ThoraxNet: a 3D U-Net based two-stage framework for OAR segmentation on thoracic CT images

Area of Science:

• Medical Imaging
• Artificial Intelligence in Healthcare
• Radiation Oncology

Background:

• Accurate delineation of organs at risk (OAR) is crucial for radiation treatment planning and dose calculation.
• Manual contouring is time-consuming, labor-intensive, and susceptible to observer variability.
• Automated contouring tools are needed to improve efficiency and consistency in clinical practice.

Purpose of the Study:

• To develop and evaluate a novel two-stage deep learning model for automatic delineation of OARs in thoracic CT images.
• To enhance the accuracy and efficiency of organ segmentation compared to existing methods.
• To validate the clinical utility of the automated tool through dosimetric analysis.

Main Methods:

• A two-stage deep learning approach using 3D U-Net architectures with an attention mechanism was proposed.
• The model first locates organs to generate cropped images, then segmenting large and small organs separately.
• Post-processing integrates segmented organs, followed by dosimetric analysis for clinical validation.

Main Results:

• The proposed model achieved superior performance, outperforming state-of-the-art methods in Dice Similarity Coefficient (DSC) for lungs and heart.
• A DSC of 0.941 was achieved for the heart, surpassing previous benchmarks by 1.1%.
• The model processed a CT scan for five organs in just 8.61 seconds, demonstrating significant speed improvement.

Conclusions:

• The developed deep learning model provides accurate and rapid automatic contouring of OARs on thoracic CT images.
• The tool's efficiency and accuracy can accelerate radiation therapy planning, reduce costs, and potentially improve patient outcomes.
• The open-source nature of the tool facilitates wider adoption and further research in automated medical image segmentation.