Abnormalities detection from wireless capsule endoscopy images based on embedding learning with triplet loss

Deep learning techniques can accurately detect and grade abnormal findings on images from Wireless Capsule Endoscopy (WCE). However, the prediction accuracy of handcrafted or deep learning in red Lesion, polyp and ulcer diseases is still under investigation. Knowing the utility of an automatic method for abnormalities detection from WCE images and how helpful it might be for the physicians, we proposed a new methodology in approaching this field. In this paper, patches with fixed size are extracted from WCE images, then, encoded using linear projection and position embedding and passed through an embedding model in a forward pass. Moreover, triplet loss is employed to adjust the embeddings. Afterwards, the trained embedding model is exploited for classification. Two strategies are followed in the design of the embedding model namely; training from scratch and fine tuning. The presented scheme, attains satisfactory results in different datasets compared to existing approaches. The detection accuracy has reached 99.9% in some used datasets.