USE-Net

Incorporating Squeeze-and-Excitation blocks into U-Net for multi-institutional prostate zonal segmentation.

USE-Net is a novel Deep Learning architecture that enhances the classic U-Net by incorporating Squeeze-and-Excitation (SE) blocks to tackle the challenges of multi-institutional medical imaging.

The project focuses on the precise zonal segmentation of the prostate—differentiating between the Central Gland (CG) and the Peripheral Zone (PZ). By using adaptive feature recalibration, USE-Net effectively handles the heterogeneous characteristics of MRI datasets collected across different clinical institutions, achieving superior generalization.

Feature Recalibration Multi-Institutional

Beyond Whole Gland Segmentation

While many tools focus on the whole prostate, USE-Net addresses the zonal compartment system (Central Gland vs. Peripheral Zone). This is clinically vital because:

  • PCa Localization: Roughly 70% of prostate cancers originate in the Peripheral Zone.
  • Diagnostic Refinement: Zonal volume ratios are key indicators for monitoring hyperplasia and refining biopsies.
  • Heterogeneity: MRI appearance varies wildly between scanners and institutions; USE-Net uses channel-wise relationships to normalize these differences.

Key Architectures

We proposed two main variants of the network:

  • Enc USE-Net: SE blocks are integrated only within the Encoder path. This variant shows excellent overall generalization across various training conditions.
  • Enc-Dec USE-Net: SE blocks are integrated into both Encoder and Decoder blocks. This model remarkably outperforms state-of-the-art methods when trained on multi-institutional datasets simultaneously.

Research Highlights

Adaptive Mechanisms

Leverages Squeeze-and-Excitation to recalibrate feature maps, focusing on the most relevant anatomical signals.

Cross-Dataset Generalization

The first study to evaluate prostate zonal segmentation across multiple institutional datasets (T2w MRI).

Benchmarked Performance

Outperforms U-Net, pix2pix, and Mixed-Scale Dense Networks in complex multi-dataset scenarios.

Request Information

Core Publication

2019

  1. USE-Net: Incorporating Squeeze-and-Excitation blocks into U-Net for prostate zonal segmentation of multi-institutional MRI datasets
    Leonardo Rundo, Changhee Han, Yudai Nagano, and 8 more authors
    Neurocomputing, 2019
    DOI