Scholarlink

<p>While free flap reconstruction is essential for repairing post-surgical defects or oncologic defects, current monitoring methods lack standardization and rely on subjective assessments. This study aims to introduce an artificial intelligence model that quantitatively analyzes and classifies the status of intraoral free flaps to enhance clinical monitoring capabilities. In this study, a total of 1862 clinical photographs from 131 patients who underwent intraoral free flap reconstruction were analyzed. Based on final flap outcomes and expert evaluation, images were classified into three ordinal categories: “Normal,” “Suspicious,” and “Compromised.” The ordinal classification model was developed using a novel approach using dynamic triplet margin loss technique. Using 1489 images for training and 373 for testing, our model achieved 0.9571 accuracy, significantly outperforming the baseline model (0.8847). The model showed F1 scores of 0.98, 0.85, and 0.73 for normal, suspicious, and compromised classes respectively, with AUC values exceeding 0.97 for all classes.</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

Deep learning-based ordinal classification overcomes subjective assessment limitations in intraoral free flap monitoring

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年12月25日

查看全文
  • Hyounmin Kim,
  • Dongwook Kim,
  • Juho Bai

摘要

While free flap reconstruction is essential for repairing post-surgical defects or oncologic defects, current monitoring methods lack standardization and rely on subjective assessments. This study aims to introduce an artificial intelligence model that quantitatively analyzes and classifies the status of intraoral free flaps to enhance clinical monitoring capabilities. In this study, a total of 1862 clinical photographs from 131 patients who underwent intraoral free flap reconstruction were analyzed. Based on final flap outcomes and expert evaluation, images were classified into three ordinal categories: “Normal,” “Suspicious,” and “Compromised.” The ordinal classification model was developed using a novel approach using dynamic triplet margin loss technique. Using 1489 images for training and 373 for testing, our model achieved 0.9571 accuracy, significantly outperforming the baseline model (0.8847). The model showed F1 scores of 0.98, 0.85, and 0.73 for normal, suspicious, and compromised classes respectively, with AUC values exceeding 0.97 for all classes.