Research And Implementation Of CT Image Segmentation Method Applied To Lung Surgery Planning

Histological analysis of lung lesions is the main screening modality in the early stages of cancer.Its diagnosis is usually made by biopsy of the lesion using a thoracoscopic procedure.Because of the tendency to cause collapse of the lung when the tools needed for the procedure need to be placed during the procedure,resulting in a strong displacement of the lesion.Therefore,the surgeon needs to reposition the lung lesion under CBCT guidance,but the lesion is not clearly visible in CBCT,and even collapse can cause the lesion to be invisible.CT images of the lesion are clearly visible,so it is highly relevant to use image alignment to predict the lung deformation from preoperative CT images to intraoperative CBCT images and thus pinpoint the lung nodule.The first step to achieve the above-mentioned deformation simulation is accurate image segmentation,including: lung parenchyma and lobe segmentation in preoperative CT,and lung parenchyma segmentation in intraoperative CBCT images before and after lung collapse.In this paper,this thesis studies the segmentation algorithms of lung parenchyma and lung lobes in CT and intraoperative CBCT images,design two traditional segmentation algorithms according to the application requirements,and build several segmentation frameworks with U-Net as the base network.For lung parenchyma segmentation,in order to solve the problem of few and unlabeled lung CBCT data before collapse,a strategy of converting some CT data into CBCT data is proposed,and the effectiveness of this strategy for lung parenchyma segmentation in lung CT and CBCT is verified using 3D U-Net as the segmentation network model.To solve the problem of small sample size and unlabeled post-collapse lung CBCT data,the SLIC superpixel-based and GVF Snake-based CBCT lung parenchyma segmentation algorithms were proposed,and the comparison and algorithm screening were performed based on the experimental results.In lung lobe segmentation,to address the problem of low accuracy of right middle lung lobe due to small sample size and intra-class imbalance phenomenon in lung lobe segmentation in CT images,a lung lobe segmentation framework based on adversarial network(named LLASN)was designed by introducing the concept of generative adversarial network starting from the network structure,and the quantitative evaluation results verified that LLASN produced more accurate and continuous segmentation than the underlying network U-Net results.Also inspired by Grad-CAM++ and SEG-GRAD-CAM,a gradient-weighted Class Activation Mapping(named MO Grad-CAM++)for multi-objective segmentation that explains the segmentation model’s focus on different categories is proposed.To further refine the interpretation of the segmentation model,the attention of the segmentation model to the categories is quantified.To evaluate the right middle lung lobe without the involvement of real labels,a qualitative evaluation criterion(named LHS)based on MO Grad-CAM++was proposed in this paper.To address the poor performance of the right middle lung lobe in CT images,a heat map loss function is proposed starting from the loss function,and the effectiveness of combining three commonly used loss functions is further verified in 2D U-Net and 3D U-Net.The experimental results show that the performance of 3D U-Net is due to 2D U-Net for the same 3D data with the participation of the heat map loss function.