To evaluate a machine learning radiomics model integrating PET/CT and MRI for risk stratification in patients with suspected prostate cancer.
Key Findings:
Multimodal models outperformed single-modality approaches for clinically significant prostate cancer, indicating a potential for improved diagnostic accuracy.
In the internal test cohort, logistic regression, support vector machine, and LightGBM achieved AUCs of 0.91, suggesting strong model performance.
In external validation, LightGBM achieved AUCs of 0.82 and 0.89, while logistic regression achieved AUCs of 0.80 and 0.85, highlighting variability in model performance across cohorts.
Interpretation:
The model shows promise for risk stratification but its clinical role remains uncertain due to modest negative predictive values, which may limit its utility in clinical practice.
Limitations:
Retrospective design and small external validation cohorts.
Possible selection bias and variability in imaging protocols across centers.
Lack of prospective validation and generalizability limited to Chinese academic centers.
Ambiguity in automated segmentation analysis regarding feature selection thresholds.
Conclusion:
Further study of multimodal radiomics is warranted as a decision-support tool for prostate cancer risk stratification, not as a replacement for histopathologic confirmation, and should include direct comparisons with established methods.
