Deep Learning for Prostate Segmentation
Segmentation of the prostate in computed tomography (CT) is used for planning and guidance of prostate treatment procedures. However, due to the low soft-tissue contrast of the images, manual delineation of the prostate on CT is a time-consuming task with high inter-observer variability. In this study, we developed an automatic, three-dimensional (3D) prostate segmentation algorithm based on a customized U-Net architecture. An automatic 3D deep learning-based segmentation algorithm was designed to address the need for a fast, accurate, and repeatable 3D segmentation of the prostate on CT images, which does not depend on the intra-patient data for training. The proposed neural network has not been trained on previously acquired images from the target patient. We evaluated the performance of the algorithm to the manual references from two expert radiologists, and we used the complementary region-based (DSC, SR, and PR), surface-based (MAD), and volume based (ΔV) error metrics to measure the segmentation error of our algorithm against manual references. We also compared the results to the measured inter-expert observer difference in manual segmentation. The proposed algorithm showed robustness to some of the image artifacts caused by metallic implanted objects.
The four-level 3D U-Net FCNN architecture. The numbers above the feature maps indicate the number of feature channels and the numbers below the feature maps indicate the size of each feature channel.
Qualitative segmentation results for three, sample cases. Each row shows the results for one patient. For each image, the algorithm segmentation results are shown with yellow contours and the reference contours with blue dashed and green dotted contours on three, sample 2D slices. The bottom row shows the results from the patient in our test dataset with the lowest DSC value (62%).