Computer Vision News - November 2019

2 Summary Oral Presentation 10 Zongwei Zhou is a third-year PhD student in Biomedical Informatics at Arizona State University, where he is working on deep learning for medical imaging. His advisor is Jianming Liang. He spoke to us about his oral and poster session at MICCAI. Zongwei tells us that this work provides a pre-trained 3D model for medical image analysis. A pre-trained model gives a better starting point. It ’s easier to optimise, eventually leading to better or more stable performance. Nobody has done this before in 3D. Common practice uses Professor Fei- Fei Li’s ImageNet , which has millions of hand-annotated images. Models are pre-trained on that and use transfer learning to the medical domain. On the other hand, this work employs a method that, rather than transfer learning from natural image to medical domain, does it directly from medical domain to medical domain instead. It ’s more straightforward if the models are built in 3D directly because most medical imaging, such as CT and MRI, is in 3D. Another advantage of this method is it doesn’t need any human annotation. Zongwei thinks ImageNet is great, but they had to ask people to label millions of images, whereas this method doesn’t need that. The approach is self- supervised learning. It downloads 3D CT or MRI raw data from the Internet and lets the model learn directly from it. Zongwei explains the practical consequences of this work: “Our work is for disease detection. We don’t focus like doctors do on helping Models Genesis: Generic Autodidactic Models for 3D Medical Image Analysis "We want a machine, image by image, to really quickly and really accurately detect a disease." Best of MICCAI