Performance Analysis in Children of Traditional and Deep Learning CT Lung Nodule Computer-Aided Detection Systems...

Area of Science:

• Medical imaging analysis
• Artificial intelligence in healthcare
• Pediatric oncology

Background:

• Chest CT scans are crucial for detecting lung metastases in pediatric cancer patients.
• Existing computer-aided detection (CAD) systems for lung nodules are primarily trained on adult data.
• The effectiveness of adult-trained CAD systems in pediatric populations remains largely unexamined.

Purpose of the Study:

• To evaluate the diagnostic performance of traditional and deep learning CAD systems trained on adult data for pediatric lung nodule detection on chest CT.
• To compare the generalization ability of these CAD systems between pediatric and adult patient cohorts.

Main Methods:

• Retrospective analysis of pediatric (n=59) and adult (LUNA 2016 subset 0, n=89) chest CT scans.
• Utilized traditional (FlyerScan) and deep learning (MONAI) CAD systems trained on adult CT data.
• Calculated sensitivity and false-positive (FP) frequency for nodules measuring 3-30 mm.

Main Results:

• Adult-trained CAD systems demonstrated significantly lower detection sensitivity on pediatric data (FlyerScan: 68.4%, MONAI: 53.1%) compared to adult data (FlyerScan: 83.9%, MONAI: 95.5%) at two FPs per scan.
• Pediatric lung nodules were significantly smaller (mean 5.4 mm) than adult nodules (mean 11.0 mm).
• Performance disparities suggest adult-trained models struggle with the characteristics of pediatric lung nodules.

Conclusions:

• Adult-trained lung nodule CAD systems exhibit reduced sensitivity in pediatric patients compared to adults.
• Smaller nodule size in children is a potential factor contributing to decreased CAD performance.
• Development of pediatric-specific lung nodule CAD systems trained on relevant pediatric data is essential.