Abstract
Medical image registration for PET-CT scans is essential for longitudinal disease tracking but is often hampered by the intensive manual labor required to generate 3D ground-truth labels. We present a novel framework that utilizes "key curves" derived from sparse 2D annotations for evaluation and a self-supervised training loop with cross-visit semantic consistency. This approach achieves high-accuracy 3D alignment while significantly reducing the expert effort needed for model development and validation.
Introduction
PET-CT registration aligns scans taken at different time points, a process complicated by physiological changes such as weight loss or postural shifts. Traditional deep learning methods for this task require massive amounts of perfectly aligned 3D volumes, which are difficult and time-consuming for clinicians to produce. This "data bottleneck" has limited the adoption of automated registration tools in clinical workflows.
Methodology
The proposed method addresses the data bottleneck through two primary innovations: key-curve evaluation and weakly supervised training.
Key-Curves for Evaluation
Instead of full 3D volume alignment, experts provide corresponding points on several 2D slices. These points are fitted to a second-order polynomial to create a 3D "key curve." This method is more intuitive for humans and provides a robust metric for measuring misalignment without requiring dense point-to-point correspondence.
Weakly Supervised Training and Semantic Consistency
The training process is entirely data-driven, utilizing self-supervision where the network learns to undo synthetic 3D deformations. To handle real-world variations between patient visits, the authors employ Linear Centered Kernel Alignment (LCKA). This technique enforces semantic consistency in the network's feature maps across different visits, ensuring the model focuses on structural alignment rather than transient biological changes.
Results and Discussion
The method was evaluated on the QIN-HEADNECK dataset. The full proposed method reduced the Root Mean Squared Error (RMSE) of the key curves from 29.25 mm (unaligned) to 13.41 mm. Ablation studies confirmed that the LCKA component is crucial, as its removal increased the error to 15.89 mm. Qualitative analysis showed that the model successfully aligns complex structures like the spine and ribs, which are typically challenging for automated systems.
Conclusion
By decoupling training from human-intensive labeling and introducing a curve-based evaluation metric, this research significantly lowers the barrier to developing high-performance PET-CT registration models. Future work will explore making the key-curve metric differentiable to allow for direct optimization during training.
Summary updated