Multi-classification of Lung Diseases Using Lung Ultrasound Imaging

Now a days, in medical imaging, the need for safe and efficient diagnostic modalities has become increasingly crucial, particularly in emergency situations and at the patient’s bedside. While the medical community has traditionally depend on X-rays and CT scans, our research focuses on the power of Deep Learning (DL) techniques to analyze Lung Ultrasound (LUS) images. LUS offers a promising alternative for assessing pulmonary congestion, making it an ideal approach for innovative applications. Within the realm of LUS analysis, the identification of A-line and B-line artifacts holds highest importance in disease detection. However, the interpretation of LUS images has proven to be a difficult task, compounded by the challenges faced by the pandemic times. Visual quantification of these artifacts, prone to observer variability, necessitates a novel approach to enhance accuracy and efficiency. In our study, we embarked on a journey to explore and train various deep learning models, with a particular focus on automatic A-line and B-line detection and localization in LUS datasets sourced from diverse origins. Leveraging these datasets and employing a pre-trained ResNet50 model, we have achieved a significant breakthrough in automating the detection and classification of LUS images. Our approach predict disease severity scores associated with lung ailments, differentiating between A-line and B-line pathologies. While our experiments with the pre-trained model and the collected datasets have yielded satisfactory results, we acknowledge that further research is required by the expansion of our models to accommodate on larger datasets.