Computer Assisted Diagnosis
Alistair Young1

1Anatomy and Medical Imaging, University of Auckland, Auckland, New Zealand

Synopsis

Artificial intelligence (AI) and machine learning (ML) are advancing rapidly, as evidenced by the recent success of AI systems in the automated diagnosis of cancerous skin lesions from images or autism from brain MRI. This talk will give course participants an overview of current capability and future applications in the fields of image interpretation, classification and analysis.

Highlights

· Artificial intelligence and machine learning methods are advancing rapidly.

· These systems are now essential to add value to the scan, and to reduce costs

· Immediate applications can be found in cases where there is a large amount of “ground truth” for training

Target Audience

Clinicians, physicists and engineers seeking to learn more about the current state of the art in computer aided diagnosis, and prepare for future innovations in this area.

Outcomes/Objectives

Artificial intelligence (AI) and machine learning (ML) are advancing rapidly, as evidenced by the recent success of AI systems in the automated diagnosis of cancerous skin lesions from images [1] or autism from brain MRI [2]. This talk will give course participants an overview of current capability and future applications in the fields of image interpretation, classification and analysis.

Methods

AI encompasses a wide field, from image classification [1], to playing poker [3]. Neural nets, support vector machines, random forests, probabilistic boosting trees, and ridge regression methods have all been applied to automatically generate features relevant to interpretation. The advances in the last few years have been driven by improved processing power and improved learning algorithms. These methods are not limited to image interpretation. MR fingerprinting, for example, can use ML methods to generate parametric maps from sparse sampling.

Results

Deep convolutional neural networks perform very well in classification and identification tasks that have a lot of ground truth (annotated images). The need for substantial amounts of training data limits some of these methods in wider applications. However, several multinational companies now have access to very large data repositories. Alternatively, the business model for several startup companies involves cloud processing of data, where the system continually learns as more data is uploaded. Although this technology has been applied as a post-processing tool, it can just as well be applied to the image acquisition process. In many cases, using MRI to get a human-interpretable image may not be the most cost effective use of a scanner. If MRI is replaced by MR-AI, scanners may not need to make images at all, and the cost of hardware may fall accordingly. Challenges now and in the future will involve incorporation of these methods into the clinical workflow, and critical examination of results.

Conclusions

AI systems are poised to disrupt all aspects of ISMRM activities, from image interpretation to scanner design. However: “Natural stupidity, rather than artificial intelligence, remains the greatest risk” [4].

Acknowledgements

No acknowledgement found.

References

[1] A. Esteva, B. Kuprel, R. A. Novoa, J. Ko, S. M. Swetter, H. M. Blau, et al., "Dermatologist-level classification of skin cancer with deep neural networks," Nature, vol. 542, pp. 115-118, Feb 02 2017. [2] H. C. Hazlett, H. Gu, B. C. Munsell, S. H. Kim, M. Styner, J. J. Wolff, et al., "Early brain development in infants at high risk for autism spectrum disorder," Nature, vol. 542, pp. 348-351, Feb 15 2017.

[3] M. Moravcik, M. Schmid, N. Burch, V. Lisy, D. Morrill, N. Bard, et al., "DeepStack: Expert-level artificial intelligence in heads-up no-limit poker," Science, Mar 02 2017.

[4] S. Cave. (2017, 28 March). Intelligence: A history. Available: https://aeon.co/essays/on-the-dark-history-of-intelligence-as-domination

Proc. Intl. Soc. Mag. Reson. Med. 25 (2017)