, providing a radiation-free alternative for surgical planning.

Main Methods:

• A 3D deep learning cascade framework combining a 3D fully convolutional network (FCN) for coarse segmentation and a convolutional neural network (CNN) for fine-grained refinement was developed.
• The CNN utilized concatenated probability maps from the FCN and original MRI data to enhance segmentation context and accuracy.
• The framework was evaluated for its performance and clinical feasibility in segmenting CMF bony structures from MRI.

Main Results:

• The proposed 3D deep learning cascade framework demonstrated good performance in segmenting CMF bony structures from MRI.
• The method successfully addressed challenges like low signal-to-noise ratio and partial volume effects inherent in MRI bone imaging.
• Experimental results confirmed the clinical feasibility of the developed deep learning approach.

Conclusions:

• The 3D deep learning cascade framework offers a promising, radiation-free solution for CMF bony structure segmentation from MRI.
• This approach enhances diagnostic and treatment planning capabilities in CMF surgery by leveraging MRI's safety profile.
• The study highlights the potential of advanced AI techniques to overcome imaging modality limitations in clinical practice.