Edge vs Cloud Models in App-Based AI for Orthodontic Assessment: An Analysis of Architecture, Performance Metrics, and Clinical Effectiveness in AI-driven Malocclusion Detection on Edge and Cloud Platforms

Abstract

This thesis investigates the trade-offs between edge and cloud deployment of CNN models for AI-driven malocclusion detection in a mobile application. The study explores the CNNs VGG-19, YOLOv8 and ResNet-50 and their latency, accuracy and efficiency in Cloud and Edge environments. A mobile prototype was developed to capture dental images, preprocess them using ARKit-based alignment and validation mechanisms, and classify the severity of malocclusion according to the Skåneindex scale. Each model was evaluated using performance metrics such as F1-score, specificity, and inference latency, while also considering computational resource usage. The results indicate that edge deployment reduces latency, improving user responsiveness, while cloud deployment offers marginally higher classification accuracy, particularly with VGG-19. YOLOv8 demonstrated strong overall performance and robustness across environments. Additionally, expert stakeholder validation confirmed the clinical potential of the application in streamlining orthodontic screening and referral processes. These findings highlight key considerations in selecting an appropriate deployment strategy for mobile AI applications in healthcare.