Perspectives on the usage of AI in Medical Image Analysis.

Having worked in the healthcare industry on the medical image analysis area for close to 8 years I developed certain impressions on the applicability of AI in this area.

One of the most striking things I noticed was the dependence of the AI algorithms on a specific make of the machine.

In general, algorithms developed for one made of a machine do not scale well across manufacturers. This is very unlike the computer vision case where (at least) image classification works well across different cameras. This makes the task of data collection strictly company-specific.

The trouble does not end here of course. Keep in mind that the hardware of machines keeps getting upgraded. Algorithms developed for version 1 or 2 usually will not work for versions 3 or 4. This, of course, applies to healthy cases for a normal adult. How do these algorithms scale across ethnicities? Ages of patients? Disease conditions? After all, when you build an algorithm the expectation is that this should work across geographies and pathologies.

Here comes the small question of getting annotations for data. A CT volume for specific anatomy can contain anywhere between 40-50 slices. If the algorithm pertains to say blood vessels, viewing this and annotating it is a very difficult task. Certainly not suitable for a non-specialist (read not a medical doctor).

The cost of annotating such a volume ranges anywhere from 60-80$ per hour. In contrast, even semantic segmentation for a reasonable number of classes is less than a tenth of this cost, considering that it is largely common-sensical and can be done by a non-expert (even a high school education suffice. Furthermore, remember the range of organs to be imaged and the abnormalities to be identified therein.

As we move from CT to MR the range of imaging possibilities can widely vary based on different types of sequences. Things are somewhat simpler for 2D imaging such as X-ray and ultrasound. However, there too expert knowledge is a must when annotating. Considering the above challenges I mention, obviously, this is an expensive affair. Finally, the question is if such algorithms are built, how do the clinicians view them. In my experience, it is a mixed bag.

There is a subset of doctors that are enthusiastic to adopt new technologies, especially if they were influential in the seeding of the ideas for the algorithm. Overall the impression I got was that most of the doctors did not see value in the AI-driven algorithms. All they wanted was to reduce the IT hassles in getting the image to their machine so they could read it. They were also receptive to visualizations which were more intuitive.

Considering the investments to be made in generating ground truth of medical data and the readiness of the hospitals to pay for such algorithms, the cost-benefit great-looking trade-off is terrible. The doctors see value in a machine that can create great-looking images but that is all. However, they do not trust or like to use AI assistance much. On top of it, alarmist comments which say that radiologists will get replaced by AI further alienate the doctors.

The usage of AI, however, is beneficial in other albeit a little less glamorous areas such as fraud detection, optimizations of patient flows, harmonization of flows across multiple vendors, etc.