Multi-Level Classification of Lung Pathologies in Neonates using Recurrence Features.

Area of Science:

• Medical Imaging
• Neonatal Medicine
• Artificial Intelligence

Background:

• Lung Ultrasound (LUS) is increasingly used for neonatal lung disease diagnosis due to its accessibility and lack of radiation.
• Expert interpretation of LUS images requires specialized training, posing challenges in rural and remote areas.
• An automated diagnostic aid could significantly improve access to care in underserved communities.

Purpose of the Study:

• To develop and evaluate a computer-aided screening method for classifying neonatal LUS images.
• To quantify recurrent image features for improved diagnostic accuracy.
• To assist clinicians in the first-level diagnosis of common neonatal lung conditions.

Main Methods:

• A novel feature extraction method using virtual scanlines from LUS images.
• Conversion of scanlines into signals for analysis.
• Application of Recurrence Quantification Analysis (RQA) to extract features.
• Classification using a linear classifier with and without clinical data.

Main Results:

• The proposed method achieved 69.3% classification accuracy using only LUS image features.
• Accuracy increased to 77.6% when incorporating clinical features.
• The system successfully classified LUS images into six common neonatal lung conditions.

Conclusions:

• Automated analysis of LUS images using RQA is a viable approach for neonatal lung disease screening.
• This technology can serve as a valuable first-level diagnostic tool, particularly in resource-limited settings.
• Further development could enhance diagnostic capabilities and accessibility in remote medical communities.