Automated image analysis for esophageal cancer detection using volumetric laser endomicroscopy


Volumetric Laser Endomicroscopy (VLE) is a recent endoscopic imaging technique that allows gastroenterologists to inspect the inner tissue layers of the esophageal wall. This provides endoscopists with a very useful tool for the detection of early esophageal cancer, which typically develops in these subsurface layers that cannot be imaged by traditional endoscopic imaging techniques. During a VLE image capturing procedure, a balloon is endoscopically inserted in the esophagus. Next, the balloon is inflated to stretch the tissue into a fixed position after which a laser scans a cylindrical segment of up to 6 cm in 90 seconds, resulting in 1,200 circumferential scans of 4,096 × 1,024 pixels.

The fast and efficient procedure VLE offers, is particularly beneficial for patients with a so called Barrett’s Esophagus (BE). In patients with BE, due to consistent gastric reflux, a segment of the esophagus proximal to the gastro-esophageal junction is covered with an acid resistant tissue-type that is not native to the esophagus. This process results in an over 30 times higher chance of developing esophageal adenocarcinoma, which accounts for the majority of esophageal cancers in the Western World.

Early cancer in BE has proven to be very hard to recognize using traditional white-light endoscopy. Hence, over the past years, studies have presented automated analysis algorithms for endoscopic imagery in order to aid the physician in detecting early esophageal cancer. Although these methods might increase the detection ratio of early Barrett’s cancer, they suffer from a major drawback: what is not imaged by the endoscopists, is not processed by the detection algorithm. Consequently, if a BE segment is overlooked, developing cancers might go unnoticed. Therefore, VLE offers a promising alternative, as it can image the entire Barrett segment in a single scan. On top of that, it allows the analysis of the subsurface tissue layers, which is impossible with conventional white-light endoscopy.


Although the above-mentioned aspects of VLE are very attractive, there are two major difficulties that this novel technique faces: (1) physicians struggle to interpret the complex VLE signal and (2) carefully inspecting 1,200 high-resolution VLE slices is very laborious and time consuming. Therefore, in this project, we aim to develop a Computer-Aided Detection (CADe) algorithm that automatically inspects a full VLE volume for signs of early cancer. This involves a number of research activities, which are listed below.
• Efficient implementation of adequate preprocessing to address the large and noisy VLE signal;
• Identification of discriminative features and machine learning methods for robust and accurate cancer detection;
• Evaluation of the developed cancer detection methods on in vivo VLE data and a comparative validation to alternative approaches proposed in literature.
Recent publications from our group show the feasibility of CADe algorithms for VLE on ex vivo data and the employed Matlab framework is available for the student. However, for investigation the potential of Convolutional Neural Networks (CNN) for our purpose, other programming environments should be used instead, such as TensorFlow and Python.

This project will be executed internally at the TU/e.
Machine learning, Image analysis, Computer-Aided Diagnosis (CAD), Optical Coherence Tomography (OCT), Healthcare/oncology
Contact person: