Endoscopic Landmark Detection

During my undergraduate studies, I contributed to a medical robotics research project on automatic anatomical landmark detection in endoscopic images. The project aimed to support autonomous navigation of gastrointestinal robots by detecting the natural anatomical centre of the pylorus, the opening between the stomach and the duodenum.

Project background

For a gastrointestinal robot to move autonomously from the stomach into the duodenum, it needs to recognise the pylorus and identify a safe passage point. The pyloric centre is an important target because it provides a natural anatomical reference for guiding robot movement while reducing potential soft-tissue injury.

However, endoscopic images are visually complex. The appearance of the pylorus can vary between healthy and diseased cases, and the target structure may change shape, contrast, and orientation. This project therefore explored whether computer vision methods could automatically detect the pylorus centre from endoscopic images.

What we developed

The method combined convolutional neural network classification with classical image-processing filters, including Sobel and Laplace operators. The CNN component was used to classify image regions, while the filtering steps helped refine anatomical boundaries and estimate the pylorus centre.

The algorithm was evaluated across healthy and diseased pylorus images, including multiple anatomical appearances and pathological settings.

Main results

The published study reported an average pylorus-centre detection error of 22.33 pixels, corresponding to a 2.33% relative error against the 960-pixel diagonal length of the endoscopic image. The average processing time was 26.51 ms per image, supporting the feasibility of real-time tracking.