Introduction
In the field of image-guided surgery and robotics, real-time visualization of target organs is essential for precision and safety. While CT and MRI provide high-resolution anatomical maps, they are not feasible for intra-operative use. Ultrasound (US) is an ideal companion due to its portability and lack of ionizing radiation, yet it remains difficult to interpret because of low contrast, speckle noise, and acoustic shadows from the rib cage. Inter-Modal Image Registration (IMIR) attempts to bridge this gap by overlaying pre-operative CT scans onto real-time US feeds. However, progress has been limited by a lack of publicly available, paired human datasets for training and benchmarking algorithms.
The TRUSTED Dataset
The Tridimensional Renal Ultra Sound TomodEnsitometrie (TRUSTED) dataset is the first public collection of paired 3D transabdominal US and CT scans from human patients specifically for kidney research. It includes data from 48 patients, providing a total of 96 kidneys for study.
- CT Data: Three-phase scans (non-contrast, venous, and portal).
- US Data: 59 high-quality 3DUS volumes where the kidney is fully contained within the field of view.
- Annotations: The dataset features binary masks (segmentations), surface meshes, and 3D coordinates for seven anatomical landmarks per kidney.
Methodology
To establish a "gold standard" ground truth, the researchers employed a rigorous annotation pipeline. Two experienced radiographers independently annotated each kidney. To reconcile inter-annotator differences, the team used the Simultaneous Truth and Performance Level Estimation (STAPLE) algorithm to generate consensus segmentations. The dataset was then used to benchmark five state-of-the-art deep learning (DL) models—including nnUNet and CoTr—and several registration methods such as Iterative Closest Point (ICP) and Bayesian Coherent Point Drift (BCPD).
Key Results and Insights
The "US Difficulty Gap"
The benchmarking results highlighted a significant performance gap between modalities. While AI models achieved nearly 90% accuracy (Dice score) on CT data, they peaked at approximately 79% for ultrasound. This discrepancy underscores that automated US segmentation remains a complex, unsolved challenge due to the modality's inherent noise and artifacts.
Segmentation Performance
Among the tested models, nnUNet was the most robust for ultrasound segmentation, achieving a Dice score of 79.4%. For CT scans, the CoTr (Convolutional Transformer) model performed best with a Dice score of 89.1%.
Registration and Surface-Based Approaches
The study found that surface-based registration—aligning the extracted shapes of the kidneys—generally outperformed intensity-based registration (aligning raw voxels). The Bayesian Coherent Point Drift (BCPD) method using affine transforms emerged as the top performer for registration, achieving an average Target Registration Error (TRE) of 4.53mm. Sensitivity analyses further revealed that surface-based methods are significantly more resilient to poor initial alignments than intensity-based methods.
Discussion and Future Work
The TRUSTED dataset reveals that even human experts exhibit slight disagreements (94% agreement), highlighting the inherent ambiguity in medical imaging. This "label uncertainty" is a critical feature for researchers developing AI that must navigate real-world clinical noise.
While the 59 3DUS volumes represent a major step forward for public data, the dataset size remains a limitation for training massive deep learning models from scratch. Future work will focus on developing "shadow-aware" architectures to handle rib interference and exploring "Double Target" training strategies that utilize multiple human annotations to improve model robustness.
Summary updated