Learning-based Pulmonary Disease Detection Using Weak Labels for Volume Sweep Lung Ultrasound Imaging.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Pulmonary Medicine

Background:

• Pulmonary disease causes significant global morbidity and mortality.
• Accessible medical imaging for lung conditions is limited worldwide.
• Lung ultrasound (LUS) is accurate but requires trained sonographers; Volume Sweep Imaging (VSI) simplifies acquisition but needs expert interpretation.

Purpose of the Study:

• To develop a novel learning-based approach for automated interpretation of lung ultrasound VSI.
• To reduce the dependence on expert radiologists for diagnosing pulmonary abnormalities from VSI scans.
• To investigate the efficacy of deep learning architectures with weak labels for VSI analysis.

Main Methods:

• Ultrasound B-mode cine loops from VSI exams of the posterior lungs were analyzed.
• Two hybrid Convolutional Neural Network (CNN)-based deep learning architectures were developed and compared.
• A combined CNN and random forest classifier utilized weak labels (entire cine loop diagnosis) for training.

Main Results:

• The best performing model, a combined CNN and random forest classifier, achieved 91.67% accuracy.
• Precision was 92.31%, recall was 85.71%, and F1-score was 88.89%.
• The approach demonstrated robustness to extraneous information within the ultrasound sweep.

Conclusions:

• Automated interpretation of lung ultrasound VSI using learning-based approaches shows significant potential.
• This method can enhance the accessibility and efficiency of diagnosing pulmonary conditions.
• The findings support the clinical relevance of AI in medical imaging interpretation for lung diseases.