Deep learning for real-time multi-class segmentation of artefacts in lung ultrasound

• DOI: 10.1016/j.ultras.2024.107251
• Publication Date: 2024-01-29T17:12:19Z
• Authors: Lewis Howell; Nicola Ingram; Roger Lapham; Adam Morrell; James R. McLaughlan
• Keywords: Transfer of learning; Triage; Similarity (geometry)

Lung ultrasound (LUS) has emerged as a safe and cost-effective modality for assessing lung health, particularly during the COVID-19 pandemic. However, interpreting LUS images remains challenging due to its reliance on artefacts, leading operator variability limiting practical uptake. To address this, we propose deep learning pipeline multi-class segmentation of objects (ribs, pleural line) artefacts (A-lines, B-lines, B-line confluence) in training phantom. Lightweight models achieved mean Dice Similarity Coefficient (DSC) 0.74, requiring fewer than 500 images. Applying this method real-time, at up 33.4 frames per second inference, allows enhanced visualisation these features This could be useful providing helping skill gap. Moreover, masks obtained from model enable development explainable measures disease severity, which have potential assist triage management patients. We suggest one such semi-quantitative measure called Artefact Score, is related percentage an intercostal space occupied by B-lines turn may associated with severity number conditions. show how transfer used train small datasets clinical images, identifying pathologies simple effusions consolidation DSC values 0.48 0.32 respectively. Finally, demonstrate DL translated into practice, implementing phantom alongside portable point-of-care system, facilitating bedside assessment improving accessibility LUS.