3D spatial priors for semi-supervised organ segmentation with deep convolutional neural networks.

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Fully Convolutional Neural Networks (FCNs) are widely used for medical image segmentation but lack explicit spatial organ position integration.
• Integrating spatial organ position information is critical for accurate medical image labeling, particularly in complex cases.

Purpose of the Study:

• To develop a novel method, STIPPLE, that combines 3D organ position priors with FCN predictions for enhanced medical image segmentation.
• To leverage spatial prior information for improved pseudo-label selection in low-data regimes through a self-labeling process.

Main Methods:

• A generalized prior-driven prediction function was developed to integrate 3D organ position probabilities with FCN visual predictions.
• A self-labeling strategy was implemented, utilizing the spatial prior to enhance pseudo-label quality in low-data scenarios.

Main Results:

• The STIPPLE model demonstrated significant performance improvements over baseline FCNs on CT pancreas segmentation datasets.
• STIPPLE outperformed state-of-the-art semi-supervised segmentation methods by effectively utilizing spatial prior information.
• Performance gains were particularly notable in scenarios with limited training images.

Conclusions:

• STIPPLE offers an effective solution for medical image segmentation using few labeled examples, addressing a key challenge in the medical domain.
• The method provides an intuitive approach to incorporating absolute spatial information, mimicking the process of expert annotation.