[Paper Review] A Cascaded Residual UNET for Fully Automated Segmentation of Prostate and Peripheral Zone in T2-weighted 3D Fast Spin Echo Images

Lavanya Umapathy, Wyatt Unger+5 more|arXiv (Cornell University)|Dec 25, 2020

Advanced Neural Network Applications8 references4 citations

TL;DR

This paper proposes a cascaded residual U-Net (Cascaded MRes-UNET) for fully automated segmentation of the prostate gland and peripheral zone in T2-weighted 3D fast spin echo MRI scans. The architecture uses residual blocks and a two-stage cascaded design to improve boundary localization, achieving high Dice scores of 0.91±0.02 and sub-5% error in total prostate volume estimation compared to radiologist annotations.

ABSTRACT

Multi-parametric MR images have been shown to be effective in the non-invasive diagnosis of prostate cancer. Automated segmentation of the prostate eliminates the need for manual annotation by a radiologist which is time consuming. This improves efficiency in the extraction of imaging features for the characterization of prostate tissues. In this work, we propose a fully automated cascaded deep learning architecture with residual blocks, Cascaded MRes-UNET, for segmentation of the prostate gland and the peripheral zone in one pass through the network. The network yields high Dice scores ($0.91\pm.02$), precision ($0.91\pm.04$), and recall scores ($0.92\pm.03$) in prostate segmentation compared to manual annotations by an experienced radiologist. The average difference in total prostate volume estimation is less than 5%.

Motivation & Objective

• To eliminate time-consuming manual segmentation of the prostate and peripheral zone in T2-weighted 3D MRI by radiologists.
• To develop a fully automated deep learning framework that accurately segments both the prostate gland and its peripheral zone in a single inference pass.
• To improve segmentation accuracy, especially at tissue boundaries, using residual learning and a cascaded network architecture.
• To achieve high consistency in prostate volume estimation, minimizing deviation from manual annotations.

Proposed method

• A cascaded U-Net architecture is employed, where the first stage predicts the prostate gland and the second stage refines the segmentation of the peripheral zone.
• Residual blocks are integrated into the encoder and decoder pathways to stabilize training and improve gradient flow.
• The network is trained in an end-to-end fashion using binary cross-entropy loss with focal weighting to address class imbalance.
• The input is 3D T2-weighted fast spin echo MRI data, and the model predicts voxel-wise segmentation masks for the prostate and peripheral zone.
• A two-stage inference process allows the second stage to focus on fine-grained peripheral zone boundaries using features from the first stage.
• Data augmentation techniques including rotation, flipping, and intensity jittering are applied to improve generalization.

Experimental results

Research questions

• RQ1 Can a cascaded deep learning architecture achieve high-precision segmentation of both the prostate gland and peripheral zone in T2-weighted 3D MRI?
• RQ2 Does the integration of residual blocks improve segmentation performance and training stability in prostate MRI segmentation?
• RQ3 To what extent does the cascaded design enhance boundary localization compared to a single-stage U-Net?
• RQ4 How accurate is the automated volume estimation compared to manual radiologist measurements?

Key findings

• The proposed Cascaded MRes-UNET achieved a Dice score of 0.91±0.02 for prostate segmentation, indicating strong overlap with manual annotations.
• Precision and recall for prostate segmentation were 0.91±0.04 and 0.92±0.03, respectively, demonstrating high segmentation reliability.
• The average difference in total prostate volume estimation between automated and manual segmentation was less than 5%.
• The cascaded architecture improved boundary delineation, particularly for the peripheral zone, compared to baseline U-Net models.
• The use of residual blocks contributed to stable training and better feature learning in deep layers.