MS-Net: Multi-Site Network for Improving Prostate Segmentation With Heterogeneous MRI Data.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer-Aided Diagnosis

Background:

• Automated prostate segmentation in MRI is crucial for computer-assisted diagnosis.
• Deep learning methods show promise but require large datasets, which are often scarce in medical imaging.
• Aggregating multi-site data is essential for robust model training but faces challenges due to data heterogeneity.

Purpose of the Study:

• To develop a novel multi-site network (MS-Net) for improved prostate segmentation.
• To address the challenge of inter-site heterogeneity in multi-site MRI datasets.
• To enhance the learning of robust and generic representations from diverse data sources.

Main Methods:

• Proposed a novel multi-site network (MS-Net) incorporating Domain-Specific Batch Normalization layers.
• Implemented a Multi-site-guided Knowledge Transfer paradigm to capture shared knowledge across datasets.
• Evaluated the method on three heterogeneous prostate MRI datasets.

Main Results:

• The proposed MS-Net consistently improved segmentation performance across all tested datasets.
• The Domain-Specific Batch Normalization effectively compensated for inter-site heterogeneity.
• Multi-site-guided Knowledge Transfer enhanced the extraction of generic representations.

Conclusions:

• The MS-Net offers a robust solution for automated prostate segmentation using multi-site MRI data.
• The proposed methods effectively mitigate challenges posed by data heterogeneity in multi-site learning.
• MS-Net outperforms existing state-of-the-art methods in multi-site prostate MRI segmentation.