Cooperative AI training for cardiothoracic segmentation in computed tomography: An iterative multi-center annotation...

Area of Science:

• Medical imaging analysis
• Artificial intelligence in radiology

Background:

• Quantitative radiological reporting requires validated biomarkers.
• Voxelwise annotation of medical images is a bottleneck for large-scale studies.
• High-resolution thoracic CT data necessitates efficient segmentation methods.

Purpose of the Study:

• To develop and evaluate an iterative training workflow for AI-based segmentation.
• To accelerate and improve the accuracy of heart and mediastinum segmentation in thoracic CT scans.
• To reduce human involvement in the annotation process for multi-center studies.

Main Methods:

• Utilized 132 thoracic CT scans annotated by 13 radiologists.
• Implemented three iterative training experiments using a nnU-Net model.
• Incorporated AI pre-segmentation and human correction in subsequent iterations.

Main Results:

• Iterative training consistently improved AI model quality and segmentation accuracy (Dice Similarity Coefficient > 0.90).
• AI models reduced human interaction time by up to 70% for segmentation tasks.
• Satisfactory results were achieved even with models trained on as few as five datasets.

Conclusions:

• The iterative training workflow offers an efficient solution for developing AI segmentation models in multi-center research.
• This approach enhances model accuracy and reduces the need for extensive human annotation.
• Future research will focus on minimizing initial datasets and exploring advanced pre-processing techniques.