Training Convolutional Networks for Prostate Segmentation With Limited Data.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Multi-zonal segmentation of prostate cancer on MRI is crucial for computer-aided diagnosis.
• Convolutional Neural Networks (CNNs) like U-Net achieve expert-level prostate segmentation but require extensive manual training data.
• Limited labeled data at institutions hinders the development of effective site-specific segmentation models.

Purpose of the Study:

• To compare transfer learning and aggregated training strategies for improving prostate MRI segmentation with limited site-specific data.
• To evaluate the impact of varying amounts of internal training data on segmentation performance.
• To optimize the U-Net architecture (2D vs. 3D, fine-tuning depth) for transfer learning.

Main Methods:

• Compared transfer learning and aggregated training using public external data against internal prostate MRI datasets.
• Investigated the effect of data quantity (5-40 internal datasets) on segmentation performance.
• Optimized U-Net architecture, including 2D/3D variants and fine-tuning depth, for transfer learning.

Main Results:

• Both transfer learning and aggregated training significantly improved segmentation performance compared to using limited internal data alone.
• Expert-level segmentation results were achieved with approximately 20 site-specific MRI datasets using these strategies.
• Cross-training experiments highlighted the impact of differences between internal and external datasets.

Conclusions:

• Transfer learning and aggregated training are effective strategies for developing site-specific prostate MRI segmentation models with limited data.
• These methods enable the creation of models comparable to human experts, even with small datasets.
• Findings provide guidance for clinical and research applications requiring accurate prostate cancer segmentation.