Pseudo Random Forests for Tube Identification

Abstract

Random forests are widely used in machine learning as they can potentially offer higher accuracy than individual decision trees by the averaging of multiple independent models. We propose a modification called “Pseudo-random forests” that combines stochastic feature selection with dynamic problem-specific feature generation. As proof of concept, we apply the method to the problem of edge detection and classification in radiographic images. In particular, we use the method to detect feeding tubes in pediatric patients, which are inserted to deliver food and medicine. Since multiple layers of tissues and medical objects are overlaid in a single image, these can be difficult to read on x-rays, even for trained radiologists. The placement of these tubes is critical to the well-being and care of the patient. Automating the recognition of these tubes can help confirm the correct placement of these tubes, as an improperly placed tube could delay treatment or jeopardize the health of the patient. It can also save time by enhancing the visibility of tubes for interpretation by radiologists, as hospitals may have to validate tens to hundreds of these x-rays a day. We report an average recall of 85% for tube pixel identification by using Pseudo-random forests for classification, based on leave-out-one cross-validation. Further improvement is possible by post-processing for tube continuity and the incorporation of other techniques developed as part of the research.