TRUSTED: The Paired 3D Transabdominal Ultrasound and CT Human Data for Kidney Segmentation and Registration Research

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Dataset Development

Background:

• Inter-modal image registration (IMIR) and segmentation of abdominal Ultrasound (US) images are crucial for clinical applications like image-guided surgery.
• Research in these areas is hindered by a lack of publicly available, annotated datasets.

Purpose of the Study:

• To introduce the TRUSTED dataset, a novel resource for developing and validating AI algorithms for kidney image analysis.
• To provide paired 3D US and CT kidney images with expert annotations for segmentation and landmark identification.

Main Methods:

• The TRUSTED dataset includes paired transabdominal 3D US and CT kidney images from 48 patients (96 kidneys).
• Expert radiographers provided segmentation and anatomical landmark annotations; gold-standard segmentations were created using the STAPLE algorithm.
• The dataset was used to benchmark four deep learning models for kidney segmentation and four IMIR methods.

Main Results:

• Inter-rater segmentation agreement exceeded 93% Dice score.
• Benchmarked deep learning models achieved average Dice scores from 70.51% to 90.09% for US and CT images, respectively.
• Coherent Point Drift demonstrated the best performance for IMIR, with an average Target Registration Error of 4.47 mm and Dice score of 84.10%.

Conclusions:

• The TRUSTED dataset is a valuable, freely available resource for advancing research in medical image segmentation and inter-modal image registration.
• The dataset facilitates the development and validation of AI-driven tools for kidney analysis, potentially improving surgical guidance and organ measurement.