Attention-Unet: A Deep Learning Approach for Fast and Accurate Segmentation in Medical Imaging

Abstract

Accurate extraction of the bronchial tubes from lung computed tomography (CT) images is crucial for evaluating respiratory function and diagnosing diseases. Current bronchial segmentation methods often rely heavily on substantial human-computer interaction to improve segmentation accuracy. Although deep learning has been widely applied in medical image processing, especially in lung nodule detection and diagnosis of malignancy, its use in bronchial segmentation in lung CT images faces challenges such as image noise and partial volume effects, which lead to segmentation leakage and difficulties in identifying small bronchi. Additionally, original lung CT images contain non-relevant regions like bones and the patient bed, which increase data processing time and risk errors. By leveraging the anatomical structure of the bronchial tree, we propose a stepwise approach to bronchial segmentation and introduce an Attention-Unet-based method. Experimental results demonstrate that applying the deep learning-based Attention-Unet network to bronchial segmentation in lung CT images enhances segmentation speed and accuracy while effectively preventing leakage.