Nuit Blanche: A Perspective on Deep Imaging

Monday, March 06, 2017

A Perspective on Deep Imaging

In the previous entry on The Great Convergence goes mainstream: CfP, "Machine Learning for Image Reconstruction" in IEEE Trans. on Medical Imaging. I noted the reference to this interesting paper: A Perspective on Deep Imaging by Ge Wang (and yes IEEE seems to have an Open access capability). Here is the abstract:

The combination of tomographic imaging and deep learning, or machine learning in general, promises to empower not only image analysis but also image reconstruction. The latter aspect is considered in this perspective article with an emphasis on medical imaging to develop a new generation of image reconstruction theories and techniques. This direction might lead to intelligent utilization of domain knowledge from big data, innovative approaches for image reconstruction, and superior performance in clinical and preclinical applications. To realize the full impact of machine learning for tomographic imaging, major theoretical, technical and translational efforts are immediately needed.

I note the following tidbit on the parallels between compressive sensing and deep learning:

As another example, in the area of compressed sensing it was shown that while compressed sensing produced visually pleasing images, tumor-like features were hidden or lost [41]. Nevertheless, those features were constructed based on the known imaging geometry and algorithmic details, which would not likely be encountered in clinical settings. Indeed, most theoretical analyses on compressed sensing methods suggest the validity of the results with the modifier “with an overwhelming probability”, such as in [42]. Actually, multiple iterative image reconstruction algorithms for medical imaging already have CS components and show excellent results. As long as a method most likely delivers decent results, it is a great tool unless we have an even better method.
Still, there are more theoretical limitations of compressed sensing that have yet to be resolved. When the claim was made that compressed sensing generates valid results “with an overwhelming probability”, important caveats cannot be ignored. Especially, the problem sizes need to be large for most theoretical results to become valid, and the probabilistic sampling schemes have to be designed according to distributions that may not be easily verifiable. Even if there is a high-probability of ”success” in the theoretical settings, the involved constants of proportionality are not always favorable. Although the current theory cannot give the imaging performance guarantee for most medical imaging problems, the theoretical insights have enabled a large range of applications.
Overall, we feel that the story for deep learning will be similar to that for compressed sensing; that is, dependably-good results are feasible in the absence of full-fledged theory, and eventually we will have a satisfactory theory. Encouragingly, good results are constantly emerging such as [43].

Two items:

I think Deep Learning will prove itself on an empirical basis without ever needing a theory or a full fledged theory. Compressed sensing existed before the papers of Tao, Romberg, Candes and that of Donoho. It just was field specific and was used because it yielded interesting results in those fields. Currently Deep Learning is already accepted as a de facto better technique and will get into medical imaging without the need for a theory like it was required for compressive sensing. By that I mean it is likely that it will be easier to publish in the medical imaging literature with a Deep Learning approach now than it was to publish in that literature on compressive sesning before the 2004-2006 papers.
As mentioned yesterday in Sunday Morning Insight: How can you tell the world is changing right before your eyes ?, we seem to already have some very interesting success as mentioned on the Google blog on Friday with Assisting Pathologists in Detecting Cancer with Deep Learning by Martin Stumpe and and Lily Peng.
Ge Wang who's outstanding work we have covered here on Nuit Blanche before, states in his bio:

First paper on spiral cone-beam CT (1991) to solve the long object problem. Cone-beam CT scanners perform 100-million scans annually in the world.

100 million scans is not impressive, you know what's impressive ? a billion scans a year :-) Let's hope that he and others using Deep Learning get to these scales.