Artefact detection in video endoscopy using retinanet and focal loss function

Endoscopic Artefact Detection (EAD) is a fundamental task for enabling the use of endoscopy images for diagnosis and treatment of diseases in multiple organs. Precise detection of specific artefacts such as pixel saturations, motion blur, specular reflections, bubbles and instruments is essential for high-quality frame restoration. This work describes our submission to the EAD 2019 challenge to detect bounding boxes for seven classes of artefacts in endoscopy videos. Our method is based on focal loss and Retina-net architecture with Resnet-152 backbone. We have generated a large derivative dataset by augmenting the original images with free-form deformations to prevent over-fitting. Our method reaches a mAP of 0.2719 and a IoU of 0.3456 for the detection task over all classes of artefact for 195 images. We report comparable performance for the generalization dataset reaching a mAP of 0.2974 and deviation from the detection dataset of 0.0859.