Anatomical landmarks serve as important reference points that guide gastroenterologists during endoscopic procedures and help ensure the thorough examination of the gastrointestinal tract. The automatic detection of these anatomical landmarks in endoscopic video frames can be a valuable tool for assisting physicians during GI tract screenings.

This study introduces a novel automatic method for detecting anatomical landmarks in the GI tract from endoscopic video frames. The method relies on a semisupervised deep convolutional neural network (CNN) and is compared to the results obtained from a supervised CNN model. The study utilizes anatomical landmarks from the Kvasir dataset, which comprises 500 images for each class of Z-line, pylorus, and cecum. These images have varying resolutions, ranging from 750 × 576 up to 1920 × 1072 pixels.

Experimental findings indicate that the supervised CNN model achieved an outstanding accuracy rate of 100%. Furthermore, our proposed semisupervised CNN model demonstrates competitive performance, albeit with a slight difference compared to the fully supervised CNN model. The semisupervised model, when trained using 1%, 5%, 10%, and 20% of the training data labeled as the training dataset, achieved average accuracy rates of 83%, 98%, 99%, and 99%, respectively.

The primary advantage of our proposed method lies in its ability to achieve high accuracy even with a limited amount of labeled data, eliminating the need for extensive data labeling efforts. This approach not only saves labor but also reduces costs and the time required for data labeling, making it a practical and efficient solution for anatomical landmark detection in endoscopic procedures.