Medical Image Segmentation Using Deep Neural Networks with Pre-trained Encoders

Abstract

With the growth of popularity of deep neural networks for image analysis, segmentation is the most common subject of studies applying deep learning to medical imaging and establishing state-of-the-art performance results in many applications. However, it still remains a challenging problem, for which performance improvements can potentially benefit diagnosis and other clinical practice outcomes. In this chapter, we consider two applications of multiple deep convolutional neural networks to medical image segmentation. First, we describe angiodysplasia lesion segmentation from wireless capsule endoscopy videos. Angiodysplasia is the most common vascular lesion of the gastrointestinal tract in the general population and is important to detect as it may indicate the possibility of gastrointestinal bleeding and/or anemia. As a baseline, we consider the U-Net model and then we demonstrate further performance improvements by using different deep architectures with ImageNet pre-trained encoders. In the second example, we apply these models to semantic segmentation of robotic instruments in surgical videos. Segmentation of instruments in the vicinity of surgical scenes is a challenging problem that is important for intraoperative guidance that can help the decision-making process. We achieve highly competitive performance for binary as well as for multi-class instrument segmentation. In both applications, we demonstrate that networks that employ ImageNet pre-trained encoders consistently outperform the U-Net architecture trained from scratch.