Deep learning-based estimation of respiration-induced deformation from surface motion: A proof-of-concept study on 4D...

Area of Science:

• Medical Imaging
• Radiotherapy Physics
• Computational Biology

Background:

• Four-dimension computed tomography (4D-CT) is crucial for thoracic radiotherapy, but respiratory patterns challenge its quality and increase radiation exposure.
• Synthesizing 4D-CT images using estimated tissue deformation offers a potential solution to improve image quality and reduce radiation dose.
• Accurate estimation of respiration-induced thoracic tissue deformation is essential for effective 4D-CT synthesis.

Purpose of the Study:

• To propose a non-patient-specific cascaded ensemble model (CEM) for estimating thoracic tissue deformation from surface motion.
• To develop a deep learning-based approach for synthesizing 4D-CT images.
• To provide a solution that does not require patient-specific breathing data or additional pre-treatment training.

Main Methods:

• The CEM utilizes three cascaded deep learning models to output a deformation vector field (DVF) from surface motion.
• Surface motion was simulated using body contours derived from 4D-CT data.
• The model was trained on a private database (62 4D-CT sets) and tested on a public database (80 4D-CT sets).

Main Results:

• CEM synthesized CT images with an average root mean square error (mRMSE) of 61.06 ± 10.43 HU and an average mean absolute error (mMAE) of 26.80 ± 5.65 HU.
• The synthesized images achieved a high average structural similarity index measure (mSSIM) of 0.990 ± 0.004.
• CEM outperformed other published methods in synthesized CT quality.

Conclusions:

• The CEM effectively estimates thoracic tissue deformation (DVF) from surface motion.
• The model's non-patient-specific nature and lack of pre-treatment training requirement facilitate broad clinical application.
• CEM shows significant potential for improving 4D-CT synthesis in thoracic radiotherapy planning.