Deep Learning for Accurate B-Line Detection and Localization in Lung Ultrasound Imaging

Nixson Okila¹Andrew Katumba^1*Joyce Nakatumba-Nabende¹Cosmas Mwikirize¹Sudi Murindanyi¹Jonathan Serugunda¹Samuel Bugeza¹Anthony Oriekot²Juliet Bossa²Eva Nabawanuka^2*

• ¹Makerere University, Kampala, Uganda
• ²Mulago Hospital, Kampala, Uganda

Lung ultrasound (LUS) has become an essential imaging modality for assessing various pulmonary conditions, including the presence of B-line artifacts. These artifacts are commonly associated with conditions such as increased extravascular lung water, decompensated heart failure, dialysis-related chronic kidney disease, interstitial lung disease, and COVID-19 pneumonia.Accurate detection of the B-line in LUS images is essential for effective diagnosis and treatment.However, interpreting LUS is often subject to observer variability, requiring significant expertise and posing challenges in resource-limited settings with few trained professionals. To address these limitations, deep learning models have been developed for automated B-line detection and localization. However, existing models do not precisely capture the artifacts' shapes and often fail to reveal their intricate textures and patterns. In addition, none has been integrated into mobile LUS screening tools for use in resource-constrained environments. This study introduces YOLOv5-PBB and YOLOv8-PBB, two modified models based on YOLOv5 and YOLOv8, respectively, designed for precise and interpretable B-line localization using polygonal bounding boxes (PBBs). YOLOv5-PBB was enhanced by modifying the detection head, loss function, nonmaximum suppression, and data loader to enable PBB localization. Meanwhile, YOLOv8-PBB was customized to convert segmentation masks into polygonal representations, displaying only boundaries while removing the masks. Additionally, we incorporated an image preprocessing technique into the models to enhance LUS image quality. Moreover, we integrated the smaller, faster inference model into a mobile LUS screening tool, making it accessible for resource-limited settings. The models were trained on a diverse dataset from a publicly available repository and Ugandan health facilities. Experimental results showed that YOLOv8-PBB achieved the highest precision (0.947), recall (0.926), and mean average precision (0.957). YOLOv5-PBB, while slightly lower in performance (precision: 0.931, recall: 0.918, mAP: 0.936), had advantages in model size (14 MB vs. 21 MB) and average inference time (33.1 ms vs. 47.7 ms), making it more suitable for real-time applications in low-resource settings.