Anatomical Feature-Based Lung Ultrasound Image Quality Assessment Using Deep Convolutional Neural Network

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Pulmonary Medicine

Background:

• Lung ultrasound (LUS) is a valuable point-of-care tool for diagnosing respiratory conditions like COVID-19 due to its safety and portability.
• Current LUS image assessment is operator-dependent, leading to variability in image quality and interpretation.
• Existing algorithmic methods for image quality lack clinical context, focusing solely on pixel data.

Purpose of the Study:

• To develop a novel method for objectively quantifying the clinical context and diagnostic quality of LUS images.
• To introduce an Anatomical Feature-based Confidence (AFC) Map that incorporates anatomical feature visibility.
• To address the limitations of existing image quality assessment methods in LUS.

Main Methods:

• A deep convolutional neural network, specifically two U-net models, was developed to segment crucial anatomical features in LUS images.
• The models identify and quantify 'Bright Features' (pleural and rib lines) and 'Dark Features' (rib shadows).
• The output confidence values from the segmentation models form the AFC Map, which is then correlated with image quality scores.

Main Results:

• The feature segmentation models achieved an average Dice score of 0.72, indicating effective identification of anatomical structures.
• A correlation was observed between the calculated confidence values in the AFC Map and subjective image quality scores.
• The AFC Map demonstrated its relevance in quantifying the clinical context and diagnostic utility of LUS images.

Conclusions:

• The proposed Anatomical Feature-based Confidence (AFC) Map offers an objective approach to assess LUS image quality by considering clinical context.
• This method has the potential to standardize LUS image interpretation and improve diagnostic accuracy.
• Further development and validation of AFC Maps can enhance the reliability of AI-assisted LUS analysis.