A deep learning based automatic segmentation approach for anatomical structures in intensity modulation radiotherapy.

Area of Science:

• Medical Physics
• Radiotherapy
• Medical Imaging

Background:

• Accurate segmentation of organs at risk (OARs) is crucial for effective radiotherapy planning.
• Manual contouring of OARs is time-consuming and subject to inter-observer variability.
• Deep learning-based automatic segmentation offers a potential solution to improve efficiency and consistency.

Purpose of the Study:

• To evaluate an automatic segmentation approach for OARs in radiotherapy.
• To compare dose volume histogram (DVH) parameters derived from automatic versus manual contours.
• To assess the accuracy and efficiency of a U-Net-based automatic segmentation method.

Main Methods:

• A U-Net-based automatic segmentation approach was applied to OARs for nasopharyngeal carcinoma (NPC), breast, and rectal cancer cases.
• Automatic contours were transferred to the Pinnacle System for accuracy evaluation and DVH parameter comparison.
• Manual contouring times were recorded and compared against automatic segmentation times.

Main Results:

• Automatic segmentation significantly reduced contouring time (e.g., 1.5 min vs. 56.5 min for NPC OARs).
• Dice-similarity coefficients (DSC) varied, with high values for some structures (e.g., Eye DSC 0.907) and lower values for others (e.g., Spinal Cord DSC 0.459).
• Poor DSC for Spinal Cord (NPC) and Femoral heads (rectal) were noted, attributed to anatomical complexities and manual contouring variability.

Conclusions:

• The automatic contouring approach demonstrates sufficient accuracy for research purposes in radiotherapy.
• While DSC is a useful metric, it may not fully reflect dose distribution accuracy.
• The method offers significant time savings and a valuable supervisory role in OAR contouring.