Classification of Dehiscence Defects in Titanium and Zirconium Dental Implants

Abstract

In oral health, the accurate diagnosis of conditions like periapical lesions and dehiscences, especially those associated with titanium and zirconia implants, presents significant challenges due to the complex nature of such dental pathologies, which often manifest with subtle and overlapping symptoms, making them difficult to distinguish in traditional imaging methods. Moreover, the intricate interaction between these conditions and the surrounding oral structures, compounded by the varied responses to different implant materials, further complicates the diagnostic process. This paper introduces innovative multilabel classification methods aimed at enhancing diagnostic precision. We employ an adapted EfficientNet-B0 model with a new loss function, achieving 97% of accuracy. We automatically select and segment the best DICOM file slice from the cone-beam computed tomographies. This autonomous approach contributes to creating a new dataset that will aid in the diagnosis made by healthcare professionals. This new dataset attained an average Structural Similarity Index Measure (SSIM) of 0.6 compared to images selected by expert radiologists. The paper also explores the model’s explainability and addresses the handling of files originating from Cone Beam Computed Tomography (CBCT).