Deep learning augmented pathological grading and intraductal carcinoma of the prostate signatures improve recurrence prediction after intensity-modulated radiation therapy.

Locally advanced prostate cancer (PCa) is associated with a high recurrence rate even after curative treatment. We aimed to develop a precise risk model by integrating the status of intraductal carcinoma of the prostate (IDC-P) and the International Society of Urological Pathology Grade Group (ISUP GG) with deep learning (DL)-based grading to predict clinical recurrence after intensity-modulated radiation therapy (IMRT). Furthermore, we identified key pathological features associated with IDC-P. We retrospectively analyzed 165 patients treated with high-dose IMRT for high- and very high-risk PCa at two institutions. Pathological features were extracted from hematoxylin and eosin-stained specimens from 100 cases using a DL-based model, from which an expert pathologist selected 10 cancer-related features. The area under the curve (AUC) values derived from split-sample validation for predicting clinical recurrence using ISUP GG and IDC-P status were 0.732 and 0.735, respectively, and 0.789 for their combination. Integrating 10 key features further improved the AUC to 0.850, with robust performance in external validation (AUC = 0.833). Kaplan-Meier analysis showed a significant difference (p < 0.05) in clinical recurrence between high- and low-risk prediction groups. Additionally, we identified four DL-derived pathological features that are significantly associated with IDC-P (p < 0.05). Incorporation of ISUP GG, IDC-P status, and DL-derived pathological features improves the prediction of clinical recurrence after high-dose IMRT in high- and very high-risk PCa. Our findings underscore the clinical significance of IDC-P and support optimized treatment strategies.