Research On Accurate Segmentation Method Of Sensitive Tuberculosis Lesions For CT Images

According to the World Health Organization(WHO),tuberculosis(TB)remains one of the greatest threats to human health.In 2020,there were over 8 million new cases of tuberculosis reported worldwide,with an average incidence rate of 130 cases per100,000 people.In China,tuberculosis is a major infectious disease with high morbidity and mortality rates.The large number of TB patients and infected individuals,who primarily transmit the disease through the respiratory tract,poses a significant risk for epidemic spread,resulting in severe consequences for both society and public health security.The incidence of tuberculosis has been increasing in recent years.Therefore,the accurate diagnosis,effective treatment,and advanced medical intelligence in tuberculosis management play a critical role in the diagnosis,prevention,and control of infectious diseases.Among the various types of pulmonary tuberculosis,sensitive pulmonary tuberculosis constitutes a significant proportion.In this study,accurate segmentation of sensitive pulmonary TB lesions is achieved by combining image processing,graph theory,and deep learning techniques on lung CT images.The primary research contributions are outlined as follows:Firstly,a two-stage segmentation method is proposed for the segmentation of lung parenchyma and tuberculosis lesions,particularly focusing on isolated tuberculosis lesions.In the lung parenchyma segmentation stage,a combined digital image processing algorithm is employed.This algorithm integrates threshold segmentation,region growing,and level set segmentation techniques,incorporating morphological operations,connected domain analysis,and flood filling.In the tuberculosis segmentation stage,a combination of threshold segmentation and the U-net deep learning model is utilized for the accurate segmentation of tuberculosis lesions.Through the application of multiple segmentation techniques,the segmentation of isolated tuberculosis lesions is successfully achieved with high accuracy.Secondly,we propose a modified convex hull algorithm combined with nonuniform rational B-spline curves to address the challenge of accurately segmenting pleural adhesion-type tuberculosis(TB)lesions that lack clear borders.The proposed approach consists of two main stages: lung parenchyma contour repair and tuberculosis lesion segmentation.In the lung parenchyma repair stage,we accurately localize the lesion by utilizing a convex hull point distance threshold and analyzing the frequency of edge slope changes to capture the complexity of the lesion’s boundary.To repair the defective lung parenchyma contour,we fit the edges that exhibit high similarity to the lung contour using non-uniform rational B-spline curves.This process effectively improves the accuracy of pleural adhesion-type TB lesion segmentation,resulting in an average intersection over union(IOU)of 97.3% for lung parenchyma segmentation.Building upon the acquisition of an accurate lung parenchyma contour,we employ a deep learning model for the precise segmentation of pleural adhesion-type TB lesions.By leveraging the rich features learned by the deep learning model,we achieve a significant improvement in segmentation accuracy,with an average Dice coefficient of90.17%.Finally,to address the problems of sensitive TB lesions with blurred boundaries,cloudy flocculent shape and overlapping with blood vessels.AFF-Net,a deep learning model for adaptive feature fusion,is proposed in this work.The AFF-Net incorporates several key components to enhance feature extraction perceptibility and improve local detail supervision.These components include a multi-level perceptual field module,an AFF module,an attention gate,and an edge segmentation similarity loss function.In the early layers of down-sampling,a combination of two-channel conventional convolution and inflated convolution schemes is employed,with the inflated convolution rate decreasing progressively.The bottom AFF module facilitates adaptive learning across different scales by incorporating successive down-sampling and the previous feature layer.During up-sampling,an attention gate is introduced to enhance local information perception.Furthermore,the loss function incorporates an edge segmented similarity loss to provide additional supervision for edge information and penalize non-target regions.Experimental results demonstrate the effectiveness of the proposed AFF-Net model in segmenting four types of TB lesions.The summed average Dice coefficient,accuracy,and recall achieved by the AFF-Net are 89.45%,88.87%,and 89.83%,respectively.These results indicate a significant improvement in the accuracy of TB lesion segmentation.Lastly,we develop a sensitive tuberculosis assisted diagnosis system using the encapsulated segmentation algorithm implemented with the PyQt framework.