SRS257 - Automated segmentation of prostate anatomy for the efficient provision of 3D models during robotic-assisted radical prostatectomy

Purpose Evidence suggests that PSMs and functional outcomes of patients undergoing Robotic-assisted radical prostatectomy (RARP) can be improved when surgeons are presented with 3D prostate models during surgery. However, the creation of such models is time-consuming and inefficient for clinical use. We aim to provide an automated segmentation model of prostate anatomy to facilitate the creation of 3D models. Methods 108 cases were prospectively identified at Guy’s Hospital, London (REC: 23/LO/0492). Automated segmentation of the prostate gland was done with AutoProstate in MONAILabels on T2w images. An expert radiologist manually delineated prostate cancer lesions, external sphincter, urethra and neurovascular bundles (NVBs) on mp-MRI sequences. We trained an nnU-Net model for image segmentation. Cases were split into training (N = 84), validation (N = 20) and hold-out test (N = 4) following 5-fold cross-validation. 3D surface models were created in Python for visualisation and printing. Results Mean Dice scores for segmentation model predictions after cross validation for the prostate (94.8%), lesions (78.8%), external sphincter (83%), urethra (71.2%) and NVBs (69.9%). We then successfully generated 3D models that were loaded into Innersight Labs and 3D printed. Conclusions We present an automated segmentation model of patient-specific prostate anatomy to automate 3D models creation. We demonstrated that predictions can efficiently be used for the creation and deployment of 3D models. Future work will use mp-MRI as input to improve accuracy of lesions, pre-train a model via a self-supervised task, and introduce an efficient human-in-the-loop mechanism for the automated correction of predictions.