Benchmarking deep learning-designed inlay restorations across operator experience: An in vitro comparison of time efficiency, contact intensity, and contour quality

Objectives

This in vitro study aimed to evaluate the performance of a deep learning (DL)-based workflow for designing inlays, in terms of time efficiency, contact intensity, and contour quality, by comparing it with human-based workflows. The impact of operator experience was also assessed to examine whether the DL-based workflow could reduce experience-related variability.

Methods

A total of 25 digital scans of maxillary and mandibular arches, including posterior abutments prepared for mesial-occlusal (MO) or distal-occlusal (DO) cavities, were used to design inlays using five different workflows. Two conventional human-based workflows were used for the controls: one by a master-level user (CM) and one by a beginner-level user (CB). A commercial DL-based workflow was tested in three distinct forms: as-generated (TD), optimized by the master-level user (TDM), and optimized by the beginner-level user (TDB). The workflows were compared using evaluation metrics including time efficiency, occlusal and proximal contact intensity, and overall contour quality. The Kruskal-Wallis test and the chi-square test were used to detect statistical differences among the groups. Post hoc analyses were performed using Dunn’s test and analysis of standardized residuals (α = 0.05).

Results

The design time varied significantly among the groups (P < 0.001), with TD showing a significantly shorter time than CM (P < 0.001), except between CM and TDB. Significant differences in occlusal and proximal contact intensities were observed among groups (P < 0.001). Regarding operator experience, occlusal contact intensity differed significantly between TDM and TDB (P = 0.003), and proximal contact intensity between CM and CB (P = 0.002). Contour quality also differed across workflows (P < 0.001); however, no significant difference was found between CM and CB (P = 0.470) or between TDM and TDB (P = 0.059).

Conclusions

DL-based workflows may complement operator experience by improving time efficiency and reducing variability between technicians. Significant differences were observed in time efficiency, contact intensity, and contour quality across workflows, though human expertise remains crucial for both quantitative and qualitative outcomes.

Clinical Significance

Compared to conventional human-based design workflows, DL-based design workflows may enhance time efficiency and reduce skill disparities among technicians of varying experience levels. However, operator experience remains essential for achieving optimal clinical outcomes, including favorable occlusal and proximal contacts and optimal contour in posterior inlays designed to restore occlusal and proximal surfaces.