Deep Learning in Radiology: Does One Size Fit All?

Bradley J Erickson,Panagiotis Korfiatis,Timothy L Kline,Zeynettin Akkus,Kenneth Philbrick,Alexander D Weston

doi:10.1016/j.jacr.2017.12.027

Bradley J Erickson, Panagiotis Korfiatis + Show 4 more

Open Access

https://doi.org/10.1016/j.jacr.2017.12.027

Copy DOI

Journal: Journal of the American College of Radiology	Publication Date: Jan 31, 2018
Citations: 101	License type: cc-by

Affiliation: Mayo Clinic

Abstract

Deep learning (DL) is a popular method that is used to perform many important tasks in radiology and medical imaging. Some forms of DL are able to accurately segment organs (essentially, trace the boundaries, enabling volume measurements or calculation of other properties). Other DL networks are able to predict important properties from regions of an image—for instance, whether something is malignant, molecular markers for tissue in a region, even prognostic markers. DL is easier to train than traditional machine learning methods, but requires more data and much more care in analyzing results. It will automatically find the features of importance, but understanding what those features are can be a challenge. This article describes the basic concepts of DL systems and some of the traps that exist in building DL systems and how to identify those traps.

Highlights

HOW MUCH DATA IS ENOUGH?Deep learning (DL) and big data are often confused, because it is thought that any DL system requires huge amounts of data
Deep learning (DL) is a popular method that is used to perform many important tasks in radiology and medical imaging
Traditional machine learning identifies patterns that are present in training sets

Summary

HOW MUCH DATA IS ENOUGH?

DL and big data are often confused, because it is thought that any DL system requires huge amounts of data. One way to reduce the need for training data is to train the first layers of the system on other images that are expected to have similar features that are important—a technique referred to as transfer learning This might sound restrictive, there has been great success in using networks trained on photographic images and training the last few layers to make radiological diagnoses [14,15]. Whether there is significant room for improvement by starting with network trained on medical images is not clear and remains an area of active investigation Another way to reduce the need for training data is to create variants of the original data—a technique called data augmentation [16]. We and others have shown this to be quite effective for medical images

ADAPTING DL MODELS TO BEST ADDRESS A CLASS OF PROBLEMS

CONCLUSION