Automated multi-lesion annotation in chest X-rays: annotating over 450,000 images from public datasets using the...

Area of Science:

• Artificial Intelligence in Medical Imaging
• Radiology and Diagnostic Imaging
• Computational Pathology

Background:

• Accurate annotation of pulmonary lesions on chest X-rays (CXRs) is crucial for AI development.
• Existing annotation methods can be time-consuming and labor-intensive.
• Large, high-quality datasets are essential for training robust AI models.

Purpose of the Study:

• To develop and evaluate an AI-based image annotation tool, Smart Imagery Framing and Truthing (SIFT), for pulmonary abnormalities on CXRs.
• To assess the efficiency and accuracy of SIFT compared to traditional semi-automatic methods.
• To improve the quality and quantity of annotated CXR datasets for AI training.

Main Methods:

• Utilized the SIFT tool, based on Multi-task, Optimal-recommendation, and Max-predictive Classification and Segmentation (MOM ClaSeg) technologies, to annotate 452,602 CXR images from four public datasets.
• SIFT identifies and delineates 65 abnormal regions of interest (ROIs), provides confidence scores, and offers recommendations.
• The underlying MOM ClaSeg system integrates Mask R-CNN and Decision Fusion Network, trained on over 300,000 CXRs.

Main Results:

• SIFT demonstrated high efficiency, improving annotation speed by 7.92 times when used by radiologists compared to traditional methods.
• The system achieved an average sensitivity of 89.38%±11.46% across four diverse CXR datasets.
• SIFT accurately predicts abnormality types and boundary locations for ROIs on CXR images.

Conclusions:

• The SIFT system offers a highly efficient and accurate solution for annotating pulmonary lesions and abnormalities on CXRs.
• This AI tool significantly enhances the process of creating training and testing datasets for medical AI development.
• SIFT has the potential to accelerate advancements in AI-driven diagnostic imaging.