Radiomics is the science of extracting thousands of quantitative features from routine medical images - features invisible to the human eye - and using them to build prediction models. In endometrial cancer, radiomics could one day tell doctors preoperatively what a tumor looks like at the molecular and histological level, potentially replacing the need for some invasive tests.
Despite rapid growth in the number of radiomics studies, the field has struggled to move from research into actual clinical use. The main reason is a lack of standardization: different labs use different methods, software, and reporting practices, making it nearly impossible to compare or reproduce results.
To address this, two quality scoring tools have been developed. The Radiomics Quality Score (RQS), proposed in 2017, has 16 items covering methodological rigor. The newer METRICS (METhodological RadiomICs Score), developed by the European Society of Medical Imaging Informatics, has 30 items across nine categories and is designed to be more flexible and comprehensive. This study applied both tools to all available endometrial cancer radiomics research to take stock of the field's quality.
Researchers searched PubMed and Scopus for all endometrial cancer radiomics studies published through October 2023. After removing duplicates, non-original articles, and non-English papers, 68 studies were included. These covered CT, MRI, PET/CT, and ultrasound imaging.
Four reviewers assessed each study using both the RQS and METRICS tools - two were novices in radiomics and two were experienced experts. To ensure everyone was calibrated, 30 randomly selected articles were scored independently by all four reviewers before comparing results. Remaining articles were then divided between the pairs.
The agreement between reviewers was measured using the intraclass correlation coefficient (ICC), where values above 0.90 indicate excellent reliability. Subgroup analyses tested whether scores varied based on imaging modality, study topic, year of publication, or journal quality ranking.
The median RQS was 11 out of a maximum 36, equivalent to just 30.6%. This is widely considered unsatisfactory. Only two studies reached the top score of 19 (52.8%). Under the RQS framework, the field appears to have significant methodological shortcomings.
The median METRICS score was 67.6%, which the authors describe as reflecting generally good overall quality. Two studies achieved near-maximum METRICS scores of 90.7%. The METRICS tool tends to score studies more favorably because it weights items more evenly and accommodates retrospective and exploratory research designs without the heavy penalties that RQS imposes.
Importantly, the two tools are not directly comparable - they measure overlapping but distinct aspects of quality. RQS emphasizes study design features like prospective enrollment and external validation. METRICS emphasizes documentation, transparency, and practical applicability. Neither tool is definitively better; they highlight different weaknesses.
Under RQS evaluation, the most striking findings were near-universal absences: no study used a prospective design, no study conducted phantom testing for imaging reproducibility, and no study included a cost-effectiveness analysis. These three items alone account for a large portion of the RQS penalty.
Only 17.6% of studies detected and discussed biological correlates - meaning most studies predicted clinical outcomes from imaging features without linking those features to any known tumor biology. In endometrial cancer, this gap is particularly important given that molecular features now drive treatment decisions under the updated FIGO staging system.
External validation - testing a model on a completely separate dataset - was performed in only about 30% of studies combined (external only plus both internal and external). Without external validation, there is no way to know whether a radiomics model will work in a different hospital, scanner, or patient population. Only 11.8% of studies made their code or data publicly available, limiting reproducibility.
The most common research goal was tumor characterization - predicting histological features like myometrial invasion depth, lymphovascular space invasion, tumor grade, or molecular subtype. This accounted for 35.3% of all studies. The clinical motivation is clear: if imaging alone could reliably predict these features, surgery could be better planned without waiting for post-operative pathology results.
Recurrence risk stratification was the second most common goal at 19.1%, followed by cancer classification (distinguishing malignant from benign lesions) at 14.7%, lymph node metastasis prediction at 13.2%, prognosis prediction at 11.8%, and segmentation or treatment planning at 5.9%.
MRI was dominant, used in 82.3% of studies. CT and PET/CT each accounted for 7.4%, and ultrasound for 2.9%. Machine learning methods were used in 88.2% of studies, while deep learning was employed in only 11.8% - reflecting its relatively recent emergence in the field. The volume of published work grew rapidly, with nearly 30% of all studies published in 2023 alone.
Both scoring tools showed excellent inter-rater reliability. The ICC for METRICS among all four reviewers was 0.959, and for RQS it was 0.897 - both above the 0.90 threshold for excellent agreement. This means the tools produce consistent results regardless of whether the reviewer is a novice or an expert, which is important for their adoption as quality benchmarks.
Expert readers and novice readers produced equivalent scores on both tools (confirmed by statistical equivalence testing), suggesting that after a brief training session, even researchers without deep radiomics expertise can apply these tools reliably.
No statistically significant differences in quality scores were found when comparing studies by imaging modality, year of publication, research topic, or journal quartile. This was somewhat surprising - one might expect newer studies or those in higher-ranked journals to show better quality, but the data did not support this pattern.
The apparent contradiction - poor quality by RQS, acceptable quality by METRICS - reflects fundamental differences in how the tools were designed. RQS weights heavily toward a few key items, particularly prospective design (+7 points) and external validation. Because virtually no endometrial cancer radiomics study is prospective, the RQS systematically penalizes the entire field.
METRICS distributes weight more evenly across 30 items and uses conditional scoring that adapts to the study design. A retrospective pilot study can still score well on METRICS if it clearly documents its methods, appropriately handles feature selection, and uses valid performance metrics. This makes METRICS more useful as a practical guide for improvement rather than as a blunt quality judgment.
The authors suggest that METRICS may also serve as a study design checklist - researchers planning new radiomics studies can use METRICS as a step-by-step guide to ensure their work meets methodological standards, increasing the likelihood that their model will eventually translate into clinical practice.
The review highlights that endometrial cancer radiomics research is methodologically active but not yet ready for routine clinical use. The most important improvements needed are external validation in independent datasets, discussion of how imaging features relate to known biology, and eventual prospective study designs.
The updated FIGO staging system for endometrial cancer now incorporates molecular features that can only be determined through biopsy or surgery. Radiomics that can predict these features preoperatively - from a routine MRI scan - would be genuinely transformative. Achieving this requires not just better models but better-validated models with biological grounding.
The authors conclude that routinely applying METRICS and RQS - both as evaluation tools and as design guides - can accelerate the maturation of the field. Journals, grant agencies, and clinical implementers should require compliance with these standards before radiomics tools are considered for clinical adoption.