Item 4 — Study aims, objectives, and hypotheses

Define clearly the clinical or scientific question to be answered; avoid vague statements or descriptions of a process. Limit the chance of post hoc data dredging by specifying the study’s hypothesis a priori. The study’s hypothesis and objectives should guide appropriate statistical analyses, sample size calculations, and whether the hypothesis will be supported or not.

Explanation and Elaboration: The effectiveness and usefulness of an AI should be evaluated in the context of the specific problem it is intended to solve. In the absence of clear statement of the problem, it is difficult to determine the importance and impact of the work. The objectives and hypotheses (if not a data-driven approach) provide context for the reader to evaluate whether the methods and tests employed in the investigation are appropriate for achieving the stated aims of the work.

Examples

The current study is motivated to investigate this critical screening task because clinicians often wonder whether existing public datasets, such as NLST, are sufficient to train effective AI models or if newer datasets, such as the NELSON study, are necessary to be pixel-annotated to further improve performance. Similarly, AI developers need insights into whether collecting more NLST-like data will significantly enhance their models or if diminishing returns set in. Although larger datasets and deeper neural networks generally lead to improved performance, the prediction of nodule malignancy risk may encounter inherent limitations. These limitations may arise from the nature of the morphologic biomarkers themselves, suggesting that there could be a maximum level of predictive accuracy that cannot be surpassed, regardless of the volume of data or the complexity of the model. This article investigates the relationship between the volume of training data and the performance of an AI model designed for pulmonary nodule malignancy risk assessment. Specifically, we investigated whether increasing the size of the AI training dataset leads to continuous performance improvements on the external dataset, or if the performance eventually plateaus, indicating a limit to the benefits of adding similar data. [Introduction (12), © RSNA, 2025]

However, CT-derived morphologic measurements can be difficult to interpret. Reference charts have long underpinned clinical decision-making; pediatric growth charts, for example, offer indispensable benchmarks for monitoring development ... In neuroimaging, analogous MRI-based approaches have emerged …, such as the creation of brain charts … that can standardize age-related structural changes and support early detection of neuropsychiatric conditions. However, equivalent reference standards for the rest of the body remain largely absent, although the standardized assessment of anatomy in CT imaging lends itself well to anatomic benchmarking. Large-scale efforts such as the UK Biobank have built extensive MRI repositories …, but no comparable CT-based cohorts exist. Hospital picture archiving and communication systems (PACS) represent a rich yet underutilized data repository that holds substantial promise for quantifying anatomic variability. Using these data could enable the establishment of comprehensive, data-driven reference models of human anatomy based on CT imaging. In this study, we generate comprehensive body charts for 104 anatomic structures across adulthood from more than 200 000 CT scans and over 7.8 million segmented anatomic structures. Our primary aim is to characterize the nonlinear, age-dependent trajectories of organ volumes and to establish reference models that convert raw measurements into age- and sex-standardized centile scores. To achieve this, we identify optimal parametric forms of fractional polynomials (FPs) that capture complex aging patterns in both median and variance, going beyond traditional linear models. [Introduction (13), © RSNA, 2026]