Joint Segmentation of Multiple Thoracic Organs in CT Images with Two Collaborative Deep Architectures

Area of Science:

• Medical Imaging
• Radiotherapy Planning
• Deep Learning

Background:

• Computed Tomography (CT) is essential for radiotherapy planning.
• Accurate delineation of Organs at Risk (OARs) is critical to protect healthy tissues during radiation therapy.
• Multi-organ segmentation in thoracic CT is challenging due to low contrast and ill-defined borders.

Purpose of the Study:

• To develop a novel framework for automatic, collaborative delineation of multiple OARs in thoracic CT images.
• To address the challenge of low contrast and ill-defined organ borders by incorporating spatial relationships.
• To improve the accuracy of OAR segmentation for enhanced radiotherapy planning.

Main Methods:

• Proposed a novel framework with two collaborative deep architectures for joint OAR segmentation.
• Utilized a deep SharpMask architecture combined with Conditional Random Fields (CRF) to capture low-level and high-level features and spatial relationships.
• Employed a second deep architecture to refine segmentation using anatomical constraints learned from the first network.

Main Results:

• The proposed framework demonstrated superior performance in segmenting multiple OARs (esophagus, heart, aorta, trachea) compared to state-of-the-art methods.
• Achieved improved segmentation accuracy on 30 thoracic CT scans.
• Effectively leveraged spatial relationships between organs to overcome low contrast challenges.

Conclusions:

• The novel collaborative deep learning framework offers a significant advancement in automatic OAR delineation for radiotherapy planning.
• Joint segmentation considering inter-organ spatial relationships enhances accuracy, addressing limitations of single-organ segmentation methods.
• This approach holds promise for improving the safety and efficacy of radiotherapy treatments.