RESUMEN
We aimed to apply a potent deep learning network, NAFNet, to predict adverse pathology events and biochemical recurrence-free survival (bRFS) based on pre-treatment MRI imaging. 514 prostate cancer patients from six tertiary hospitals throughout China from 2017 and 2021 were included. A total of 367 patients from Fudan University Shanghai Cancer Center with whole-mount histopathology of radical prostatectomy specimens were assigned to the internal set, and cancer lesions were delineated with whole-mount pathology as the reference. The external test set included 147 patients with BCR data from five other institutes. The prediction model (NAFNet-classifier) and integrated nomogram (DL-nomogram) were constructed based on NAFNet. We then compared DL-nomogram with radiology score (PI-RADS), and clinical score (Cancer of the Prostate Risk Assessment score (CAPRA)). After training and validation in the internal set, ROC curves in the external test set showed that NAFNet-classifier alone outperformed ResNet50 in predicting adverse pathology. The DL-nomogram, including the NAFNet-classifier, clinical T stage and biopsy results, showed the highest AUC (0.915, 95% CI: 0.871-0.959) and accuracy (0.850) compared with the PI-RADS and CAPRA scores. Additionally, the DL-nomogram outperformed the CAPRA score with a higher C-index (0.732, P < 0.001) in predicting bRFS. Based on this newly-developed deep learning network, NAFNet, our DL-nomogram could accurately predict adverse pathology and poor prognosis, providing a potential AI tools in medical imaging risk stratification.
RESUMEN
We aimed to study radiomics approach based on biparametric magnetic resonance imaging (MRI) for determining significant residual cancer after androgen deprivation therapy (ADT). Ninety-two post-ADT prostate cancer patients underwent MRI before prostatectomy (62 with significant residual disease and 30 with complete response or minimum residual disease [CR/MRD]). Totally, 100 significant residual, 52 CR/MRD lesions, and 70 benign tissues were selected according to pathology. First, 381 radiomics features were extracted from T2-weighted imaging, diffusion-weighted imaging, and apparent diffusion coefficient (ADC) maps. Optimal features were selected using a support vector machine with a recursive feature elimination algorithm (SVM-RFE). Then, ADC values of significant residual, CR/MRD lesions, and benign tissues were compared by one-way analysis of variance. Logistic regression was used to construct models with SVM features to differentiate between each pair of tissues. Third, the efficiencies of ADC value and radiomics models for differentiating the three tissues were assessed by area under receiver operating characteristic curve (AUC). The ADC value (mean ± standard deviation [s.d.]) of significant residual lesions ([1.10 ± 0.02] × 10-3 mm2 s-1) was significantly lower than that of CR/MRD ([1.17 ± 0.02] × 10-3 mm2 s-1), which was significantly lower than that of benign tissues ([1.30 ± 0.02] × 10-3 mm2 s-1; both P < 0.05). The SVM feature models were comparable to ADC value in distinguishing CR/MRD from benign tissue (AUC: 0.766 vs 0.792) and distinguishing residual from benign tissue (AUC: 0.825 vs 0.835) (both P > 0.05), but superior to ADC value in differentiating significant residual from CR/MRD (AUC: 0.748 vs 0.558; P = 0.041). Radiomics approach with biparametric MRI could promote the detection of significant residual prostate cancer after ADT.
