The integration of artificial intelligence (AI) into dental prosthetics has revolutionized the design and fabrication of implant-supported crowns, offering unprecedented precision and efficiency.^[1] Among the most promising tools in this domain are three-dimensional generative adversarial neural networks (3D GANs)^[2], which enable the automated generation of complex 3D prosthetic structures based on extensive training datasets. These AI-driven systems hold the potential to streamline workflows in prosthetic dental medicine, reducing human error and optimizing treatment timelines. However, the validation of these AI-generated designs remains a critical challenge, as the success of a crown extends beyond its digital accuracy to its clinical performance in the human mouth. This raises an essential question: how do we assess the efficacy of AI-designed crowns when traditional validation metrics may not fully capture their real-world applicability?

In AI-driven design, validation typically relies on quantitative metrics such as the intersection over union (IoU) and F1 score, which measure the similarity between the generated 3D files and a predefined “ground truth”—a reference model used as training data and assumed to represent the ideal outcome. “Ground truth”^[3] is usually a 3D model designed manually by a trained specialist with the aid of CAD software. The visual metrics evaluate how closely the neural network replicates the target design, with values approaching 1 indicating near-perfect alignment. Yet, the process of validating 3D-generated files in neural network assessments is inherently tied to this ground truth, which may not account for the variability inherent in clinical scenarios. For instance, a generated crown with a lower IoU or F1 score—indicating deviation from the ground truth—might still yield a successful clinical outcome when applied in vivo. This discrepancy suggests that while these metrics are valuable for assessing algorithmic performance, they may not fully reflect the functional and aesthetic success of prosthetic restoration.

The design of an implant crown involves multiple prosthetic parameters, such as those describing the emergence profile, which includes the contour of the crown^[4], and angles such as the mucosal emergence angle (MEA), deep emergence angle (DA), and total contour angle (CA). These parameters influence not only the crown’s fit and stability but also its biological integration and long-term prognosis.‌^[5] In restoring a missing tooth, there is rarely a single “correct” design; rather, multiple configurations may achieve excellent clinical outcomes depending on patient-specific factors, such as occlusion, soft tissue architecture, and bone morphology. Thus, an AI system such as 3D GANs, trained to approximate a singular ground truth, might undervalue designs that deviate from this reference yet remain clinically viable. This highlights a key insight: the success of AI-designed crowns may depend less on achieving perfect metric scores and more on how prosthetic parameters align with real-world functionality.

This research paper explores the role of prosthetic parameters in validating AI-designed implant crowns, building on prior studies that correlate emergence angles with validation metrics. By examining how these parameters interact with neural network outputs, we aim to bridge the gap between digital precision and clinical efficacy. Our hypothesis posits that even crowns with suboptimal IoU or F1 scores can achieve favorable outcomes in the human mouth if their prosthetic features meet critical clinical thresholds. This investigation seeks to redefine validation in AI-driven prosthetics, emphasizing the synergy between computational design and practical application.

For the purposes of the study, a database of 200 training sets was created, each consisting of five 3D images, 20 of which were used as a control group. The 3D GAN developed and trained by G. Kostadinov and A. Naydenov was used to generate 20 screw-retained implant crowns replacing a missing tooth in position 36. The abutment used for all cases was Neodent GM.^[6] The initial voxelized files of all 20 crowns were processed with the MeshMixer software, followed by the creation of medio-distal and vestibulo-lingual longitudinal cuts using the ExoCad software. For the emergence angle analysis, CorelDRAW was used. Grok, an artificial intelligence language model developed by xAI, was used to aid with the statistical calculations and graphics.

This study comprised four sequential stages: first stage – refining the voxelized data; second stage – obtaining sectional views; third stage – outlining and measuring; fourth stage – analyzing the data.

Overall median scores: MEA: 27.79°, DA: 56.70°, CA: 43.54

Median scores per surface: MEA: vestibular: 34.50°, lingual/palatal: 35.48°, mesial: 16.59°, distal: 15.81°; DA: vestibular: 59.23°, lingual/palatal: 61.61°, mesial: 51.54°, distal: 53.35°; CA – vestibular: 53.20°, lingual/palatal: 51.38°, mesial: 36.08°, distal: 37.53°. Maximum values for MEA, DA, and CA are 68.78°, 93.55°, and 83.01°, and minimum values are -0.01°, 27.51°, and 12.35°.

For visual metrics, the overall median IoU was 0.835, and the median F1 score was 0.91 across the 20 teeth. The median precision score was 0.945, and the median recall score was 0.91. To assess variability, teeth were grouped based on performance relative to the overall median. For IoU, teeth with values ≥0.835 (n=11) had a median of 0.86, while those of <0.835 (n=9) had a median of 0.80. For the F1 score, teeth with values ≥0.91 (n=11) had a median of 0.92, and those of <0.91 (n=9) had a median of 0.89. For the precision score, teeth with values ≥0.945 (n=10) had a median of 0.96, while those of <0.945 (n=10) had a median of 0.91. For recall, teeth with values ≥0.91 (n=10) had a median of 0.94, while those of <0.91 (n=10) had a median of 0.84.

The mean deviation from optimal values (equal to 1 for all visual metrics) is 0.165 for IOU, 0.055 for precision, 0.09 for recall, and 0.09 for the F1 score.

The comparative analysis according to the values we used for interpretation shows that the MEA mean score is “excellent,” the CA mean score is “satisfactory,” the DA mean score is “average,” and the IoU, F1 score, precision, and recall scores are not considered excellent – their mean score is less than 1.0, but it can be analyzed as appropriate.

The integration of artificial intelligence (AI) in prosthetic dental medicine is transforming the design and fabrication of implant-supported crowns, offering enhanced precision and efficiency. A systematic review by Revilla-León^[10] highlights AI’s potential in diagnosis, treatment planning, and prosthesis design. Among the most promising tools in this domain are the three-dimensional generative adversarial neural networks (3D GANs)^[11], which enable the automated generation of complex three-dimensional prosthetic structures based on extensive training datasets. These AI-driven systems hold the potential to streamline workflows in prosthetic dental medicine, reducing human error and optimizing treatment timelines. However, the validation of these AI-generated designs remains a critical challenge, as the success of a crown extends beyond its digital accuracy to its clinical performance in the human mouth. Kong et al.^[12] support the need for robust validation beyond digital accuracy. This raises an essential question: how exactly do we assess the efficacy of AI-designed crowns when traditional validation metrics may not fully capture their real-world applicability?

Validating AI-designed crowns involves assessing 3D-generated files through neural network evaluations, often using metrics such as Intersection over Union (IoU) and F1 score. These metrics measure the similarity between the AI-generated design and a predefined ground truth.^[13] However, research suggests that these metrics may not fully capture clinical efficacy, as evidenced by Rohrer et al.^[14], who reported high F1 scores (97%), while clinical outcomes depend on more than digital accuracy.

The process of validation, as explored by Piedra-Cascón et al.^[15], involves comparing AI outputs against a reference model. However, this approach may overlook variability in clinical scenarios. Revilla-León et al.^[10] indicate that neural networks trained on a single ground truth may undervalue designs that deviate from this reference yet remain clinically viable.

In prosthetic dental medicine, one of the most important predictors for successful treatment is the use of critical prosthetic parameters as guidelines in design. Mattheos et al.^[5] reported that “Emergence Profile (EP), Emergence Angle (EA), and Cervical Margin (CM), as well as the design of the implant–abutment and abutment–prosthesis junctions and their location in relation to the tissues of the implant supracrestal complex, can have a significant impact on the maintenance of stable and healthy peri-implant tissues in the long term.”

Considering the evolution of implant crown design and the predictors for successful treatment, this study was conducted to investigate whether it is useful to assess the outcome of automated design by GAN networks using not only visual parameters but also clinical ones.

The present findings indicate that commonly used visual metrics such as IoU and F1 score are predictive factors for the quality of newly generated three-dimensional objects. However, it remains questionable whether a newly generated crown can be used in a real clinical situation even if IoU and F1 scores show perfect results (=1).^[8,9] This issue is related to the method used for training such GANs and the widely accepted approach to visual assessment. At present, the so-called ground truth is compared with newly generated 3D objects to assess inaccuracies.^[3] Further investigations are needed to support this statement. Real clinical performance must be assessed using specific science-based parameters for prosthetic constructions, such as screw-retained implant crowns, among others. In this study, angles describing the emergence profile were measured as some of the most important clinically relevant parameters for successful prosthetic treatment. It was observed that one of the most important predictors, the mucosal emergence angle (MEA), demonstrated an excellent mean score, suggesting that the newly generated crowns may have excellent clinical applicability with respect to this parameter.

It is suggested that further investigations should include additional clinically and prosthetically relevant parameters, as well as their relationship to the visual metric scores used for validation. A standardized protocol incorporating all clinically relevant parameters and visual metrics should be adopted to improve the assessment of newly generated crowns and enhance their translation into clinical practice.

In conclusion, while AI offers immense potential, validation must evolve to include clinical parameters beyond digital metrics. Visual parameters used to validate the design of implant crowns are not universal but rather serve as supplementary measures for assessing the real clinical value of newly generated crowns. The present findings demonstrate that even in cases where IoU, F1 score, precision, and recall are not perfect (<1), clinically significant parameters such as MEA can still demonstrate excellent mean values (27.79°), which represent a predictive factor for peri-implant tissue health.