CASHNet: Context-Aware Semantics-Driven Hierarchical Network for Hybrid Diffeomorphic CT-CBCT Image Registration

Area of Science:

• Medical Imaging
• Computer Vision
• Machine Learning

Background:

• Accurate Computed Tomography (CT) to Cone-Beam Computed Tomography (CBCT) image registration is vital for image-guided radiotherapy and surgery.
• Challenges include inconsistent intensities, low contrast, and imaging artifacts, hindering precise alignment.

Purpose of the Study:

• To develop an advanced deep learning network, CASHNet, for robust CT-CBCT image registration.
• To enhance semantic structural perception and anatomical plausibility during image alignment.

Main Methods:

• Proposed a Context-Aware Semantics-driven Hierarchical Network (CASHNet) integrating context-aware semantics-encoded features.
• Employed diffeomorphisms for unified rigid and non-rigid registration in an end-to-end trainable network.
• Utilized a Siamese Mamba-based encoder and a coarse-to-fine decoder with Semantics-guided Velocity Estimation and Feature Alignment (SVEFA) modules.

Main Results:

• CASHNet demonstrated superior performance on challenging CT-CBCT datasets involving soft and hard tissues.
• The method achieved enhanced registration accuracy compared to existing state-of-the-art techniques.
• The network successfully enabled anatomically plausible deformations and preserved topological consistency.

Conclusions:

• CASHNet offers a significant advancement in CT-CBCT image registration accuracy and reliability.
• The proposed hierarchical, semantics-driven approach effectively addresses key challenges in medical image alignment.
• The method holds promise for improving image-guided interventions in radiotherapy and surgery.