Gary Birnbaum, MD
MRI versus AI
Updated: Mar 27, 2019
Automated Image Quality Evaluation
of Structural Brain MRI Using an Ensemble
of Deep Learning Networks
Sheeba J. Sujit, PhD, Ivan Coronado, MS, Arash Kamali, MD, Ponnada A. Narayana, PhD, and Refaat E. Gabr, PhD
J Magn Reson Imaging. 2019 Feb 27. doi: 10.1002/jmri.26693. [Epub ahead of print]
1. Artificial intelligence has received a lot or press recently regarding possible application to medicine.
2. Using computer programs called “neural networks” it has been possible to “teach” computers to recognize abnormalities of medical images, such as pictures of the back of the eye (the retina).
3. This is done via a process called “deep learning,” in which the computer is shown thousands of normal and abnormal images and “learns” what is normal and abnormal.
4. The program can decide whether a particular image is abnormal within fractions of a second, allowing a radiologist to quickly verify the choice without first having to find it in a group of normal images.
5. The important questions are: a) how accurate and sensitive are the computer’s choices in determining an image is abnormal? b) how specific are they for the changes that are being looked at? C) and how do the choices made compare with those made by a trained neuro-radiologist?
6. This paper describes the use of computer “deep learning” to evaluate the quality of central nervous system MRIs from persons with either autism or MS.
7. The specificity of the program (the ability to detect good MRI image quality) was good, as was the accuracy (the ability to distinguish good from bad images), but the sensitivity of the program (the ability to separate patients from one another) was poor.
8. This paper does not address the complex issue of using artificial intelligence and deep learning to identify MRIs that meet diagnostic criteria for MS, from those that don’t. (https://www.ncbi.nlm.nih.gov/pubmed/26822746)However, this paper is an important first step in the development of computer programs that could do this.
9. If successful, and if results agree with those of trained neuro-radiologist, the accuracy and speed of diagnosing MS should improve greatly.
The field of artificial intelligence (AI) has been revolutionized by the use of computer programs called “neural networks” that can “learn” from being shown large amounts of information, a process called “deep learning”. Such programs have been shown to beat world champion chess players and world champion players of Go. Their use in medicine is now being applied to the evaluation of medical images, such as determining whether images of the back of the eye (the retina) are abnormal and what may be causing the abnormality (e.g. diabetes, bleeding, a tumor). “Teaching” a computer to identify normal from abnormal images, and to distinguish different kinds of abnormalities, is a long and difficult task, requiring tens of thousands of images from many groups of individuals. However, once “taught” such programs can quickly identify individuals that need more careful evaluation and hopefully also lead to more accurate diagnosis.
While “deep learning” has been of great benefit in evaluating eye images obtained in emergency rooms, and may revolutionize the field of ophthalmology, its use in evaluating the more complex, 3D images of the brain is just developing. The current paper is one of the most recent to use AI and deep learning to identify changes in brain MRIs, this case, those that are of poor quality and need to be retaken. While the program was generally successful, 25% of the computer decisions were still different from those of two independent radiologists. Thus, while the techniques being developed are powerful and of enormous potential, their use in clinical practice remain to be developed.
The abstract of this article is available.