Deep Learning in Radiology: Now the Real Work Begins.
J Am Coll Radiol. 2018 Feb;15(2):364-367. doi: 10.1016/j.jacr.2017.08.007. Epub 2017 Dec 29. Lugo-Fagundo C1, Vogelstein B2, Yuille A3, Fishman EK4.
Machine learning systems build statistically based mathematical models that program computers to optimize a performance criterion from sample data [1. Deep learning is a form of machine learning that attempts to mimic the interactivity of layers of neurons in the neocortex, the part of the cerebral cortex where most higher thinking occurs [2. Deep learning methods can now be used to recognize patterns in digital representations of sound and images as well as in other data types.
Recognizing its power, medical researchers and scientists have been exploring the role of deep learning methods to create tools to alert clinicians to focus on outliers identified in data sets and as aids for earlier and more accurate diagnosis. Recent studies have highlighted the potential power of deep learning in a variety of medical applications such as imaging-guided survival time prediction of patients with brain tumors [3, bladder cancer treatment response assessment [4, lung nodule [5 and mammographic classifications [6, diagnosis of breast lesion pathologies [7, and body part recognition [8. However, many studies fail to emphasize the importance of the data collection required to build the algorithms.
In general, the scientific community agrees that large volumes of data are required for building robust deep learning algorithms [9, 10]. Most recently, Esteva et al developed an algorithm to separate gross images of skin lesions into nonproliferative benign or malignant lesions [9. The investigators utilized a set of 129,450 images of skin lesions for computer training; 2,032 of those images were of skin diseases diagnosed using noninvasive visual analysis or biopsy testing by dermatologists [9. Esteva et al specified that the machine’s performance can be enhanced if it is trained with a larger data set. Similarly, for deep learning to be successful in radiology, it will be necessary to collect and annotate potentially thousands of cases representing a range of pathologies.
These detailed data are used to create deep networks that are trained in a supervised way, with a method that uses more knowledge and information. We have previously proven that weak supervision, which specifies a bounding box around an object instead of per pixel labeling, does not work for deep network building of complex data [11. This method might function in simpler tasks like recognizing a dog from a cat, but not for compound data like pancreatic anatomy and pathology.