Dental abutments are components used in dental implantology. They attach to a dental implant embedded in the patient's skull or jawbone, and a prosthetic restoration is then attached to the abutment. Each component used in implantology, including dental abutments, affects the success of implant therapy. Custom dental abutments, tailored to individual patients, exhibit better mechanical properties and fit compared to standardised, stock abutments. Analysing the geometry of custom dental abutments is a crucial step in achieving a better understanding of how the geometry affects therapeutic success. Unsupervised machine learning can analyse the geometry of custom dental abutments without human bias. Although machine learning has been successfully implemented in various fields, its applications with 3D models remain limited due to irregular data discretisation and insufficient labeled datasets. 3D models are typically described by point clouds or meshes, but machine learning often works with voxel models due to their uniform data discretisation. Custom dental abutments are suitable candidates for analysis using unsupervised machine learning to simplify their design process. This study applies several unsupervised machine learning methods to voxel models of custom dental abutments in order to analyse their geometry. Three unsupervised machine learning methods were used: K-Means for segmentation, a convolutional autoencoder for reconstruction and encoding, and principal component analysis combined with K-Means for the classification of custom dental abutments. K-Means effectively segmented abutments into sections but struggled with wider coronal sections. Principal component analysis combined with K-Means identified important features of custom dental abutments and classified them into six categories. The convolutional autoencoder generated a 16-value encoded representation, enabling classification, though the quality of reconstructions varied. Custom abutments for premolars and molars had better reconstructions than those for incisors and canines. These findings indicate the potential for automating the labelling and simplifying the design of custom dental abutments through unsupervised ML.
