A view of AI in healthcare is discerning patterns from a statistically rich set of data. These data are used either in supervised learning where, data are used retrospectively, after a diagnosis or prospectively such as in screening-mammograms.
The process of diagnosis involves acquisition of data, followed by storage and subsequent interpretation by a Physician who may be assisted by AI. All these processes must be performed by doing as little harm to the patient as possible. An example is acquisition of CT or Radiographic images with as low a dose as possible and yet yield images that have a rich diagnostic-content.
We are all familiar with Moore’s-law in computing that deals with the density of electronic components. Beyond Moore’s law is a law postulated by Alvy Ray Smith, one of the founders of Pixar, when he stated that everything that is good in computers improves by a factor of 10, every five years. Since 1969, when we went to the moon, there have been several 5-year increments in computing leading to an exponential explosion in computing power, storage and transmission of information. The increase in computing power, today, is afforded by advances in personal computers and personal devices. The earliest IBM PC had only 64 Kilobytes of memory and had instructions that, each, took a microsecond. The storage was on floppy Disks or a 10 Megabyte hard disk. In comparison, today’s PC’s have several million instructions/second and possess terabytes of memory. If consumers were not interested in faster computing, gaming, greater storage, and very high-speed internet we would not be witnessing the AI revolution.
I will illustrate the influence of consumer-technology with a few examples, in medical imaging.
Digital Radiography (DR) has become ubiquitous in the acquisition, storage and processing of X-Ray images. The acquisition in DR uses technology that has been developed for Digital TV. It is necessary in DR to acquire data from several million pixels in a few seconds to afford resolution that is necessary for a usable image. Digital TVs use technology to address several million pixels, several times a second. These DR-data must be digitized, for every pixel, with a high resolution, to discern diseases. This digitization is now possible using high-speed and high-resolution Analog to Digital converters that emanated from consumer-technology and are miniature. Each medical image involves several megabytes of data which must be stored, processed, and transmitted for analysis and diagnosis. In comparison to the early 80’s, this was not possible at that time.
Another example is in modern CT scanners that acquire data at a very high rate to render 2D or 3D data, almost in real time. To “image wisely” it is necessary to subject patients to as low a dose as possible. The computational-complexity in acquisition as well dose-reduction requires computing power that was unthinkable in the early days of CT, when it took several minutes to acquire and render CT-images due to lack of computing power. AI is being used for dose-reduction by discerning common patterns in projections as well as rendered images.
Yet another example is in the acquisition and rendering of MR-data. There is a computational burden in acquiring and rendering MR-Data, that is usable, without having to subject the patient to long exam-times.
Using AI in medical diagnostics has its own challenge. Each human being is unique in presentation of the same disease. Discerning patterns in a large set of data involves information from several sources that have been stored and transmitted from several locations to effect diagnoses that minimizes false-positives and more importantly, false-negatives. In screening procedures such as mammography, where there is absence of indication of disease, reader-bias can be eliminated or reduced. Screening mammography is an example of unsupervised-learning in contrast with retrospective analysis such as after a biopsy, which can be categorized as supervised-learning. In each of these instances a rich set of data that are statistically significant must be made available to the AI.
I view AI as observing data with several million, different, eyes. Each layer of a neural network involves a few more million more “eyes “after the results of the first look are stored. This process requires high-speed computation, storage and transmission of data from several datasets to eliminate or reduce institutional as well as physician bias in recognizing diseased states, particularly for screening procedures.
AI has made impressive strides today in medical diagnostics. We can only imagine the advances in consumer technology on AI in the next five to ten years where the goodness of computers will increase 10-fold and 100-fold. It is possible that in the next few years AI will augment or replace some of the tasks that are performed by diagnostic-physicians. Prediction is difficult. I could not have predicted in the late 70’s the advances in medical imaging today. In other areas such as reading EKG’s AI has already taken over in diagnostics.
