# A Deep Metric Learning and Multimodal Gated Fusion Framework for AI‐Driven Risk Assessment of Lingual Plate Perforation and Mandibular Canal Injury in Posterior Mandible Implants

## TL;DR

This paper introduces an AI model that automatically assesses risks for dental implant procedures, aiming to reduce human variability and manual measurements.

## Contribution

A novel deep metric learning and multimodal gated fusion framework for automated implant risk assessment.

## Key findings

- The model achieved over 0.87 accuracy in detecting implant site risks.
- Gated fusion improved F1-scores by 3%-5% and R² by 5%-8%.
- Grad-CAM visualizations confirmed the model's ability to localize relevant features.

## Abstract

To provide an automated risk assessment for lingual plate perforation (LPP) and mandibular canal injury (MCI) in dental implants. Also, to reduce interoperator variability in risk classification and eliminate manual measurements during implant planning.

A dataset of 896 CBCT‐implant records was used for training our model (DentaRisk‐Net). CBCT images were processed using ResNet‐18 encoders for each implant site view. Patient‐specific metadata was processed through a fully connected network. These modalities were fused using a gated fusion mechanism (GFM). Deep metric learning (DML) was employed to enhance distinguish of risk classes. The model was trained on two risk tasks: bone plate (BP) and mandibular canal (MC) tasks. Classification and regression metrics (accuracy, precision, recall, F1‐score, R
2, and ablation studies) were used to evaluate performance. True assessment rate (TAR) was used to assess the model’s agreement with human annotators. Grad‐CAM visualizations were conducted for qualitative evaluation.

The model achieved an accuracy exceeding 0.87 in detecting risk in implant sites, with precision and recall above 0.84. The GFM increased F1‐scores by 3%–5% and R
2 by 5%–8% for both risk tasks. The model achieved high performance for the Safe and Risk classes, whereas the Caution class was comparatively lower in BP and MC. TAR ≤1 values demonstrated acceptable alignment with human (BP: r = 0.63; MC: r = 0.72). Grad‐CAM visualizations confirmed the model’s ability for feature localization.

By integrating imaging and metadata through DML and GFM, the model attempted to reduce interoperator variability in decision‐making and enhance risk‐classified assessment in the treatment planning process.