An unsupervised multi-scale framework with attention-based network (MANet) for lung 4D-CT registration.

Area of Science:

• Medical Imaging
• Radiotherapy
• Computational Anatomy

Background:

• Deformable image registration (DIR) of 4D-CT is crucial for radiotherapy applications like tumor definition and dose accumulation.
• Accurate and fast DIR of lung 4D-CT images is challenging due to complex deformations.

Purpose of the Study:

• To propose an unsupervised multi-scale DIR framework with an attention-based network (MANet) for improved 4D-CT lung image registration.
• To enhance accuracy and reduce computation time for DIR in radiotherapy.

Main Methods:

• Developed a cascaded multi-scale network (MANet) with attention gates to distinguish moving and non-moving structures.
• Employed unsupervised learning by minimizing dissimilarity and DVF regularization loss functions.
• Integrated an adversarial network to enforce realistic deformations and incorporated joint training for optimal performance.

Main Results:

• MANet achieved an average Target Registration Error (TRE) of 1.53 ± 1.02 mm, outperforming conventional and deep learning methods.
• The method demonstrated a fast execution time of approximately 1 second for DVF estimation.
• No manual parameter tuning was required, indicating robustness and ease of use.

Conclusions:

• The proposed MANet framework offers superior performance for deformable image registration of 4D-CT lung images.
• MANet provides an accurate, fast, and robust solution for critical radiotherapy applications.
• The attention-based multi-scale approach effectively handles complex lung deformations.