Demarcation Line Determination for Diagnosis of Gastric Cancer Disease Range Using Unsupervised Machine Learning in...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Gastroenterology

Background:

• Accurate demarcation line (DL) identification is crucial for diagnosing cancerous lesions in M-NBI endoscopy.
• Novice endoscopists often struggle with precise DL determination.
• Automated DL identification can improve diagnostic accuracy and consistency.

Purpose of the Study:

• To develop and evaluate an unsupervised machine learning method for automatically determining the DL in M-NBI images.
• To assess the accuracy and reliability of the proposed automated system compared to expert endoscopists.

Main Methods:

• An unsupervised machine learning approach was employed, involving superpixel segmentation (Simple Linear Iterative Clustering).
• Image features were extracted per superpixel, followed by k-means clustering.
• Cluster boundaries were identified as DL candidates, and performance was evaluated using 23 M-NBI images.

Main Results:

• The automated system produced DLs comparable to those identified by experienced endoscopists, with calculated Euclidean distances.
• The system's ability to generate pathologically valid DLs was confirmed by adjusting the number of clusters.
• Average Euclidean distances indicated a high degree of similarity between AI-identified and expert-identified DLs.

Conclusions:

• The developed machine learning system effectively determines accurate DLs in M-NBI images.
• This automated tool can serve as a valuable aid for training inexperienced endoscopists.
• The system also has the potential to enhance the diagnostic knowledge of experienced clinicians.