Artificial intelligence (AI), founded as an academic discipline in 1956 during the Dartmouth Workshop, has recently undergone one of the most rapid developments in human history. AI is gradually being applied to many fields, including medicine and radiology in particular.
The most frequently asked question in medical imaging is: will AI replace the radiologist one day? It is a legitimate issue, given that, very recently, a Google medical chatbot passed the medical exam in the USA. Yet, the equation is not so simple.
One observation remains certain: AI will simplify the radiologist’s work, by enabling the automation of repetitive and time-consuming tasks, such as identifying patients’ files or segmenting anatomical structures in the image. This will save time, which will be hence devoted to higher value-added activities by the radiologist and, above all, to the patient. AI will also help improve diagnostic accuracy, for example by detecting micro-calcifications in mammograms that are difficult to identify visually, or by reducing the number of false negatives. Time savings and improved accuracy are in line with the demographic evolution of today’s world, which tends towards more patients and fewer radiologists.
Beyond these performances, AI could also contribute to the implementation of more screening campaigns to develop preventive medicine, and to reduce medical deserts.
However, AI imitates a single part of human intelligence – learning, understanding, deciding – and can make mistakes – just like humans, which is a point in common. Algorithms can be biased by the data used to train them, sometimes with a risk of discrimination, particularly against minorities. Besides, data tend to come from people with privileged access to healthcare, with all the problems of representativeness and disparities this entails [1].
Furthermore, the quality of labeling, particularly for segmentation algorithms, is absolutely essential to obtain high-performance results. The famous adage “Garbage in – Garbage out” takes on its full meaning, with all the consequences one can imagine.
As for Deep Learning image reconstruction, while offering very high-quality visualizations, it may fail to capture a small tumor in its reconstructed version. It may also attenuate or even erase anatomical zones and anomalies which, without visual control of the raw data, may be critical for the patient [2]. The ultimate goal is not to obtain a beautiful image, but a clinically relevant one.
These tools must therefore remain what they are: a fantastic way of saving time and improving quality, under human control. As Curtis Langlotz [3] stated, AI will not replace the radiologist. But the radiologist who uses AI will replace the radiologist who does not.
[1] Garin SP, Parekh VS, Sulam J, Yi PH. Medical imaging data science competitions should report dataset demographics and evaluate for bias. Nature medicine. 2023;29(5):1038-1039.
[2] Antun V, Renna F, Poon C, Adcock B, Hansen AC. On instabilities of deep learning in image reconstruction and the potential costs of AI. Proceedings of the National Academy of Sciences. 2020;117(48):30088-30095.
[3] Langlotz CP. Will artificial intelligence replace radiologists?. Radiology: Artificial Intelligence. 2019;1(3):e190058.
