3D-WDA-PMorph: Efficient 3D MRI/TRUS Prostate Registration using Transformer-CNN Network and...

Area of Science:

• Medical Imaging
• Computer Vision
• Biomedical Engineering

Background:

• Multimodal image registration, especially for MRI and TRUS in prostate cancer, is vital but challenging due to modality differences.
• CNNs excel at local features but miss global context; Transformers capture long-range dependencies but struggle with fine details.
• Accurate registration requires integrating both local and global information for precise spatial alignment.

Purpose of the Study:

• To develop a novel 3D image registration framework addressing limitations of CNN and Transformer methods.
• To enhance multimodal registration accuracy and robustness for prostate cancer diagnosis and treatment planning.
• To introduce an innovative Wavelet-3D-Depthwise-Attention (WDA) module for improved feature fusion and detail preservation.

Main Methods:

• A Swin Transformer (ST)-CNN encoder-decoder architecture was employed for 3D image registration.
• A novel Wavelet-3D-Depthwise-Attention (WDA) module was integrated into the skip connections.
• WDA utilizes wavelet transforms for multi-scale spatial-frequency analysis and 3D-Depthwise convolution for efficiency and fusion.

Main Results:

• The proposed method achieved a median Dice score of 0.94 and a target registration error of 0.85 on clinical MRI/TRUS datasets.
• Demonstrated significant improvements in registration accuracy and robustness compared to state-of-the-art methods.
• WDA-enhanced skip connections effectively preserved critical anatomical details crucial for accurate alignment.

Conclusions:

• The novel ST-CNN framework with WDA module offers a promising advancement for multimodal prostate image registration.
• The method effectively combines local and global feature extraction, overcoming limitations of existing approaches.
• The framework shows potential for generalization to other medical image registration tasks.