Robust and Generable Vertebrae Instance Segmentation with Domain Adaptation.

Area of Science:

• Medical Imaging
• Computer Vision
• Radiology

Background:

• Accurate identification and segmentation of vertebrae are crucial for diagnosing and treating spinal conditions.
• Challenges include similar vertebral appearance and anatomical variations, limiting clinical data utility.

Purpose of the Study:

• To present a streamlined framework for precise spine localization and segmentation.
• To enhance model performance in clinical settings through iterative domain adaptation.

Main Methods:

• Utilized a U-Net architecture variant for spine localization and segmentation.
• Developed an iterative domain adaptation strategy to improve clinical applicability.
• Evaluated performance on the VerSe'20 dataset and a PUTH dataset (Peking University Third Hospital spine CT).

Main Results:

• The proposed method significantly improved spine localization and segmentation performance on both datasets.
• Domain adaptation notably enhanced vertebrae localization accuracy on the PUTH dataset.
• Validated effectiveness across various models on the PUTH dataset.

Conclusions:

• The streamlined framework offers accurate spine localization and segmentation.
• Iterative domain adaptation is effective for improving model performance in clinical environments.
• The approach holds promise for advancing spinal imaging analysis in healthcare.