.png){kind=icon}
.png){kind=icon}
Return
Deep Learning-Based Dental Image Analysis Using Grad Cam Convolutional Network
.png){kind=icon}
PRE
AI
DOI：10.1922/ejprd.v34i2.1321.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Panoramic radiographs have been regularly applied in the field of prosthodontic and restorative dentistry to aid in treatment planning, tooth morphology, edentulous space evaluation, and preliminary screening of the implant site. Nevertheless, panoramic images may not be easily interpreted because of anatomy overlap, distortion, and varying quality of images. This paper provides a deep learning architecture that can be explained and used to aid the radiographic evaluation of prosthodontics by means of automated localization of teeth and interpretable visualisation. It applied a multi-stage pipeline, which involved dataset validation, COCO-based annotation auditing, instance segmentation with a Mask R-CNN backbone based on ResNet-FPN, and incorporation of explainable artificial intelligence (XAI) methods. Grad-CAM, occlusion sensitivity mapping, and mask confidence visualisation were automatically used to segment and analyse tooth regions to get transparent decision-support outputs. Qualitative data showed that there was anatomical localization and activation of teeth in patterns that were consistent with morphologically relevant structures. Despite the fact that quantitative metrics of segmentation were affected by rigid confidence levels, explainability analysis showed that model attention was mainly focused on tooth anatomy instead of background artefact. The suggested framework offers a reproducible and interpretable base of AI-assisted panoramic radiograph interpretation, which has possible applications in the field of prosthodontic planning, restorative assessment, and implant-oriented screening procedures.
Keywords:
Deep learning
Explainable artificial intelligence (XAI)
Grad-CAM
Instance segmentation
Panoramic dental radiographs
Journal
E
IF:
0
Papers:
49
Citations:
0
