Annotated normal CT data of the abdomen for deep learning: Challenges and strategies for implementation

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Deep learning algorithms require large, annotated datasets for training.
• Accurate annotation of abdominal structures in CT images is crucial for developing reliable AI models.
• Previous methods for pancreas segmentation lacked standardized annotation procedures.

Purpose of the Study:

• To develop and report procedures for annotating abdominal CT images of subjects without pancreatic disease.
• To create a dataset suitable for training deep convolutional neural networks (DNNs).
• To facilitate the development of deep learning algorithms for automatic recognition of a normal pancreas.

Main Methods:

• Retrospective assessment of dual-phase contrast-enhanced volumetric CT scans from 575 potential kidney donors.
• Manual annotation of 22 abdominal structures by four trained annotators, confirmed by expert radiologists.
• Utilization of commercial software supporting 3D segmentation for efficient data management.

Main Results:

• Annotation of 1150 CT datasets from 575 subjects (229 men, 346 women; mean age 45±12 years).
• High fidelity in deep network predictions for multi-organ segmentation, achieving 89.4% Dice similarity coefficient and 1.29mm mean surface distance.
• Low mean intra-observer intra-subject volume difference of 4.27mL (7.65%) for annotated structures.

Conclusions:

• A reliable data collection and annotation process for abdominal structures was successfully developed.
• The established process is suitable for generating large datasets essential for deep learning applications.
• This methodology supports the advancement of AI in medical image analysis for pancreas assessment.