RATIONALE AND OBJECTIVES

To develop an automatic deep-radiomics framework that diagnoses and stratifies prostate cancer in patients with prostate-specific antigen (PSA) levels between 4 and 10 ng/mL.

MATERIALS AND METHODS

A total of 1124 patients with histological results and PSA levels between 4 and 10 ng/mL were enrolled from one public dataset and two local institutions. An nnUNet was trained for prostate masks, and a feature extraction module identified suspicious lesion masks. Radiomics features were extracted from the biparametric magnetic resonance imaging using these masks. Machine learning models were developed to diagnose prostate cancer (PCa), clinically significant PCa (csPCa), and high-risk csPCa based on radiomics and clinical features. The models were evaluated in both internal and external cohorts. The best model was further compared with PSA density (PSAD), free to total PSA (F/T PSA), and Prostate Imaging Reporting and Data System (PI-RADS) scores in the external cohort.

RESULTS

The models based on both radiomics and clinical features outperformed those based on radiomics or clinical features alone. The top-performing models achieved areas under the curve of 0.80, 0.88, and 0.83 on internal testing, and 0.79, 0.80, and 0.82 on external testing for diagnosing PCa, csPCa, and high-risk csPCa. Our deep-radiomics model surpassed PSAD, F/T PSA, and PI-RADS scores in an external cohort. Decision curve analysis indicated that our model offers greater net benefit than these methods.

CONCLUSION

The deep-radiomics model automatically segments prostate and suspicious lesions, diagnoses, and stages of PCa in patients with PSA levels between 4 and 10 ng/mL. Our method addresses the shortcomings of manual segmentation and inconsistency, delivering outstanding performance. It provides multilevel predictions to assist clinical decision-making and benefit patients with gray zone PSA.