Parse Challenge 2022: Pulmonary Arteries Segmentation Using Swin U-net Transformer(Swin UNETR) and U-net

In this paper, we describe a deep neural network architecture based on Swin UNETR and U-Net for segmenting the pulmonary arteries from CT scans. The final segmentation masks were created using an ensemble of six models, three based on Swin UNETR and three based on 3D U-net with residual units. Using this strategy, our group scored 84.36 % on the multi-level dice. We conducted additional investigation and separated the task into three major subtasks: Task 1: Use the default hyperparameters for plain UNET segmentation and experiment with the patch size, a key hyperparameter for UNET segmentation models. Task 2 : Develop a lung segmentation model that distinguishes between the major pulmonary artery and the branches in order to precisely assess the model’s performance. Task 3 : Examining the mask by extracting small patches near the branches and large patches around the major pulmonary artery.The code of our work is available on the following link: https://github.com/akansh12/parse2022

1. INTRODUCTION

An important biomarker for predicting and diagnosing hypertension is the modified main pulmonary arteries(PA), which is typically found in patients with pulmonary hypertension (PH) and has a diameter that is significantly bigger than that of a normal person. A broadening of the PA in chronic obstructive pulmonary disease (COPD) is linked to higher exacerbation risk and lower survival rates. Blockage of one of the pulmonary arteries, mostly caused by blood clots, causes Pulmonary Embolism (PE). Therefore the early diagnosis of these pulmonary diseases, assessing the risk, and planning treatment for the patients at the early stage is required. Recently, advances in cardiac imaging have been accepted as good tools to help clinicians with early diagnosis and advanced planning in surgery. For the evaluation of many Pulmonary Vascular Diseases(PVD), the morphological examination of the Pulmonary Artery (PA) is crucial. However, even for specialists, the diagnosis procedure frequently takes a long time due to the enormous size and complexity of various imaging modalities; the pulmonary arteries must therefore be correctly and effectively segmented out. Angiography offers insights into the blood flow and conditions of the vascular tree. Three-dimensional volumetric angiography information can be obtained using magnetic resonance (MRA), ultrasound, or x-ray-based technologies like computed tomography (CT). Currently, it is standard procedure to evaluate coronary and pulmonary artery diseases with computed tomography angiography (CTA) as it provides high-resolution 3D imaging as non-invasiveness. Various undermentioned reasons account for the difficulty of the pulmonary artery segmentation: