Progressive multi-task learning for fine-grained dental implant classification and segmentation in CBCT image.

Area of Science:

• Oral Medicine
• Medical Imaging
• Artificial Intelligence

Background:

• Digital technology is transforming oral medicine towards computer-assisted diagnosis and treatment.
• Identifying dental implants without prior records is challenging and time-consuming.
• Accurate dental implant identification is vital for treatment success and reliability.

Purpose of the Study:

• To propose a deep learning method for multi-task fine-grained classification and segmentation of dental implants in CBCT images.
• To address challenges in differentiating similar implant features and handling intra-class variations.
• To achieve automatic and synchronized classification and segmentation of implant systems.

Main Methods:

• Developed MFPT-Net, a deep learning model using progressive training with multiscale feature extraction and enhancement.
• Utilized a dataset of 437 CBCT sequences with 723 dental implants from three centers.
• Employed fine-grained classification and segmentation techniques to differentiate subtle implant features.

Main Results:

• Achieved 92.98% accuracy, 93.15% average precision, 93.31% average recall, and 93.18% F1 score for classification.
• Reached a 98.04% Dice similarity coefficient for segmentation, outperforming state-of-the-art methods.
• External validation with 252 implants confirmed clinical feasibility and superior performance over existing models.

Conclusions:

• MFPT-Net demonstrates high accuracy and efficiency in classifying and segmenting dental implants from CBCT images.
• The method assists dentists in implant identification, especially when patient records are unavailable.
• This computer-assisted approach enhances the reliability and precision of dental implant treatments.