Nuclear Medicine

Most commercial radiology AI validations lack per‑subgroup performance reporting.

This scoping review found 392 of 545 validation studies reported demographic subgroup data, but only 77 presented subgroup performance results. Reporting of sex, age, and race/ethnicity performance was fragmented across modalities and targets, with skeletal and lung studies more likely to report results. No convincing temporal improvement or effect of company sponsorship on subgroup reporting was observed (OR 1.039 and OR 1.010 respectively). The authors conclude current validation literature is inadequate to quantify algorithmic bias and calls for standardized, thorough subgroup reporting. 1

Radiomics-based ML discriminates bladder cancer risk but suffers high bias and low certainty.

This meta-analysis of 57 studies (n=11,933) reports pooled validation AUROCs around 0.893 for muscle invasion and up to 0.921 when MRI and clinical features were combined. Performance varied by imaging modality, with MRI+clinical models showing the highest discrimination in some analyses. Risk of bias, methodological shortcomings, and low GRADE certainty limit clinical translation of these models. The data support further rigorous, prospective model development and external validation before clinical adoption. 2

Endocrine‑organ radiomics adds little beyond clinical variables in PSMA‑negative prostate cancer.

In 101 men with biochemically recurrent PSMA‑negative prostate cancer, multimodal radiomics plus clinical variables reached AUC 0.758 versus 0.727 for the best clinical model. Imaging‑only radiomics performed worse than models including clinical variables, and DeLong testing showed no statistically significant incremental value. Authors suggest endocrine organ radiomics may reflect tumor‑host biology but its added clinical value is modest and unproven. Larger, prospective studies are required before using such radiomics to change management in this population. 3