Related Experiment Video
Updated: Jun 25, 2026
04:23
A Swin Transformer-Based Model for Thyroid Nodule Detection in Ultrasound Images
Published on: April 21, 2023
2.4K
Comparative Performance of 3 Artificial Intelligence Systems for Lung Nodule Characterization in Low-Dose Computed
Khulan Khurelsukh1,2, Yen-Po Lin3, Hsuan-Ming Chang3
1Department of Medical Imaging and Radiological Sciences, College of Medicine, Chang Gung University.
Journal of Thoracic Imaging
|March 12, 2026
Summary
Three artificial intelligence (AI) systems showed varied performance in detecting lung nodules on low-dose computed tomography (LDCT) scans. Users must understand AI system characteristics for accurate clinical interpretation of lung nodule findings.
Area of Science:
- Radiology
- Artificial Intelligence
- Medical Imaging
Background:
- Lung nodules are common findings on low-dose computed tomography (LDCT).
- Accurate detection and characterization of lung nodules are crucial for early lung cancer diagnosis.
- Artificial intelligence (AI) systems are emerging tools for analyzing medical images, including LDCT scans.
Purpose of the Study:
- To evaluate the performance of three commercial AI systems in detecting, characterizing, and classifying lung nodules on LDCT scans.
- To assess the agreement of AI systems with a reference standard established by expert radiologists.
- To determine the inter-vendor consistency among different AI platforms for lung nodule analysis.
Main Methods:
- Retrospective analysis of LDCT scans from 100 subjects using three AI platforms (AI 1, AI 2, AI 3).
- Comparison of AI system performance against a consensus reference standard derived from two thoracic radiologists.
- Assessment of agreement for nodule presence, type, and Lung-RADS category using Cohen Kappa, and for continuous measurements (diameter, volume) using intraclass correlation coefficients (ICC).
Main Results:
- AI systems detected significantly more nodules (435, 152, 70) compared to radiologists (126).
- AI 2 and AI 3 demonstrated high accuracy (80.6%, 82.2%) and substantial agreement with the reference standard (perfect for AI 2, almost perfect for AI 3).
- AI 1 showed poor performance, while inter-AI agreement was substantial (κ=0.66-0.78) and measurement reliability was moderate to good (ICC=0.57-0.87).
Conclusions:
- Commercial AI systems exhibit variable performance in lung nodule detection and classification.
- Understanding the specific characteristics and limitations of each AI system is essential for clinical application.
- AI results should be interpreted cautiously within the broader clinical context to ensure optimal patient care.
