Introduction:

Recent years have seen a dramatic increase in magnetic resonance imaging (MRI)-guided prostate biopsy utilization. Accurate biopsy requires accurate prostate segmentation on MRI. Providing precise prostate segmentations on MRI is a tedious and time-consuming task. Recent advancements in deep learning have enabled deep neural networks to rapidly perform medical imaging analysis tasks. Despite the potential advantages of utilizing deep learning methods, challenges persist in incorporating these models into clinical care. Our objective was to develop a deep learning model to rapidly and accurately segment the prostate on MRI and implement it in the routine clinical workflow in a proof of principle fashion.

Methods:

We utilized a multi-institutional cohort of 905 patients who underwent MRI at 28 academic and private practice institutions between 2013 and 2019, followed by MRI-US fusion biopsy at Stanford University. We trained a deep learning model, ProGNet, to automatically segment the prostate on 805 T2WI MRI sequences and tested it on an independent 100-case test set. A urologic oncology expert segmented all cases. We explored whether our model would improve performance over radiology technicians by comparing segmentation overlap with the urologic oncology expert using the Dice similarity coefficient (DSC).. We compared our ProGNet model’s performance to two state-of-art deep learning networks, the U-Net and the HED models. DSCs were compared using paired t-tests. Finally, we worked with the biopsy vendor (Eigen, Grass Valley, CA) to enable model outputs to be usable on the clinical biopsy system and then utilized the model on a prospective 11 patient-cohort of men undergoing biopsy.

Results:

ProGNet (mean DSC = 0.92±0.02) outperformed two state-of-the-art deep learning networks, the U-Net and HED models (mean DSC = 0.85±0.07, p<0.0001 and 0.80±0.08, p<0.0001, respectively) in the expert-annotated 100-case test set. Furthermore, ProGNet exceeded the segmentation performance of experienced radiology technicians (mean DSC = 0.89±0.05, p<0.0001). The ProGNet model also performed more consistently, with fewer poorly performing cases than radiology technicians. Initial training of ProGNet—which is only done once—took 20 hours, but after training, it only took 50 seconds to segment the entire test set. Conversely, it took radiology technicians almost 17 hours to segment the 100 cases, meaning the ProGNet model would have saved >16 hours of segmentation time in the test set alone. We successfully utilized the ProGNet deep learning model to segment the prostate on MRI on 11 consecutive prospective patients undergoing biopsy at Stanford University.

Conclusion:

Despite the enormous potential of deep learning to perform image analysis tasks, there has been almost no clinical implementation to date. Deep learning has never been used clinically for the important and time-consuming prostate segmentation task. We have developed a deep learning model to segment the prostate gland on T2WI MRI and proved that it outperformed state-of-art deep learning networks as well as trained radiology technicians. The model saved >16 hours in segmentation time in a 100-patient test set alone. Most importantly, we successfully integrated it with biopsy software to allow clinical use in a proof of principle fashion.

Funding: N/A