Scholarlink

<p>Orbital fractures are frequently encountered in maxillofacial trauma and can be associated with severe globe injuries (SGI). Accurate and timely diagnosis is critical to prevent long-term complications, yet subtle fractures and soft tissue injuries may be overlooked. Recent advances in artificial intelligence offer promising tools to support clinical decision-making in acute settings. We collected CT images from 250 eyes with orbital wall fractures treated at a single center between May 2020 and May 2023. From CT scans of included subjects, a dataset of 18,000 annotated orbital CT scan slices was used, with segmentation preprocessing based on a U-Net framework. Two supervised convolutional neural network (CNN) classifiers were trained on axial and coronal CT images of the orbit. The first model (FxDetectCNN) was trained to detect orbital fractures, while the second model (SGIDetectCNN) was trained to classify injury severity as non-severe or severe. The models’ performance was compared with annotations by two experienced ophthalmologists. The fracture detection model achieved an overall accuracy of 81%, with 86% sensitivity and 76% specificity. The SGI estimation model achieved 64% accuracy, with a sensitivity of 80% and a specificity of 48%, compared with expert annotation, which reached a maximum accuracy of 59.5% for SGI estimation. Our study demonstrates the feasibility of CNN models in detecting fractures and classifying the severity of ocular injuries from CT scans. These tools can assist emergency physicians in triaging patients and may serve as valuable decision-support systems in emergency settings.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Automated orbital wall fracture detection and severity classification of ocular injuries using deep learning

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年5月22日

查看全文
  • Farhad Salari,
  • Zahra Meskar,
  • Farhad Fallah,
  • Maryam Fazeli,
  • Reza Samiee,
  • Hossein Arabalibeik,
  • Seyed Mohsen Rafizadeh

摘要

Orbital fractures are frequently encountered in maxillofacial trauma and can be associated with severe globe injuries (SGI). Accurate and timely diagnosis is critical to prevent long-term complications, yet subtle fractures and soft tissue injuries may be overlooked. Recent advances in artificial intelligence offer promising tools to support clinical decision-making in acute settings. We collected CT images from 250 eyes with orbital wall fractures treated at a single center between May 2020 and May 2023. From CT scans of included subjects, a dataset of 18,000 annotated orbital CT scan slices was used, with segmentation preprocessing based on a U-Net framework. Two supervised convolutional neural network (CNN) classifiers were trained on axial and coronal CT images of the orbit. The first model (FxDetectCNN) was trained to detect orbital fractures, while the second model (SGIDetectCNN) was trained to classify injury severity as non-severe or severe. The models’ performance was compared with annotations by two experienced ophthalmologists. The fracture detection model achieved an overall accuracy of 81%, with 86% sensitivity and 76% specificity. The SGI estimation model achieved 64% accuracy, with a sensitivity of 80% and a specificity of 48%, compared with expert annotation, which reached a maximum accuracy of 59.5% for SGI estimation. Our study demonstrates the feasibility of CNN models in detecting fractures and classifying the severity of ocular injuries from CT scans. These tools can assist emergency physicians in triaging patients and may serve as valuable decision-support systems in emergency settings.