Transfer Learning From Micro-CT to Periapical Radiographs for Three-Dimensional Root Canal Morphological...

Area of Science:

• Dental Imaging and Diagnostics
• Machine Learning in Medicine
• Radiology and Anatomy

Background:

• Accurate identification of root canal morphology is crucial for endodontic treatment success.
• Current methods for analyzing root canal anatomy rely on 2D radiographs, which can limit the visualization of complex 3D structures.
• Multimodal transfer learning offers a potential solution for enhancing diagnostic capabilities by integrating data from different imaging modalities.

Purpose of the Study:

• To investigate the transfer of implicit anatomical features from micro-computed tomography (micro-CT) to periapical radiographs.
• To evaluate the effectiveness of multimodal transfer learning for 3D root canal morphology identification.
• To assess the impact of task complexity on the performance of transfer learning models.

Main Methods:

• Fused-rooted mandibular second molars (MSMs) were scanned using micro-CT to create virtual radiographs.
• Clinically simulated periapical radiographs (CSPRs) were generated from ex vivo mandibles.
• Four convolutional neural network (CNN) architectures were trained using different pre-training strategies, including models pre-trained on ImageNet and virtual radiographs.
• Grad-CAM visualization was employed to interpret model attention, and results were compared with endodontic residents' performance.

Main Results:

• CNNs pre-trained on virtual radiographs achieved higher accuracy (69.68%) in a three-class classification task compared to ImageNet-pretrained models (64.36%) and endodontic residents (61.17%).
• Grad-CAM analysis showed that virtual radiograph-pretrained models focused on relevant root structures, unlike ImageNet-pretrained models.
• In a simplified two-class task, performance differences between methods were not statistically significant, suggesting transfer learning benefits are greater for complex tasks.

Conclusions:

• Implicit 3D features from micro-CT virtual radiographs can be effectively transferred to CSPRs via transfer learning.
• This approach improves CNN interpretability and diagnostic accuracy for root canal morphology identification.
• The efficacy of multimodal transfer learning is more pronounced in complex, multi-class classification tasks, providing a foundation for its application in clinical dental imaging.