Deep Cascaded Networks for Sparsely Distributed Object Detection from Medical Images

Abstract

With the development of deep learning techniques, the performance of object detection has been significantly advanced. Although various methods have been designed to detect landmarks for computer-aided diagnosis, how to efficiently and effectively leverage deep learning approaches to detect sparsely distributed objects, such as mitosis and cerebral microbleeds, from large scale medical images hasn't been fully explored. In this chapter, we introduce a two-stage cascaded deep learning framework, referred as deep cascaded networks, to detect sparsely distributed objects that provide clinical significance with both high efficiency and accuracy. Specifically, the first screening stage with coarse retrieval model rapidly retrieves potential candidates, and subsequently the second discrimination stage with the fine discrimination model focuses on those candidates to further accurately single out the true targets from challenging mimics. Furthermore, we corroborate the importance of volumetric feature representations for volumetric imaging modalities by exploiting 3D convolutional neural networks. Extensive experimental results on the challenging problems, including mitosis detection from 2D histopathological images and cerebral microbleed detection from 3D magnetic resonance images, demonstrated superior performance of our framework in terms of both speed and accuracy.