CT-ORG, a new dataset for multiple organ segmentation in computed tomography.

Area of Science:

• Medical Imaging and Computer Vision
• Artificial Intelligence in Healthcare

Background:

• Automated organ segmentation is crucial for medical image analysis but is limited by insufficient labeled training data.
• Existing datasets often lack diversity, covering few cases or organs, and may not reflect real-world clinical imaging conditions.

Purpose of the Study:

• To develop a comprehensive and diverse dataset for training and evaluating automated organ segmentation models.
• To address the scarcity of labeled data and improve the generalization of segmentation algorithms across multiple organs and imaging variations.

Main Methods:

• Created a dataset of 140 CT scans featuring six organ classes: liver, lungs, bladder, kidney, bones, and brain.
• Utilized unsupervised morphological segmentation algorithms accelerated by 3D Fourier transforms for efficient annotation of lungs and bones.
• Trained a deep neural network for simultaneous segmentation of all organs and developed a GPU library for data augmentation.

Main Results:

• The developed dataset enables rapid, simultaneous segmentation of multiple organs in CT scans, with the deep neural network achieving a processing time of 4.3 seconds per case.
• Data augmentation techniques were shown to effectively improve model generalization.
• The dataset and associated code are made available via The Cancer Imaging Archive (TCIA).

Conclusions:

• The newly created diverse CT dataset and accompanying tools significantly advance the field of automated organ segmentation.
• This resource is expected to facilitate the development and validation of more robust and clinically applicable segmentation models.