Using the Triplet Loss for Domain Adaptation in WCE

Abstract

Wireless Capsule Endoscopy (WCE) is a minimally-invasive procedure that, based on a vitamin-size camera that is swallowed by the patient, allows the visualization of the entire gastrointestinal tract. This technology was developed 20 years ago to perform useful and safe studies of different bowel disorders. However, especially the number of captured images and their difficult interpretation has hindered its deployment in some clinical scenarios. Deep learning methods have the necessary capacity to deal with WCE image interpretation, but training good models is still an open problem for some bowel disorders due to the fact that obtaining a sufficiently large set of positive cases, for the creation and validation of the model, is an arduous task. Moreover, technological advances are rapidly moving forward proposing new hardware able to obtain images with a substantially improved quality. Given these two facts, it is obvious that highly accurate models can only be built by considering heterogeneous datasets composed of images captured by different cameras, and if training methods are able to find invariances with respect to the image acquisition systems. In this paper, we study the use of deep metric learning, based on the triplet loss function, to improve the generalization of a model over different datasets from different versions of WCE hardware. The obtained results show evidence that with just a few labeled images from a newer camera set, a model that has been trained with images from older systems can be easily adapted to the new environment.