Artificial Intelligence Algorithms Are Not Clairvoyant

TO THE EDITOR: Artificial intelligence (AI) systems, and computers in general, possess several advantages over humans. They have virtually perfect recall and are not subject to fatigue, mood variations, or environmental biases such as monitor contrast or room lighting conditions. However, they are not clairvoyant. Like us, they are limited by the information provided to them.

Thus, it was with a degree of concern and trepidation that I read the review article “Artificial Intelligence for PET and SPECT Image Enhancement” highlighted in the “State of the Art” section of the January 2024 issue of The Journal of Nuclear Medicine (1). The article states that “supervised deep-learning models have shown great potential in reducing radiotracer dose and scan times without sacrificing image quality and diagnostic accuracy.” However, I believe this is fundamentally impossible. If photon counts are the source of information in a PET or SPECT image, then reducing scan time (and when not above peak noise-equivalent count rate, reducing dose) necessarily means less information about the patient currently being imaged.

I believe it is imperative to keep this simple fact in mind when promoting or evaluating the capabilities of any AI technique. AI models are generally trained using data or images from a separate cohort of patients. In this way, they can add information (prior information, therefore implicitly biased information) when processing a new image set. However, this should not be interpreted as additional information about the current patient. Only additional counts, or other sources of new information about the individual patient, can do that.

If I were to look at a noisy, low-resolution PET or SPECT image, the neural network in my head could imagine (based on prior experience) what it might look like if it were less noisy or had higher resolution. But this does not mean that I am better able to see the lesions that are otherwise buried in the noise. AI techniques can also “imagine,” and provide for us, images that appear enhanced in their resolution and noise levels. However, this raises the question of what image enhancement is in the context of medical imaging. Prettier images do not equate to images with higher levels of useful information. Instead, an AI-enhanced image may mislead a radiologist into thinking the image data contain more information about the patient (commensurate with the perceived noise level and resolution) than is in fact present.

To its credit, the article in question also focuses on the need to switch the assessment from image quality to clinical benefit. This is important and appropriate but does not circumvent the fundamental information limitations described above. Instead, it essentially performs a kind of bait and switch. It is conceivable that an AI-based postreconstruction image enhancement technique might produce images that perform better prognostically (compared with a radiologist viewing the preprocessed images); however, that would mean that the AI algorithm was doing a better job than the radiologist when assessing the original noisy image and then enhancing the image in a way that made this assessment more obvious to the radiologist. At the extreme, the AI algorithm could simply place, for example, a big red X over the lesion. Fundamentally, though, this is lesion detection, not image enhancement. It is important to understand that this is what is going on, not actual improvement in image quality or information content. The radiologist is (effectively) no longer making the assessment. Moreover, applications to other clinical tasks (e.g., calculations of SUV_max, metabolic tumor volume, or total lesion glycolysis) may no longer be accurate.

When processing data to form images, we are often very careful about the type of prior information we incorporate. Generally, we understand the degree (and sometimes direction) of the bias that the prior information imposes (e.g., expectation maximization’s constraint to positive solutions). However, we usually avoid biases in favor of gaining an image assay that is as independent and unbiased as possible. Thus, to the extent possible, the image assay provides completely new information.

For example, in PET image reconstruction of PET/MRI data, we generally forgo using the MRI as a prior even though these images will likely appear lower in noise and higher in resolution. This is because it is understood that biasing the PET image toward the MRI will result in a loss of PET information in precisely the regions containing the greatest amount of new information (i.e., the regions lacking mutual information between the PET and MRI).

There is some space within the context of PET and SPECT image reconstruction (or other means of generating medical images) where it might be appropriate to apply AI techniques. For PET raw or projection data, noise is distributed spatially along each projection. Thus, improving accuracy in one subregion along this projection can improve the accuracy in other regions. For this reason, using the MRI (or AI-derived prior information) during PET image reconstruction is to some extent defendable, whereas reducing scan time and then applying AI-based image enhancement after reconstruction simply is not.

In writing this letter, I do not intend to single out the aforementioned review article or its authors. They are merely reiterating an assumption that has become pervasive within the imaging community, that medical images can be improved without adding new information about the subject at hand. It is this seemingly unquestioned assumption that I am arguing against. Although it may be true in other contexts, it is not true for medical images that are used to assay a patient’s condition. Given the potential risk to the quality of patient care, it has been my intent to be provocative, though hopefully not overly so. I apologize in advance to anyone who might construe this letter to be critical of their work. I have no such intent. I merely wish to kick-start a debate on this important topic.

• © 2024 by the Society of Nuclear Medicine and Molecular Imaging.

• Revision received January 3, 2024.
• Accepted for publication January 29, 2024.