🤖 Generated with AI assistance · Does not replace individualized medical advice.
AI Act Transparency Notice
This content was generated by an AI assistant to summarize and structure scientific literature for educational purposes. It does not constitute medical advice, diagnosis, or treatment. Always consult with a qualified healthcare provider for clinical decisions.
ADC Radiomics Fails to Add Significant Value Over Clinical Models in PI-RADS 3-5 Prostate Lesions
A retrospective multicenter study evaluated an apparent diffusion coefficient (ADC) radiomics model for differentiating benign and malignant PI-RADS 3-5 prostate lesions. While the model showed moderate performance, it did not significantly outperform standard clinical predictors (PSA, PSAD, PI-RADS) in external validation, limiting its immediate clinical utility.
Key Facts
| Metric | Value |
|---|---|
| Study Design | Retrospective, multicenter, external validation |
| Cohort Size | 340 patients (Training: 182, Internal Test: 71, External: 87) |
| Primary Outcome | AUC for differentiating benign vs. malignant PI-RADS 3-5 lesions |
| Best Model AUC (External) | 0.808 (Combined Clinical + Radiomics) |
| Clinical Model AUC (External) | 0.777 |
| Statistical Significance | Non-significant improvement (Holm-adjusted P = 0.392) |
Why It Matters / Context
Prostate Imaging Reporting and Data System (PI-RADS) score 3 lesions represent a diagnostic “gray zone” where the probability of clinically significant cancer is intermediate. Current guidelines often recommend targeted biopsy or active surveillance, but false positives lead to unnecessary biopsies and patient anxiety. Radiomics—extracting quantitative features from medical images—has been proposed as a tool to refine risk stratification in this ambiguous category.
This study is significant because it moves beyond single-center pilot studies to include an external validation cohort, a critical step often missing in early radiomics literature. The focus on ADC maps is clinically relevant, as diffusion-weighted imaging (DWI) is a cornerstone of modern multiparametric MRI (mpMRI) for prostate cancer detection. However, the results highlight a persistent challenge in oncology AI: the difficulty of demonstrating incremental value over robust, low-cost clinical variables like PSA density (PSAD). If a complex imaging model cannot significantly beat a simple clinical formula, its adoption into routine workflow faces substantial hurdles regarding cost, complexity, and regulatory approval.
Details
The study compared four machine-learning classifiers, ultimately selecting logistic regression for the final radiomics model due to its performance in nested cross-validation. The features were extracted specifically from ADC maps, filtered for reproducibility, and selected using training data only to prevent data leakage.
The performance of the models across the internal test and external validation cohorts is detailed below:
| Model Type | Internal Test AUC | External Validation AUC | External Specificity |
|---|---|---|---|
| Radiomics Only | 0.712 PMID: 42675219 | 0.679 PMID: 42675219 | N/A |
| Clinical Only (PSA, PSAD, PI-RADS) | 0.775 PMID: 42675219 | 0.777 PMID: 42675219 | N/A |
| Combined (Clinical + Radiomics) | 0.829 PMID: 42675219 | 0.808 PMID: 42675219 | 0.538 PMID: 42675219 |
The combined model achieved the highest AUC in both cohorts, suggesting that ADC radiomics features contain complementary information not captured by clinical variables alone. However, the statistical comparison revealed that this improvement was not significant. The Holm-adjusted P-values for the difference between the combined and clinical models were 0.123 (internal) and 0.392 (external) PMID: 42675219. Furthermore, the external specificity of the combined model was only 0.538 PMID: 42675219, indicating a high rate of false positives, which could lead to unnecessary biopsies in benign cases.
Limitations
Critical Limitation: The study is retrospective and relies on a relatively small external validation cohort (n=87) PMID: 42675219. The lack of statistical significance in the incremental value of radiomics, combined with low external specificity (0.538) PMID: 42675219, suggests that the model may not be robust enough for standalone clinical decision-making. Additionally, the study did not assess the impact of different MRI vendors or field strengths on feature reproducibility, which is a known confounder in radiomics.
Clinical Takeaway
Clinicians should continue to rely on established clinical predictors (PSA, PSAD, PI-RADS) for risk stratification of PI-RADS 3-5 lesions, as current ADC radiomics models do not offer a statistically significant improvement in diagnostic accuracy sufficient to justify routine implementation.
Full Reference
Du S, Du S, Mei S, Shen Z. Apparent diffusion coefficient radiomics for differentiating benign and malignant PI-RADS 3-5 prostate lesions: external validation. International Urology and Nephrology. 2026. PMID: 42675219 [DOI: 10.1007/s11255-026-05366-z]