Opportunity
Deformable image registration is a critical task in medical imaging, particularly for aligning 3D images of organs or tissues across different scans. However, existing methods face significant challenges when dealing with images containing tumors. Traditional similarity-based registration techniques often lead to disproportionate volume changes in tumor regions, as tumors lack corresponding structures between pre- and post-treatment scans. This issue is especially problematic in clinical applications like tumor growth tracking, where preserving tumor volume is essential for evaluating treatment efficacy (e.g., radiotherapy or chemotherapy). Current learning-based methods, while achieving high registration accuracy and speed, fail to address this problem adequately. The absence of a robust solution for volume-preserving tumor registration limits the clinical utility of these advanced techniques. This patent addresses this gap by introducing a novel framework that ensures tumor volumes remain unchanged during registration while maintaining alignment accuracy in non-tumor regions.
Technology
The patent discloses a two-stage unsupervised framework for deformable medical image registration that preserves tumor volumes. In the first stage, a similarity-based registration network identifies potential tumor regions by analyzing volume changes (via Jacobian determinants) in the deformation field. These regions are then mapped to a soft tumor mask (STM), which probabilistically labels voxels as tumor or non-tumor based on their volume-change behavior. The second stage employs a volume-preserving registration network trained with an adaptive loss function. This loss function penalizes volume changes in tumor regions (guided by the STM) while promoting similarity in non-tumor regions through a weighted similarity loss. The framework balances these objectives using the STM to adjust constraints dynamically, ensuring tumors retain their original size during alignment. Key innovations include:
1. Unsupervised Tumor Mask Estimation: No manual segmentation is required; tumors are identified by their atypical volume changes during initial registration.
2. Adaptive Volume-Preserving Loss: A novel loss function that selectively enforces volume preservation based on the STM’s probability values.
3. Preregistration Step: Optional edge-preserving preprocessing to improve STM accuracy by excluding boundary artifacts.
Advantages
- Clinical Relevance: Enables accurate tumor growth tracking by preserving volumetric integrity during registration.
- Unsupervised Learning: Eliminates the need for annotated tumor masks, reducing dependency on scarce labeled data.
- Compatibility: Integrates with existing CNN- and transformer-based architectures (e.g., VoxelMorph, TransMorph).
- Robustness: Tolerates noisy STM estimates, ensuring reliable performance even with imperfect tumor region identification.
Applications
- Oncology: Longitudinal studies of tumor progression/regression post-treatment (e.g., chemotherapy, radiotherapy).
- Surgical Planning: Alignment of pre-operative and intra-operative scans for resection guidance.
- Medical Research: Atlas-based segmentation and multi-modal image fusion involving pathological tissues.
