Proposing Novel Data Analytics Method for Anatomical Landmark Identification from Endoscopic Video Frames.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Gastroenterology

Background:

• Anatomical landmarks are crucial for guiding gastroenterologists during endoscopy and ensuring complete gastrointestinal tract (GI) examinations.
• Automatic detection of these landmarks in endoscopic videos can significantly aid physicians during GI screening.

Purpose of the Study:

• To present a novel, automatic method for detecting anatomical landmarks in GI endoscopic video frames.
• To compare the performance of a semisupervised deep convolutional neural network (CNN) with a supervised CNN model for this task.

Main Methods:

• Utilized a semisupervised deep convolutional neural network (CNN) for anatomical landmark detection in endoscopic video frames.
• The Kvasir dataset, containing images of Z-line, pylorus, and cecum, was used for training and evaluation.
• Compared the semisupervised CNN model against a supervised CNN model.

Main Results:

• The supervised CNN model achieved 100% accuracy.
• The proposed semisupervised CNN demonstrated competitive performance, achieving average accuracies of 83%, 98%, 99%, and 99% with 1%, 5%, 10%, and 20% labeled training data, respectively.
• The semisupervised approach achieved high accuracy with significantly less labeled data.

Conclusions:

• The proposed semisupervised method offers a highly accurate solution for anatomical landmark detection in GI endoscopy.
• This approach reduces the need for extensive data labeling, saving labor, cost, and time.
• The method effectively aids physicians in GI screening by providing accurate landmark identification.