Segmentation of the Prostate Transition Zone and Peripheral Zone on MR Images with Deep Learning

Abstract

Purpose To develop a deep learning model to delineate the transition zone (TZ) and peripheral zone (PZ) of the prostate on MR images. Materials and Methods This retrospective study was composed of patients who underwent a multiparametric prostate MRI and an MRI/transrectal US fusion biopsy between January 2013 and May 2016. A board-certified abdominal radiologist manually segmented the prostate, TZ, and PZ on the entire data set. Included accessions were split into 60% training, 20% validation, and 20% test data sets for model development. Three convolutional neural networks with a U-Net architecture were trained for automatic recognition of the prostate organ, TZ, and PZ. Model performance for segmentation was assessed using Dice scores and Pearson correlation coefficients. Results A total of 242 patients were included (242 MR images; 6292 total images). Models for prostate organ segmentation, TZ segmentation, and PZ segmentation were trained and validated. Using the test data set, for prostate organ segmentation, the mean Dice score was 0.940 (interquartile range, 0.930-0.961), and the Pearson correlation coefficient for volume was 0.981 (95% CI: 0.966, 0.989). For TZ segmentation, the mean Dice score was 0.910 (interquartile range, 0.894-0.938), and the Pearson correlation coefficient for volume was 0.992 (95% CI: 0.985, 0.995). For PZ segmentation, the mean Dice score was 0.774 (interquartile range, 0.727-0.832), and the Pearson correlation coefficient for volume was 0.927 (95% CI: 0.870, 0.957). Conclusion Deep learning with an architecture composed of three U-Nets can accurately segment the prostate, TZ, and PZ. Keywords: MRI, Genital/Reproductive, Prostate, Neural Networks Supplemental material is available for this article. © RSNA, 2021.

Keywords: Genital/Reproductive; MRI; Neural Networks; Prostate.

Figure 1:

A, Axial T2-weighted section of a prostate with marked benign prostatic hyperplasia. B, The transition zone is shown in green, while, C, the peripheral zone is displayed in pink.

Figure 2:

Three convolutional neural networks were combined to segment the transition zone (TZ) and the peripheral zone (PZ). The T2-weighted prostate MR image is first passed through U-Net_A, which performs localization and creates a bounding box around the prostate that narrows the field of view. This output is then passed simultaneously to U-Net_B and U-Net_C. U-Net_B completes further localization by segmenting out the prostate organ itself, while U-Net_C classifies each pixel as either TZ or PZ. U-Net_B’s output is then used to identify the voxels in U-Net_C that belong in the prostate. After the appropriate voxels are selected in U-Net_C’s output, the final output with TZ and PZ segmentations is shown.

Figure 3:

The reliability and accuracy of the neural networks to estimate volume of the prostate, transition zone, and peripheral zone are shown with Bland-Altman plots. A–C, The dashed black line represents the average difference. The dashed blue lines show 2 standard deviations of the difference above and below the average difference. CNN = convolutional neural network, GT = ground truth.

Figure 4:

A, U-Net_A neural network architecture. U-Net_A localizes the prostate by creating a bounding box around it and narrows the field of view (see Appendix E1 [supplement]). The image is processed with nine layers that consist of convolutions (Convs), batch normalization (Batch Norm), and rectified linear unit (ReLU) activation. Both the contraction and expansion pathways use convolutional kernels that have 1 × 3 × 3 and 2 × 1 × 1 filters. The image is downsampled to a 1 × 1 × 1 matrix before it is upsampled. B, U-Net_B neural network architecture. U-Net_B completes prostate organ segmentation by classifying each pixel as either belonging to the prostate or background. The image is processed with six layers that consist of convolutions, batch normalization, and ReLU activation. The contraction and expansion pathways use convolutional kernels that have 1 × 3 × 3 and 2 × 1 × 1 filters. The image is collapsed to a 1 × 8 × 8 image before it is expanded. C, U-Net_C neural network architecture. U-Net_C differentiates between transition zone and peripheral zone by classifying every voxel in the image as one of these two classes. This classification then identifies the border between these two prostate regions. U-Net_C has the same architecture as U-Net_B and implements six layers that perform convolutions, batch normalization, and ReLU activation. The convolutional kernels use 1 × 3 × 3 and 2 × 1 × 1 filters. The image is downsampled to a 1 × 8 × 8 image before it is upsampled. 3D = three dimensional, 2D = two dimensional.

Figure 5:

Example segmentations in three patients with A, transition zone (TZ) Dice score of 0.940 and a peripheral zone (PZ) Dice score of 0.902, B, TZ Dice score of 0.910 and a PZ Dice score of 0.869, and C, TZ Dice score of 0.978 and a PZ Dice score of 0.907. CNN = convolutional neural network. Green and pink borders indicate TZ and PZ, respectively.

Figure 6:

Examples of challenging segmentations. A, Segmentation resulted in a peripheral zone (PZ) Dice score of 0.642 and a transition zone (TZ) Dice score of 0.893. B, Segmentation resulted in a PZ Dice score of 0.669 and a TZ Dice score of 0.955. Both the TZ and PZ are more challenging for the neural networks to segment at the prostate base. Green and pink borders indicate TZ and PZ, respectively. CNN = convolutional neural network.