Applications of the ESPNet architecture in medical imaging

Abstract Medical imaging is a fundamental part of clinical care that creates informative, noninvasive, and visual representations of the structure and function of the interior of the body. With advancements in technology and the availability of massive amounts of imaging data, data-driven methods, such as machine learning and data mining, have become popular in medical imaging analysis. In particular, deep learning-based methods, such as convolutional neural networks, now have the requisite volume of data and computational power to be considered practical clinical tools. We describe the architecture of the ESPNet network and provide experimental results for the task of semantic segmentation on two different types of medical images: (1) tissue-level segmentation of breast biopsy whole slide images and (2) 3D tumor segmentation in brain magnetic resonance images. Our results show that the ESPNet architecture is efficient and learns meaningful representations for different types of medical images, which allows ESPNet to perform well on these images.