Methods for improving colorectal cancer annotation efficiency for artificial intelligence-observer training.

Area of Science:

• Radiology
• Artificial Intelligence
• Oncology

Background:

• Missing occult cancer lesions are a primary cause of diagnostic errors in radiology.
• Artificial intelligence (AI) as a second observer offers an economical solution to reduce these errors.
• Large annotated datasets are crucial for effective AI model training in cancer detection.

Purpose of the Study:

• To compare skip-slice annotation and AI-initiated annotation for decreasing AI model training time.
• To evaluate the efficiency of different annotation methods in establishing ground truth for AI development.

Main Methods:

• Developed a 2D U-Net AI model for colorectal cancer (CRC) detection.
• Employed an ensemble of 2D U-Nets for enhanced performance.
• Trained and tested models using The Cancer Imaging Archive dataset, comparing skip-slice and AI-initiated annotation techniques.

Main Results:

• Sparse annotation, particularly skipping two slices, significantly reduced annotation time (P < 0.001).
• Reduced annotation (up to 2/3) did not negatively impact AI model sensitivity or false positive rates.
• AI-initiated annotation provided minimal time reduction, even with an ensemble AI approach.

Conclusions:

• Sparse annotation is an efficient technique for reducing the time required to establish ground truth for AI models.
• This method supports faster development of AI tools for improved cancer lesion detection.