Multi-Instance Multi-Stage Deep Learning for Medical Image Recognition

Abstract Over the last few decades, as the amount of annotated medical data is increasing speedily, deep learning-based approaches have been attracting more attention and enjoyed a great success in the medical imaging field, including computer-aided diagnosis, image segmentation, image registration, image database retrieval, and so on. In medical image recognition problems, by using beautiful biologically-inspired architectures, deep learning is able to learn a hierarchical representation of data to distinguish different image classes. However, if the discriminative information only lies in local image patches, classic deep learning framework may still have limitations in discovering them without local-level annotations. In this chapter, we introduce a multi-stage deep learning framework that aims to automatically discover local discriminative information for medical image classification and apply it on body part recognition in CT. Specifically, the pre-train stage of the proposed framework automatically discovers the local regions that are either discriminative or non-informative to the image classification problem by learning a convolutional neural network (CNN) in a multi-instance learning fashion. Then, the pre-learned CNN is fine-tuned based on these discovered local regions to produce an efficient image-level classifier. In this way, no manual annotation of local regions is required. Our method is validated on synthetic datasets and a large scale CT dataset with several thousands of images. It achieves better performances than conventional learning approaches using ad-hoc designed image features and the standard deep CNN.