Raw LLM Responses

Hello Dr. Cellini, As somebody pursuing a PhD in medical image analysis (the branch of AI that applies AI to problems in radiology), I agree with your points. Artificial intelligence is unlikely to replace radiologists any time soon and we could probably use more radiologists rather than fewer. In my opinion, AI will not render radiologists obsolete for several reasons. First, the performance of the algorithms depends very much on the modality that is being used and the image analysis task that is being considered. For example, if you want to delineate thrombi on NCCT/CTA with an algorithm, this is a very difficult task due to the object being very small, image artefacts making the scan less clear and sometimes no hyperdense artery sign being present. If however, you want to delineate very large tumors in the brain on MR scans this is a very easy task because they are large and easily visible. The latter problem would be a fairly easy to publish research paper and would draw attention from the media, but not the former problem. So this creates the impression that AI is good at everything, even though it is not. Second, there is a problem in AI that is called the domain shift problem. To put this into a medical context, if you train on a dataset that was acquired using machine A and you report your results on data acquired on the same machine it will get a certain performance. If however, you change the machine that you use, it will get a lower performance due to the images looking slightly different. This would render some algorithms that work well on relevant problems useless until they are re-trained on data that includes scans made on this new machine. Third, the data leakage problem. Many early research applying AI (deep learning specifically) included people that did not have much experience working with medical data. These were people coming from computer vision where you work with individual images (most of the time) rather than patient data. As a result, specifically CT and MR scans were not correctly distributed to training, validation and testing sets. The correct way of doing so would be assigning patients to splits. What would happen was all slices of all of the the scans would be shuffled and assigned to splits. This means that highly correlated slices would be trained and tested on, leading to inflated results in some studies. Now to provide some context, I got this information from a colleague working in the same lab as I am and did not read the original paper, but I have seen this problem in a number of studies myself. Fortunately, this is improving with better practices and the field maturing more. To end on a positive note, there are several areas that, in my opinion, AI will be very helpful for. A great example is research. If an image feature is labor intensive to annotate but could be a possible alternative endpoint AI can definitely be of use here, especially when you are working with large-scale clinical trials or registries. This makes research into the efficacy of medication and treatment much more feasible. Another application would be to allow for quantification of image features that inform treatment or are prognostic, but that are too labor intensive for a radiologist to do. Volume measurements are a good example. A third line of research that I find quite promising is that of image reconstruction. Reconstructing an MRI image from K-space is very time consuming. Thus, in addition to MRI having several contra-indications, acquisition and reconstruction time are limiting factors. There is research that is working on requiring less samples from K-space, which speeds up acquisition. Finally, AI can help make the image data more easily readable for a radiologist. For example by registering images more closely or extracting center-lines to "stretch" out arteries. Hope you find my response helpful. Kind regards, Riaan