Advanced Dental Implant System Classification with Pre-trained CNN Models and Multi-branch Spectral Channel Attention Networks

Abstract

Dental implants (DIs) are prone to failure due to uncommon mechanical complications and fractures. Precise identification of implant fixture systems from periapical radiographs is imperative for accurate diagnosis and treatment, particularly in the absence of comprehensive medical records. Existing methods predominantly leverage spatial features derived from implant images using convolutional neural networks (CNNs). However, texture images exhibit distinctive patterns detectable as strong energy at specific frequencies in the frequency domain, a characteristic that motivates this study to employ frequency-domain analysis through a novel multi-branch spectral channel attention network (MBSCAN). High-frequency data obtained via a two-dimensional (2D) discrete cosine transform (DCT) are exploited to retain phase information and broaden the application of frequency-domain attention mechanisms. Fine-tuning of the multi-branch spectral channel attention (MBSCA) parameters is achieved through the modified aquila optimizer (MAO) algorithm, optimizing classification accuracy. Furthermore, pre-trained CNN architectures such as Visual Geometry Group (VGG) 16 and VGG19 are harnessed to extract features for classifying intact and fractured DIs from panoramic and periapical radiographs. The dataset comprises 251 radiographic images of intact DIs and 194 images of fractured DIs, meticulously selected from a pool of 21,398 DIs examined across two dental facilities. The proposed model has exhibited robust accuracy in detecting and classifying fractured DIs, particularly when relying exclusively on periapical images. The MBSCA-MAO scheme has demonstrated exceptional performance, achieving a classification accuracy of 95.7% with precision, recall, and F1-score values of 95.2%, 94.3%, and 95.6%, respectively. Comparative analysis indicates that the proposed model significantly surpasses existing methods, showcasing its superior efficacy in DI classification.