Development of Deep Learning pipeline for Airways Segmentation in Human lungs

Abstract

Chronic respiratory diseases cause abnormality in the lungs, especially the airways, and are the reason for mortality around the world. These problems are more common in low and middle-income countries where resources for prevention, diagnosis, treatment, and management are scarce. Most of these diseases are due to damage in airways anatomy, including the widening of airways (bronchitis), and narrowing of airways (bronchiectasis) due to tumor or mucus production which in critical situations results in shortness of breath and ultimately death. Other disorders include bunches around airways or cavities inside the lung lumen due to air trapped inside. Therefore, early disease detection and screening are critical steps to control the situation. For diagnosis, airway anatomy and quantitative analysis are performed using a variety of approaches including manual, semi-automated, and fully automated methods. Manual methods are time-consuming and error-prone due to differences in experience performed by ra diologists using CT scans or radiographs. Semi-automated procedures are carried out with the assistance of radiologists, professionals, and computers. Automated airway segmentation is performed using digital technology without the aid of experts. In this study, we proposed a methodology for 2D airway segmentation to make it clinically applicable. The dataset used for the proposed methodology contains CT scans of lungs in NIfTI format (converted to DICOM format by 3D Slicer). The pipeline consists of shallow layers of deep learning architecture “U-Net” which is widely used in medical image segmentation tasks. Various pre-processing steps are applied to the CT scan DICOM images which include extraction of pixel information from DICOM formats, resizing, filtering, windowing (enhancement), thresholding, and lung segmentations. The results were analyzed after several hyper-parameters tunning techniques with bet xv List of Tables ter segmentation obtained after extraction of the lung portion by using the largest connected component analysis technique which discarded irrelevant portions from CT images. The evaluation of all the trained models was performed and the highest scores were obtained from the U-Net model trained on segmented lungs. The performance measures with the Dice coefficient, Jaccard coefficient, and validation loss obtained the values of 93.28%, 93.51%, and 0.010% respectively. The proposed methodology acquired results compared to previous works that used using deep learning approach with less time and memory consumption. The resulting pipeline can be used in hospi tals for accurate segmentation of airways, diagnosis of pulmonary respiratory diseases, and their severity.