Abstract / Summary
Abstract Objective: To develop a multimodal diagnostic model integrating radiomics features, deep learning features derived from prostate multiparametric magnetic resonance imaging (mpMRI), and prostate-specific antigen (PSA) levels, and to investigate its value in differentiating benign and malignant prostate nodules, thereby providing an auxiliary tool for the noninvasive diagnosis of prostate cancer and clinical decision-making. Methods: A total of 312 patients with prostate nodules confirmed by pathology were retrospectively enrolled and randomly divided into training and test sets at an 8:2 ratio. Radiomics and deep learning features were extracted from T2-weighted imaging (T2WI), diffusion-weighted imaging (DWI), and apparent diffusion coefficient (ADC) maps, and were combined with PSA levels to construct multimodal diagnostic models. Five models were developed, including the ADC model, DWI model, T2WI model, mpMRI model, and mpMRI_PSA model. Five-fold cross-validation was used for model training. Model performance was evaluated in the test set using the area under the receiver operating characteristic (ROC) curve (AUC), accuracy, sensitivity, specificity, positive predictive value, negative predictive value, and F1 score. The DeLong test was used to compare the differences in AUC between the best-performing model and the other models. A P value < 0.05 was considered statistically significant. Results: All five models demonstrated a certain ability to differentiate benign from malignant prostate nodules. The AUCs of the ADC, DWI, T2WI, mpMRI, and mpMRI_PSA models were 0.73, 0.79, 0.84, 0.94, and 0.97, respectively. The diagnostic performance progressively improved with the integration of multiple MRI sequences compared with individual sequences. Further incorporation of PSA resulted in the best diagnostic performance for the mpMRI_PSA model, which outperformed the other models in terms of accuracy, sensitivity, specificity, and F1 score. The DeLong test demonstrated that the differences in AUC between the best-performing mpMRI_PSA model and the ADC, DWI, T2WI, and mpMRI models were all statistically significant ( P < 0.05). Conclusion: The multimodal diagnostic model integrating radiomics and deep learning features from prostate mpMRI with PSA levels demonstrated high discriminative performance for differentiating benign and malignant prostate nodules. Its diagnostic performance was superior to that of the single-sequence models and the mpMRI-only model, suggesting that it may serve as a potential auxiliary tool for the noninvasive differential diagnosis of prostate cancer and clinical decision-making.