Towards an Efficient Way of Building Annotated Medical Image Collections for Big Data Studies

摘要: Annotating large collections of medical images is essential for building robust image analysis pipelines different applications, such as disease detection. This process involves expert input, which costly and time consuming. Semiautomatic labeling sourcing can speed up the collections. In this work we report innovations in both these areas. Firstly, have developed an algorithm inspired by active learning self training that significantly reduces number annotated needed to achieve a given level accuracy on classifier. iterative labeling, classifier, testing requires small set labeled at start, complemented with human difficult test cases each iteration. Secondly, built platform scale management indexing data users, well creating assigning tasks contouring big imaging studies. web-based provides tooling researchers annotators, all within simple dynamic user interface. Our annotation also streamlines iteratively algorithms learning/self described here. paper, demonstrate combination proposed workload involved collection cardiac echo images.