Lumbar and Thoracic Vertebrae Segmentation in CT Scans Using a 3D Multi-Object Localization and Segmentation CNN.

Area of Science:

• Medical Imaging
• Radiotherapy
• Artificial Intelligence

Background:

• Accurate segmentation of vertebrae in computed tomography (CT) scans is crucial for radiation treatment planning, especially for cancers affecting nearby bone structures.
• Existing methods may require extensive training data or lack the precision needed for clinical applications.

Purpose of the Study:

• To develop and validate a novel automated 3D segmentation method for individual lumbar and thoracic vertebrae in CT scans.
• To achieve high segmentation accuracy suitable for clinical use in radiation therapy.

Main Methods:

• Utilized a single, low-complexity convolutional neural network (CNN) architecture for automated 3D segmentation.
• Employed volume patch-based processing to handle arbitrary scan sizes and segment vertebrae and estimate center locations in one step.
• Implemented an advanced post-processing scheme to enhance segmentation accuracy.

Main Results:

• Achieved a Dice coefficient of 0.921 ± 0.047 and a signed distance error of 0.271 ± 0.748 mm on CT scans for radiation treatment planning.
• On the VerSe2020 dataset (129 CT scans), obtained an overall Dice coefficient of 0.940 ± 0.065 and a signed distance error of 0.109 ± 0.301 mm.
• Demonstrated superior overall segmentation performance compared to other methods validated on the VerSe dataset.

Conclusions:

• The proposed automated method provides accurate and efficient 3D segmentation of vertebrae in CT scans.
• This approach is effective even with limited application-specific training data, making it valuable for clinical radiation therapy.
• The method's high accuracy and robustness offer significant potential for improving radiation treatment planning.