DeepLabv3 + method for detecting and segmenting apical lesions on panoramic radiography.

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Computer Vision

Background:

• Apical periodontitis diagnosis relies on radiographic interpretation.
• Accurate segmentation of apical lesions is crucial for effective treatment planning.
• Deep learning models offer potential for automated analysis of dental radiographs.

Purpose of the Study:

• To evaluate and compare the performance of the DeepLabv3+ model against the U-Net model for detecting and segmenting apical lesions in panoramic radiography.
• To assess the efficacy of state-of-the-art deep semantic segmentation models in dental imaging.

Main Methods:

• A dataset of 260 panoramic images with apical lesions was curated and divided into training and testing sets.
• Manual annotation of apical lesions was performed by experienced dental radiologists.
• The DeepLabv3+ model was implemented using Python and TensorFlow and compared with the U-Net model.

Main Results:

• DeepLabv3+ achieved significantly higher AUC (29.96%) and recall (61.06%) compared to U-Net.
• U-Net demonstrated superior precision (69.17%) and F1-score (25.55%) over DeepLabv3+.
• The Intersection over Union (IoU) results between the two models were not statistically significant.

Conclusions:

• DeepLabv3+ exhibits a substantial advantage over U-Net in detecting apical lesions based on AUC and recall metrics.
• The DeepLabv3+ model holds potential for enhancing clinical diagnosis, treatment planning, and efficiency in dental practice.
• Further research and improvement of the DeepLabv3+ model are encouraged for apical lesion detection.