Deep Learning for Medical Image Analysis: Applications to Computed Tomography and Magnetic Resonance Imaging

Following the recent development in artificial intelligence, where deep learning has become the main methodology, the paradigm of medical image analysis is shifting from the previous clinical experience and knowledge-based feature engineering to the data-driven objective feature analysis of deep learning. Especially, as the application of various techniques developed for natural images to medical images is being accelerated, we are no longer simply adapting the natural image models to medical images but developing new methods, which encompasses the unique characteristics of the medical image domain. Furthermore, as the research on interpretability of decisions made by deep learning models and the way of incorporating clinical knowledge into the model progresses, we have started to obtain promising results that will allow clinical implementation of deep learning. Among various deep learning models, convolutional neural networks (CNN) have become methodology of choice for visual recognition problems. CNN is a type of feed-forward artificial neural network, which learns hierarchical features by iterating convolution and pooling layers until the output prediction layer is reached. While the convolution layers learn specific patterns in the input or intermediate feature map with locally-connected shared weights, pooling layers reduce the feature map by spatially aggregating activations. In special cases where the output of the model is same as the input or its denoised version, we call the model as convolutional auto-enconder (CAE). In medical image analysis, machine learning methods have been used in various fields such as detection and classification Corresponding Author: Kyu-Hwan Jung VUNO Inc., 6F, 507, Gangnamdae-ro, Seocho-gu, Seoul, Korea Tel: +82-2-515-6646 Fax: +82-2-515-6647 E-mail: kyuhwanjung@gmail.com