Scholarlink

Background <p>Magnifying endoscopy is a reliable method for the differential diagnosis of colorectal tumors due to its high resolution and enhanced contrast, which allow for more precise detection and characterization. However, accurate assessment of tumor differentiation still requires significant expertise. To address this, we developed an AI-assisted diagnosis model (AADM) based on the Japan NBI Expert Team (JNET) classification and evaluated its diagnostic performance in comparison to endoscopists with varying levels of experience.</p> Methods <p>A total of 2645 magnified images from 219 patients were collected for training and testing the model. The study employed an improved DeepLabV3 + model, a deep learning network that combines an encoder-decoder architecture with Atrous Spatial Pyramid Pooling (ASPP). We compared the model with six endoscopists with varying experience.</p> Results <p>The AADM achieved an accuracy of 0.938 (Macro/Micro-average), with type-specific accuracies of 0.978 for Type 1, 0.891 for Type 2A, 0.912 for Type 2B, and 0.971 for Type 3. Sensitivity was 0.915/0.876 (Macro/Micro-average), and specificity was 0.954/0.959 (Macro/Micro-average). The area under the receiver operating characteristic curve (AUC) was 0.935/0.917 (Macro/Micro-average). The diagnostic performance of the AADM was superior to that of junior endoscopists (accuracy: 0.938 vs. 0.866, <i>P</i> &lt; 0.05) and comparable to senior endoscopists (accuracy: 0.938 vs. 0.937, <i>P</i> &gt; 0.05). Additionally, junior endoscopists significantly improved their accuracy from 0.866 to 0.951 after receiving assistance from the AADM (<i>P</i> &lt; 0.01).</p> Conclusion <p>The AADM demonstrated good diagnostic capability for JNET classification, facilitating more convenient and accurate diagnoses of colorectal tumors. AADM shows potential for optimizing surgical approach selection in colorectal tumor management. Further studies are needed to assess its effectiveness and cost-efficiency in real-world clinical 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

Application of AI-assisted magnifying colonoscopy system in the diagnosis of colorectal tumors: a multicenter exploratory diagnostic study

  • 期刊论文

  • 发表日期：

    2025年11月14日

查看全文
  • Xiaofei Fan,
  • Ayixie Maihemuti,
  • Xiaohan Cai,
  • Jia Huang,
  • Siqi Zhao,
  • Tao Wang,
  • Xin Chen,
  • Haoyu Zhai,
  • Jia Li,
  • Zhongqing Zheng,
  • Chunshan Zhao,
  • Yufeng Wang,
  • Yue Feng,
  • Jinbao Mu,
  • Xiangdong Lu,
  • Hongbin Zhu,
  • Bangmao Wang,
  • Wentian Liu

摘要

Background

Magnifying endoscopy is a reliable method for the differential diagnosis of colorectal tumors due to its high resolution and enhanced contrast, which allow for more precise detection and characterization. However, accurate assessment of tumor differentiation still requires significant expertise. To address this, we developed an AI-assisted diagnosis model (AADM) based on the Japan NBI Expert Team (JNET) classification and evaluated its diagnostic performance in comparison to endoscopists with varying levels of experience.

Methods

A total of 2645 magnified images from 219 patients were collected for training and testing the model. The study employed an improved DeepLabV3 + model, a deep learning network that combines an encoder-decoder architecture with Atrous Spatial Pyramid Pooling (ASPP). We compared the model with six endoscopists with varying experience.

Results

The AADM achieved an accuracy of 0.938 (Macro/Micro-average), with type-specific accuracies of 0.978 for Type 1, 0.891 for Type 2A, 0.912 for Type 2B, and 0.971 for Type 3. Sensitivity was 0.915/0.876 (Macro/Micro-average), and specificity was 0.954/0.959 (Macro/Micro-average). The area under the receiver operating characteristic curve (AUC) was 0.935/0.917 (Macro/Micro-average). The diagnostic performance of the AADM was superior to that of junior endoscopists (accuracy: 0.938 vs. 0.866, P < 0.05) and comparable to senior endoscopists (accuracy: 0.938 vs. 0.937, P > 0.05). Additionally, junior endoscopists significantly improved their accuracy from 0.866 to 0.951 after receiving assistance from the AADM (P < 0.01).

Conclusion

The AADM demonstrated good diagnostic capability for JNET classification, facilitating more convenient and accurate diagnoses of colorectal tumors. AADM shows potential for optimizing surgical approach selection in colorectal tumor management. Further studies are needed to assess its effectiveness and cost-efficiency in real-world clinical settings.