Classification And Lesion Segmentation Of CT Images Of Nontuberculous Mycobacterial Lung Disease And Tuberculosis

The prevalence and incidence of non-tuberculous mycobacterium lung disease(NTM-LD)are increasing year by year world wide.Because its clinical manifestations are very similar to those of pulmonary tuberculosis lung disease(PTB-LD),it is prone to be misdiagnosed and lead to delays in treatment.On the other hand,the long time of strain testing,which is the gold standard for diagnosing NTM-LD and PTB-LD,increases the risk of overtreatment in patients.According to the clinical needs,this research aims to develop deep learning algorthims to read CT images of the lungs and help improve the early diagnosis of the two diseases.The main research work of this project includes the following two aspects:(1)Classification of NTM-LD/PTB-LD based on lung CT images.Due to diverse lesion types,sizes,and locations of the two diseases in CT images,it is difficult to label lesions and bring great challenges to the classification algorithm.Based on CT image data with partial lesion annotation,this research designed a multi-task deep learning model NL-MIL that can support both disease classification and lesion prediction.For the disease classification branch,this work adopts a Multiple Instance Learning(MIL)framework to cope with the challenges of a wide variety of lesions and large size variation and introduces a non-local attention module to guide the aggregators in MIL to pay attention to the global feature distribution.The lesion prediction branch is used to predict whether each instance contains lesions.This work further adopts selfsupervised learning to supervise the activation of the attention module in the aggregator based on the results of lesion prediction.(2)Lesion segmentation based on lung CT images.Lesions of NTM-LD/PTB-LD in lung CT images involve multiple types.Different from binary segmentation,the lesion segmentation problem dealt with in this work needs to predict multiple lesion categories at the same time.In order to introduce the continuous constraint of lesions in spatial distribution and suppress noise on the segmentation results,this work proposes a Med Seg GCN network,which splits the lesion segmentation problem into two stages: binary lesion area segmentation and lesion classification.The first stage uses U-Net to predict whether voxels belong to lesion areas.In the second stage,through the newly designed spatial GCN module,the predicted lesion area is transformed into an undirected graph,the nodes of the graph are supervoxels,and the connections between nodes are the adjacent relationships of supervoxels in space and features,and the classification of nodes is done by graph convolutional neural networks.In cooperation with Haihe Hospital of Tianjin University,CT image datasets of NTM-LD and PTB-LD were established,including 1103 cases.Comparative experiments with the existing SOTA methods show that the proposed method can achieve better disease classification effects,more accurate multi-label lesion segmentation results,and has good clinical application potential.