Modular Neural Networks for Osteoporosis Detection in Mandibular Cone-Beam Computed Tomography Scans

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oral and Maxillofacial Radiology

Background:

• Osteoporosis diagnosis often relies on systemic assessments.
• Cone-beam computed tomography (CBCT) provides detailed mandibular bone structure data.
• Deep convolutional neural networks (DCNNs) offer advanced image analysis capabilities.

Purpose of the Study:

• To evaluate the efficacy of DCNNs for osteoporosis diagnosis using mandibular CBCT scans.
• To develop and assess a segmented, three-phase DCNN approach for osteoporosis detection.
• To explore the potential of AI in enhancing osteoporosis identification from dental imaging.

Main Methods:

• Utilized 188 patient mandibular CBCT images with DCNN models based on the ResNet-101 framework.
• Implemented a segmented three-phase method: bone slice identification, cross-sectional view coordinate pinpointing, and bone thickness computation.
• Employed transfer learning with a modular approach for DCNN training.

Main Results:

• Stage 1 (bone slice identification) achieved 98.85% training accuracy.
• Stage 2 (coordinate pinpointing) minimized L1 loss to 1.02 pixels.
• Stage 3 (bone thickness computation) reported a mean squared error of 0.8377, indicating high accuracy in detecting osteoporotic variances.

Conclusions:

• DCNNs demonstrate significant potential for accurate osteoporosis detection via mandibular CBCT.
• The compartmentalized DCNN method enhances model transparency and training robustness.
• This AI-driven approach is effective even with limited CBCT datasets, promising improved medical care.