TRUSTED: Tridimensional Renal Ultrasound Tomodensitometry Dataset

TRUSTED provides paired 3D transabdominal ultrasound and CT kidney images with segmentation and anatomical landmark annotations for developing and validating inter-modal image registration and segmentation methods.

Key Features:

• Paired Imaging Data: The dataset contains paired transabdominal 3D ultrasound and CT kidney images from 48 human patients (96 kidneys).
• Segmentation Annotations: Kidney segmentations were annotated by two experienced radiographers with inter-rater agreement over 93% Dice score, and gold-standard segmentations were generated using the STAPLE algorithm.
• Anatomical Landmark Annotations: Seven anatomical landmarks were annotated for inter-modal image registration systems development and evaluation.
• Benchmarking Validation: Four deep-learning models for kidney segmentation were benchmarked, yielding average Dice scores from 79.63% to 90.09% for CT and 70.51% to 80.70% for ultrasound images.
• Registration Benchmarking: Four inter-modal image registration methods were benchmarked, with Coherent Point Drift performing best at 4.47 mm average Target Registration Error and 84.10% Dice score.

Scientific Applications:

• Image-Guided Surgery: The dataset supports research in image registration and segmentation for clinical applications in image-guided surgical procedures.
• Robotic Navigation: The dataset enables development of registration and segmentation methods for robotic navigation systems.
• Automatic Organ Measurement: The dataset facilitates research in automated kidney measurement through segmentation methods.
• Inter-Modal Image Registration Development: The dataset provides annotated landmarks and imaging pairs for developing and evaluating inter-modal image registration algorithms between ultrasound and CT modalities.
• Kidney Segmentation Research: The dataset supports training and validation of deep learning and other computational methods for kidney segmentation in both ultrasound and CT images.

Methodology:

Gold-standard segmentations were generated using the STAPLE algorithm from dual-rater annotations. Four deep-learning models and four inter-modal image registration methods including Coherent Point Drift were benchmarked for validation.