An unsupervised convolutional neural network-based algorithm for deformable image registration.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Deformable image registration (DIR) is crucial for aligning medical images, particularly in radiotherapy where cone-beam CT (CBCT) is used for patient positioning.
• Current DIR methods often struggle with the noise and artifacts present in CBCT images, impacting treatment accuracy.
• Deep learning approaches offer potential for more robust and efficient image registration.

Purpose of the Study:

• To develop and evaluate a deep unsupervised learning strategy for CBCT to CT deformable image registration (DIR).
• To assess the performance of a Deep Convolutional Inverse Graphics Network (DCIGN) based DIR algorithm.
• To compare the accuracy and efficiency of DCIGN against established registration methods.

Main Methods:

• A Deep Convolutional Inverse Graphics Network (DCIGN) was developed for unsupervised DIR, featuring an encoding-decoding architecture.
• The model was trained on 285 head and neck patient datasets using a distributive learning-based convolutional neural network.
• DCIGN was evaluated on synthetic and real patient data, comparing its performance against rigid registration, Demons, and landmark-guided DIR.

Main Results:

• DCIGN demonstrated superior performance across all evaluation metrics compared to rigid registration, intensity-corrected Demons, and landmark-guided DIR.
• The algorithm maintained high accuracy even with CBCT noise contamination.
• DCIGN achieved significant computational efficiency, requiring approximately 14 hours for training and 3.5 seconds for prediction on a 512x512x120 voxel image.

Conclusions:

• The proposed DCIGN strategy provides an accurate and computationally efficient solution for CBCT to CT deformable image registration.
• This deep unsupervised learning approach effectively handles noise in CBCT images, improving registration reliability.
• DCIGN represents a promising advancement for image-guided radiotherapy and other clinical applications requiring precise image alignment.