Last year, RSNA included over 130 AI companies showcasing their radiology software, and this was only the tip of the iceberg as many companies are not yet ready for show-time. The vast majority of these companies focuses on image analysis applications - they detect lesions, measure a volume or provide an image-based risk score. However, interestingly, there is a wide range of opportunities to improve the image-based diagnostic workflow even before the images are ready for reporting and potentially even before reconstruction takes place . Specifically in the context of prostate MRI, what are the various value-adds that AI can offer? How can AI be leveraged to speed improve accuracy and efficiency of the diagnostic pathway? Let’s look at this question from a few angles - walking us gently through the imaging process from pre-exam preparation to scan acquisition, image reconstruction and reading preparation.
If a PSA test raises enough suspicion for further exams, a urologist can request an MRI (which is usually a click of a button exercise). This is easy, sure, but it also ignores the opportunity to optimize the imaging request. What if we could implement an AI-based risk model that determines the best next step in the diagnostic process which offers the most diagnostic value? This AI model can use all available patient data, such as PSA and DRE test results, age, patient history, family history, co-morbidities to recommend an abbreviated MRI protocol for one patient, an extended protocol for the next and perhaps non-imaging tests (instead of MRI) such as Stockholm 3 for yet another patient. This avoids a one-size-fits-all approach and has the potential to make the entire diagnostic pathway significantly more accurate and effective.
I can already hear some physicians say that they are already taking this personalized approach and that they do not need AI for this. Here is the caveat: AI can base its assessment on much larger patient databases. A good physician has probably seen hundreds perhaps thousands of cases and most likely she has read-up on the latest literature that draws conclusions on thousands more. AI has the power to base decisions on orders of magnitude more information and arrive more consistently to the right conclusion.
An MRI machine contains a lot of hardware, including the B0-field magnet, gradient coils, and radiofrequency coils, which are all produced with the greatest of care and precision. However, it is a known fact that imperfections remain, causing distortions in the magnetic field. These small deficiencies (sometimes not so small) hamper image quality. One option is to keep optimizing the hardware, which is a route taken by many manufacturers that create stronger magnetic fields with more stable gradients. However, another option lies in the development of smart software for optimization of, for example, the different gradient fields. AI can play an interesting role in this context. Imagine a scanner-specific algorithm that is trained to correct the shortcomings of that machine. Instead of buying state-of-the-art MRI scanners, it might be a much more efficient approach to invest in AI-based finetuning of the acquisition and reconstruction protocols improving signal-to-noise ratio and reducing scanner artifacts.
Image quality assessment is another application of AI during scanning that is worth consideration. A lot of time is wasted because image quality is not up to clinical standards causing hinderance of the radiologists reading of the scan. Hence, a repeat scan cannot be avoided. This does not only cost precious scanner and physician time, but the diagnosis is also delayed, the patient needs to revisit the hospital, and costs are often not eligible to expense a second time. In other words, this is a scenario that every hospital wants to avoid. The good news is that AI can also provide a solution to this problem. While a patient is still on the table an AI algorithm could alert the scanner operator of susceptibility artifacts caused by air in the bowels or movement artifacts. If the image is not sufficient for clinical assessment, the exam can be (partially) restarted and better images can be obtained - avoiding the hassle of repeating the full exam on another day.
When observing a radiologist at work, it is clear that in many cases not all sequences are used in making a decision on a next step. This is especially true for normal cases where often one or more acquired sequences are not used. For example, in prostate MRI, only cases that show an intermediate risk (PI-RADS III) on the diffusion weighted sequence will require (according to the protocols) a contrast series to provide additional information. It would be very helpful if the scanning protocol starts with a minimum set of sequences and only extends the protocol to a contrast sequence when situations such as the one above are detected.
Another way to optimize the scanning protocol is by accelerating the scan acquisition itself. AI algorithms can support faster acquisitions by, for instance, filling k-space in an optimized way. An example of such a technique is compressed sensing, where you only acquire part of k-space and, with the use of deep learning-based reconstruction, generate an image of non-inferior quality to the traditional acquisition.
As mentioned before, k-space can be acquired partially on purpose, after which AI algorithms can reconstruct an image that is up to clinical standards. However, it can also occur that k-space has not been fully filled due to complications that occurred during acquisition. Also in this case, AI software may be able to work its magic and provide an output image of sufficient quality for a radiology reading.
Even when k-space is fully filled, we can still apply an AI algorithm that optimizes the image reconstruction. Different types of artefacts can distort the image. A well-trained AI algorithm can learn how to correct for these artefacts while creating the final image from k-space data - an optimization step that seems a waste not to implement!1
AI offers many opportunities to improve the MRI imaging process for prostate cancer patients. Whether it is before the actual MRI exam takes place, while the patient is in the scanner, or during reconstruction, we have looked at a range of promising options. Even then, we did not touch on many of the things AI can offer in the context of reading prep, for example, through AI support for image reading, automated report creation, or by diving into the world of image analysis and providing valuable input for radiologists to support their diagnosis process. AI will supercharge the future of radiology.