Segmentation of Mandibular Canal on Dental Cone Beam CT Images with AI Development Support Software for Medical Images

DOI IR Open Access
  • Torii Kohei
    Center for Design-Oriented AI Education and Research, Tokushima University
  • Nishimura Ryota
    Information Engineering, Tokushima University Graduate School of Technology, Industrial and Social Science
  • Honda Eiichi
    Oral and Maxillofacial Radiology, Tokushima University Graduate School of Biomedical Sciences Kobayashi Dental Clinic

Bibliographic Information

Other Title
  • 医用画像AI開発支援ソフトウェアを用いた歯科用CBCT画像における下顎管のセグメンテーション
Published
2024
Resource Type
journal article
DOI
  • 10.11242/dentalradiology.64.11
Publisher
Japanese Society for Oral and Maxillofacial Radiology

Search this article

Description

The primary barrier to the development of artificial intelligence （AI） in medical imaging is data collection. Medical professionals generate training data for the development of medical imaging AI; however, this process is time-consuming and labor-intensive. Consequently, insufficient data collection frequently results in the premature discontinuation of AI model development. To address this challenge, we developed a software tool named ‘Aidia’ to support the research and development of medical imaging AI. Aidia offers functionalities such as a medical image viewer, data annotation, model training for various tasks, model evaluation, and automated annotation using trained models. For data annotation, Aidia allows users to annotate images by generating polygons, rectangles, polylines, lines, and points. Aidia supports Digital Imaging and Communications in Medicine images as well as general image formats, and we optimized it for annotating medical images. Moreover, Aidia utilizes open-source Python and PyQt5 libraries to build a cross-platform graphical user interface. Thus, Aidia provides a platform where medical professionals can develop, evaluate, and create training data for AI models at no cost. In this study, we developed a segmentation model based on U-Net to predict mandibular canals in cross-sectional jawbone images using Aidia. We collected 8,287 images annotated by a radiologist and trained the segmentation model using these data. The trained model achieved a precision of 0.805, recall of 0.752, F1 score of 0.777, and average precision of 0.869 on test data, accurately generating training data for the test images. Aidia is a promising solution for AI development and image annotation in the medical field.

Journal

  • Shika Hoshasen

    Shika Hoshasen 64 (1), 11-19, 2024

    Japanese Society for Oral and Maxillofacial Radiology

Keywords

Details 詳細情報について

Report a problem

Back to top