Image Quality in the Era of Artificial Intelligence

TL;DR

AI-based image reconstruction and enhancement in radiology promises faster acquisitions and sharper images, but may produce artifacts or hallucinations and fail to add patient-specific information. The paper surveys task-based, subjective, and quantitative image-quality assessments and discusses their tradeoffs and sometimes conflicting results, including metrics such as $SSIM$, $RMSE$, and $SNR$. It outlines FDA's risk-based regulation, 510(k) substantial equivalence, and indications-for-use considerations, noting that general indications may not guarantee applicability to all clinical tasks and that postmarket surveillance is important. It highlights two key failure modes: disconnect between perceived image quality and diagnostic content, and AI-induced artifacts that mimic pathology, with real-world cases showing reduced diagnostic performance despite appealing visuals. The authors call for mindful deployment, task-specific validation, and development of AI-diagnostic safeguards to minimize patient risk.

Abstract

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.

Figures (3)

• Figure 1: Overview of image quality assessment methods. Image quality has typically been measured in one of three ways: (1) Task-based assessments, (2) subjective assessments, and (3) quantitative image metrics. Task-based, subjective, and quantitative assessments of image quality do not always agree.
• Figure 2: Example of a super-resolution task. The high-resolution image in the center was downsized by a factor of 4 and then upsized by a conventional bicubic interpolation method and by EnhanceNet, a neural network based approach. With the bicubic interpolation (far right), the image is clearly low resolution. The neural network based approach has a much higher apparent resolution, but there are significant distortions in the stripes as well as texture changes in the grass.
• Figure 3: Example of the disconnect between visual image quality and diagnostic utility. The ground truth is a brain image from the fastMRI dataset with an added lesion highlighted by the yellow box. This image was reconstructed for two different acceleration factors by a conventional method and an AI trained on the fastMRI dataset. The reconstructed images from the AI have better quantitative metrics and appearance, but no additional diagnostic information. That is, the use of AI-based reconstruction does not change the presence (or absence) of the lesion. It is readily apparent that the 8-fold accelerated conventional reconstruction is not of diagnostic quality, and the AI-reconstructed image with the same acceleration appears to be diagnostic, despite having lost critical information (the lesion).