This paper warns that progress in medical imaging driven by artificial intelligence (AI) has focused heavily on making algorithms better, while paying less attention to the basic ideas that define the problems and measures of success. The authors draw a clear distinction between algorithmic innovation — improving computational methods within a fixed problem — and conceptual innovation, which changes what question is asked, how success is measured, or why a result matters clinically.

As a Perspective article, the authors do not report a new experiment. Instead they offer an operational vocabulary and criteria to help the field recognize conceptual contributions. They list three main contributions: defining the distinction between conceptual and algorithmic innovation; proposing practical indicators of conceptual work (summarized in a boxed list in the paper); and examining systemic factors that make conceptual work harder to pursue, especially for early-career researchers.

At a high level, conceptual innovation acts upstream by shaping task definitions, assumptions about ground truth, and evaluation criteria. Algorithmic innovation acts downstream by improving models, architectures, or training under that fixed framing. The paper gives concrete examples. For image reconstruction, conventional pixel-based metrics such as peak signal-to-noise ratio (PSNR) and structural similarity index (SSIM) are imperfect proxies for diagnostic usefulness. Highly expressive deep learning reconstructions can score well on those metrics while producing images with structured distortions, suppressing diagnostically relevant features, or behaving unstably under modest changes in sampling or anatomy. These problems reveal the need to rethink what should be measured and why.

The argument matters because conceptual choices steer what counts as progress. If the field rewards only incremental algorithmic gains on narrow benchmarks, research can be well quantified but weakly connected to clinical decisions. The authors note that current incentive structures, training paths, and publication norms often favor algorithmic novelty. They conclude with recommendations for researchers, mentors, reviewers, and journals to better recognize and support conceptual work alongside algorithmic advances.