A U-Net Approach to Apical Lesion Segmentation onPanoramic Radiographs

  1. Introduction:
    • Discusses the use of artificial intelligence (AI) algorithms in medical imaging, specifically in dental radiology.
    • Highlights the advantages of panoramic radiography over other imaging techniques in terms of radiation dose and patient comfort.
  2. Materials and Methods:
    • Radiographic Data Preparation:
      • Describes the source of the 470 panoramic radiographs used in the study.
      • Mentions the exclusion criteria for images with artifacts.
    • Image Annotation:
      • Details the process of ground truth image annotation by experienced dental radiologists.
    • Deep CNN Architecture:
      • Explains the use of a U-Net model implemented with PyTorch for semantic segmentation tasks.
      • Provides an overview of the U-Net architecture with encoding and decoding sections, convolutional filters, and skip connections.
  3. Model Pipeline:
    • Model Development:
      • Utilizes the PyTorch library for model development in Python.
      • Describes the AI model (CranioCatch) developed for automatic segmentation of apical lesions on panoramic radiographs.
    • Data Split and Augmentation:
      • Divides the 470 panoramic radiographs into training, validation, and test groups.
      • Augments the training data set using horizontal and vertical flips.
    • Cropping (Preprocessing Step):
      • Divides all images into four parts for further processing.
  4. Results:
    • Segmentation Performance:
      • Reports the segmentation results of the AI model on the test data set.
      • Provides sensitivity, precision, and F1-score values for the segmentation of periapical lesions.
  5. Discussion:
    • Clinical Implications:
      • Discusses the potential of AI systems to aid in the assessment of periapical pathology based on panoramic radiographs.
    • Comparison with Other Studies:
      • Mentions the success of deep learning networks, including CNNs, in medical imaging applications.
  6. Conclusion:
    • Summarizes the study's findings regarding the U-Net approach for apical lesion segmentation on panoramic radiographs.
Read Full Article