Deep Learning with Quantitative Features of Magnetic Resonance Images to Predict Biochemical Recurrence of Radical Prostatectomy: A Multi-Center Study

Abstract

Biochemical recurrence (BCR) occurs in up to 27% of patients after radical prostatectomy (RP) and often compromises oncologic survival. To determine whether imaging signatures on clinical prostate magnetic resonance imaging (MRI) could noninvasively characterize biochemical recurrence and optimize treatment. We retrospectively enrolled 485 patients underwent RP from 2010 to 2017 in three institutions. Quantitative and interpretable features were extracted from T2 delineated tumors. Deep learning-based survival analysis was then applied to develop the deep-radiomic signature (DRS-BCR). The model's performance was further evaluated, in comparison with conventional clinical models. The model achieved C-index of 0.802 in both primary and validating cohorts, outweighed the CAPRA-S score (0.677), NCCN model (0.586) and Gleason grade group systems (0.583). With application analysis, DRS-BCR model can significantly reduce false-positive predictions, so that nearly one-third of patients could benefit from the model by avoiding overtreatments. The deep learning-based survival analysis assisted quantitative image features from MRI performed well in prediction for BCR and has significant potential in optimizing systemic neoadjuvant or adjuvant therapies for prostate cancer patients.

Keywords: MRI; biochemical recurrence; deep learning; prostate cancer; survival prediction.

Figure 1

The workflow of radiomics-based model for data preparation, modeling, signature construction, predicting BCR-free survival, and clinical applications. DSNN: deep survival neural network; BCR: biochemical recurrence; DRS: deep radiomic signature; CAPRA: Cancer of Prostate Risk Assessment; NCCN: National Comprehensive Cancer Network; GG: Gleason grade group system.

Figure 2

Performance of DRS-BCR for predicting 3-year BCR-free survival. (a–c) The Kaplan–Meier (K–M) curves of DRS-BCR-free survival for BCR-free survival within three years in the primary cohort, validation cohort 1, and validation cohort 2, respectively. The p values were calculated by log-rank test between subgroup with high-risk and low-risk, and significant discrimination was revealed by p values less than 0.05. (d–f) The calibration curves for 3-year BCR. (g–i) Clinical benefits by decision curve analysis for 3-year BCR. DRS: deep radiomic signature; BCR: biochemical recurrence.

Figure 3

The time-dependence receiver operating characteristic curve (ROC) of DRS-BCR in the primary cohort, validation cohort 1, and validation cohort 2. DRS: deep radiomic signature; AUC: area under the curve.