Improved Artifact Detection in Endoscopy Imaging Through Profile Pruning

Endoscopy is a highly operator dependent procedure. During any endoscopic surveillance of hollow organs, the presence of several imaging artifacts such as blur, specularity, floating debris and pixel saturation is inevitable. Artifacts affect the quality of diagnosis and treatment as they can obscure features relevant for assessing inflammation and morphological changes that are characteristic to precursors of cancer. In addition, they affect any automated analysis. It is therefore desired to detect and localise areas that are corrupted by artifacts such that these frames can either be discarded or the presence of these features can be taken into account during diagnosis. Such an approach can largely minimise the amount of false detection rates. In this work, we present a novel bounding box pruning approach that can effectively improve artifact detection and provide high localisation scores of diverse artifact classes. To this end, we train an EfficientDet architecture by minimising a focal loss, and compute the Bhattacharya distance between probability density of the pre-computed instance specific mean profiles of 7 artifact categories with that of the predicted bounding box profiles. Our results show that this novel approach is able to improve commonly used metrics such as mean average precision and intersection-over-union, by a large margin.