U-Net benign prostatic hyperplasia-trained deep learning model for prostate ultrasound image segmentation in prostate cancer.

Background: Prostate cancer (PCa) remains a leading cause of male morbidity and mortality globally, where transrectal ultrasound (TRUS) serves as the cornerstone imaging modality for diagnosis and therapeutic guidance. However, automated segmentation of prostatic anatomy in TRUS is persistently hindered by technical constraints. This study aimed to develop prostate segmentation model trained by deep learning from ultrasound images of patients with benign prostatic hyperplasia (BPH) and to verify if the developed model can be applied for the ultrasound image segmentation of patients with PCa. Methods: A total of 370 and 68 prostate ultrasound images were collected from 260 BPH patients and 62 PCa patients, respectively. U-Net, LinkNet and PSPNet neural network were used to train segmentation model. The Dice coefficients of the model segmentation for test set comprising BPH and PCa images were calculated. Two independent-sample Results: The study demonstrated significant radiomic differences between PCa and BPH in ultrasound imaging, with least absolute shrinkage and selection operator (LASSO) regression identifying 9 discriminative features including shape and texture parameters. The U-Net model achieved superior segmentation performance with a peak Intersection over Union (IoU) of 0.9602 and maintained robustness across resolutions. The independent-sample Conclusions: We proved that prostatic segmentation model trained on ultrasound images of BPH could be applied in PCa.