Comparative Evaluation of Conventional and Deep Learning Methods for Respiratory Signal Extraction From Clinical 3D...

Area of Science:

• Medical Imaging
• Deep Learning
• Computational Imaging

Background:

• 4D cone-beam CT (CBCT) reconstruction requires accurate respiratory signal extraction from 3D CBCT projections.
• Extracting reliable respiratory signals from CBCT projections, especially for 4D CBCT phase sorting, remains a significant challenge.

Purpose of the Study:

• To evaluate and compare conventional methods with a deep learning approach for extracting respiratory signals from clinical 3D CBCT projections.
• To assess the efficacy of these methods in handling regular and irregular respiratory motion patterns in thoracic and abdominal cancer patients.

Main Methods:

• Analyzed 70 sets of clinical 3D CBCT projections from cancer patients.
• Compared conventional methods (Intensity Analysis, Fourier Transform, Amsterdam Shroud, Local Principal Component Analysis) with a U-Net deep learning model.
• Evaluated signal extraction using correlation analysis and phase-sorting capability, referencing the diaphragm apex.

Main Results:

• The U-Net deep learning method significantly outperformed all conventional methods, achieving a high correlation coefficient (0.93 ± 0.07).
• Local Principal Component Analysis and Amsterdam Shroud showed better performance than Intensity Analysis and Fourier Transform among conventional techniques.
• LPCA was noted as superior to AS due to AS's sensitivity to bandpass filter cutoff frequencies.

Conclusions:

• The U-Net deep learning model demonstrates superior performance in extracting respiratory signals from 3D CBCT projections.
• This advancement holds significant potential for improving respiratory phase sorting and the quality of 4D CBCT reconstruction.
• Deep learning offers a robust solution for overcoming challenges in respiratory motion management in CT imaging.