Automatic deep learning-based pleural effusion segmentation in lung ultrasound images

Area of Science:

• Medical imaging
• Artificial intelligence in medicine
• Pulmonology

Background:

• Point-of-care lung ultrasound (LUS) aids in diagnosing pleural effusion (PE) but requires clinical expertise for accurate interpretation.
• A prior deep learning model achieved over 90% accuracy in detecting PE on LUS.

Purpose of the Study:

• To develop a deep learning model for segmenting PE in LUS images.
• To evaluate the performance of the segmentation model against expert interpretations.

Main Methods:

• A dataset of 3,041 LUS images from 24 patients with PE was used.
• Two experts provided ground truth segmentations for training and validation.
• The algorithm was trained using ten-fold cross-validation.

Main Results:

• The deep learning model achieved an average Dice Similarity Coefficient (DSC) of 0.70 for PE segmentation.
• Expert-to-expert segmentation agreement showed an average DSC of 0.61.

Conclusions:

• The developed algorithm demonstrates comparable performance to human experts in segmenting PE on LUS.
• This represents a significant advancement towards an AI tool for augmenting PE diagnosis and treatment.