Semi-supervised temporal attention network for lung 4D CT ventilation estimation.

Area of Science:

• Medical Imaging
• Radiotherapy
• Artificial Intelligence

Background:

• Computed tomography (CT)-derived ventilation estimation (CTVI) is vital for radiotherapy planning and response evaluation.
• Conventional CTVI methods suffer from registration errors, limiting accuracy.
• Current deep learning CTVI methods require extensive labeled data and underutilize temporal information from 4D CT.

Purpose of the Study:

• To develop a novel semi-supervised temporal attention (S²TA) network for improved lung 4D CT ventilation estimation.
• To address the limitations of existing CTVI methods, including registration errors and data requirements.
• To enhance the accuracy and temporal utilization of CTVI for clinical applications.

Main Methods:

• Proposed a semi-supervised learning framework with a teacher-student model for CT ventilation imaging.
• Utilized a temporal attention architecture to effectively capture temporal relationships in 4D CT image sequences.
• Trained the model using both labeled and unlabeled 4D CT data, with a teacher model updated via moving average to ensure stability.

Main Results:

• The S²TA network demonstrated higher estimation accuracy compared to state-of-the-art methods in extensive experiments on three public datasets.
• The method effectively utilizes temporal information from 4D CT images.
• Achieved superior performance in lung 4D CT ventilation estimation.

Conclusions:

• The proposed S²TA network offers a more accurate and robust approach to CT-derived ventilation estimation.
• This method has the potential to significantly benefit lung functional avoidance radiotherapy planning and treatment response modeling.
• The semi-supervised and temporal attention approach overcomes key limitations of previous CTVI techniques.