LUNG SEGMENTATION ON HRCT IMAGES USING DEEP LEARNING

BIO-MEDICAL

Humans possess a natural ability to swiftly recognize images and the objects they represent. Radiologists are skilled at analysing chest Computed Tomography (CT) scans, identifying various anatomical structures to diagnose illnesses. However, automated CT slices analysis systems can significantly improve the precision of diagnosis. Despite automation, CT image segmentation, which involves extracting objectives and analysing structures, remains challenging. Many techniques, such as thresholding, region-based, edge-detection, atlas-based, soft computing, clustering, and classifier-based methods, are available in the literature, but most are unreliable due to insufficient supervision. Thresholding plays a crucial role in CT image segmentation. This study's goal is to find an automated, precise, and quick deep learning segmentation method for the parenchyma utilising a limited dataset of highresolution computed tomography images from patients with idiopathic pulmonary fibrosis. This approach intends to improve radiomics investigations carried out by healthcare professionals, where operator-independent segmentation techniques are required to locate the target and develop texture-based prediction models. The research employed U-Net, a deep learning network that has been successful in numerous biological picture segmentation tasks. Just 32 of the 42 studies that made up the dataset's patient population with lung illnesses were employed in the training stage. In terms of resource needs and comparability of segmentation results to the gold standard, the study assessed the performance of two models. The findings shown that U-Net can segment the lung area with accuracy (dice similarity coefficient = 95.72%), speed (21.32 s), and clinical acceptability. The findings of the study showed that deep learning models may be effectively used to segment and measure the parenchyma of patients with pulmonary fibrosis, generating results that are user-independent and free from radiologist oversight.

This work is licensed under a Creative Commons Attribution-ShareAlike 4.0 International License.