Image Quality in the Era of Artificial Intelligence

Artificial intelligence (AI) is being deployed within radiology at a rapid pace. AI has proven an excellent tool for reconstructing and enhancing images that appear sharper, smoother, and more detailed, can be acquired more quickly, and allowing clinicians to review them more rapidly. However, incorporation of AI also introduces new failure modes and can exacerbate the disconnect between perceived quality of an image and information content of that image. Understanding the limitations of AI-enabled image reconstruction and enhancement is critical for safe and effective use of the technology. Hence, the purpose of this communication is to bring awareness to limitations when AI is used to reconstruct or enhance a radiological image, with the goal of enabling users to reap benefits of the technology while minimizing risks.

💡 Research Summary

**
The paper provides a comprehensive overview of the opportunities and emerging risks associated with the use of artificial intelligence (AI) for image reconstruction and enhancement in radiology. It begins by noting the rapid deployment of AI‑enabled devices—over a thousand have received FDA clearance—with many applications extending beyond interpretation (classification, detection, segmentation) to pre‑interpretation steps such as denoising, super‑resolution, and contrast manipulation. The authors highlight tangible clinical benefits: reduced MRI acquisition time, lower CT radiation dose, decreased radiotracer dose in nuclear medicine, and faster image review.

However, the authors stress that AI models rely on critical assumptions, chiefly that the data encountered in clinical practice follow the same statistical distribution as the training set. When this assumption fails—due to variations in scanner hardware, acquisition protocols, patient anatomy, or disease prevalence—AI may produce images that look sharper, smoother, or more detailed but actually contain distorted or missing diagnostic information.

Three principal approaches to image‑quality assessment are examined. (1) Task‑based assessment directly measures performance on a specific clinical task (e.g., lesion detection) using sensitivity, specificity, PPV, NPV, etc. While this provides the most clinically relevant information, it is labor‑intensive, requires expert reference standards, and may not generalize to other tasks. (2) Subjective assessment relies on radiologists rating overall image quality on a Likert scale. It is quick and intuitive but suffers from inter‑reader variability and does not guarantee diagnostic adequacy. (3) Quantitative metrics such as signal‑to‑noise ratio (SNR), structural similarity index (SSIM), root‑mean‑square error (RMSE), and mean‑square error (MSE) are automatically computed and widely used in computer‑vision research. These metrics, however, do not reflect task‑specific diagnostic performance and can be misleading when AI creates “hallucinations” or removes subtle lesions. The paper demonstrates that these three methods can disagree dramatically, using two illustrative experiments.

In a super‑resolution experiment, a neural network up‑scaled a low‑resolution image by a factor of four, but the resulting high‑resolution image contained significant structural distortions. In a fastMRI acceleration study, conventional 4‑fold undersampled reconstructions were compared with AI‑based 8‑fold reconstructions. The AI images exhibited lower RMSE and higher subjective scores, yet a simulated lesion present in the reference image disappeared in the AI reconstruction. A task‑based lesion‑detection assessment correctly identified the loss, whereas both subjective ratings and quantitative metrics failed. This example underscores the “perceived quality vs. information content” disconnect that can jeopardize patient care if clinicians trust visual appeal without recognizing hidden failures.

Regulatory considerations are explored in depth. Most AI‑based imaging devices are cleared via the FDA’s 510(k) pathway, which establishes “substantial equivalence” to a predicate device based largely on physical testing and global image‑quality descriptors. Specific clinical‑task performance data are rarely required, especially for devices with broad indications. Consequently, the pre‑market review cannot anticipate all possible failure modes, and post‑market surveillance (including adverse‑event reporting, device listing, and recall mechanisms) becomes essential. The authors argue that quantitative image‑quality metrics alone are insufficient for post‑market monitoring; task‑based assessments and periodic expert reviews should be incorporated.

From a technical standpoint, the paper notes that many AI reconstruction methods are trained to optimize a single quantitative loss (e.g., SSIM or MSE). While this yields visually pleasing images, it may suppress subtle features essential for diagnosis. Moreover, AI cannot inject patient‑specific information that is not present in the raw data; it can only impose learned priors. This limitation can lead to hallucinated structures that appear plausible but are diagnostically false.

To mitigate these risks, the authors propose several strategies: (1) incorporate task‑specific loss functions that directly penalize errors on clinically relevant outcomes; (2) train and validate models on diverse, multi‑institutional datasets to improve robustness to distribution shifts; (3) provide uncertainty or confidence maps alongside AI‑generated images so radiologists can gauge reliability; (4) adopt a hybrid evaluation framework that combines task‑based, subjective, and quantitative metrics throughout the product lifecycle; and (5) encourage regulatory agencies to require explicit clinical‑task validation for AI‑enhanced imaging devices, rather than relying solely on global image‑quality descriptors.

In conclusion, AI holds great promise for accelerating imaging workflows and improving visual quality, but it also introduces novel failure modes that can obscure or distort critical diagnostic information. A rigorous, task‑oriented assessment paradigm, continuous post‑market monitoring, and close collaboration between clinicians, engineers, and regulators are essential to harness AI’s benefits while safeguarding patient safety.