Research On Lesion Segmentation Of Thoracic Nuclear Medicine Images For Quantitative Assessment Of Diseases

SPECT whole body bone imaging is the most common disease diagnosis method in the field of nuclear medicine.SPECT bone imaging is also an important part of the diagnosis and evaluation of various types of tumors.potential.After considerable development,deep learning has been widely used in the automatic segmentation of medical images because of its ability to automatically learn data features from a large amount of data,showing great potential in intelligent medical care.This thesis mainly studies SPECT whole body bone imaging images,and studies and constructs a fully supervised deep learning lesion segmentation model and a semi-supervised deep learning lesion segmentation model.The main research contents and research work of this paper are as follows:(1)A method for data preprocessing for SPECT imaging is proposed.First,the thoracic region was segmented from the SPECT whole-body bone scintigraphy image and normalized;then,the geometric transformation expansion method and the adversarial generative network model were applied to expand the data,and three categories of thyroid,bone metastases,and arthritis were constructed respectively.Data set;finally,the commonly used evaluation indicators of medical image segmentation are defined to evaluate the model performance.(2)A deep learning-based lesion-supervised segmentation model for SPECT images of bone metastases is proposed.Considering the characteristics of large number,wide distribution and diverse shapes of bone metastases in the thoracic region,a lesion segmentation model based on improved Mask R-CNN and improved U-Net was constructed.Specifically,the residual structure is used to deepen the network depth,while an attention mechanism is introduced to enhance the feature space region information,and a set of real clinical diagnostic data is selected for experimental verification.The experimental evaluation shows that for SPECT bone metastases,the improved U-Net model can effectively segment the bone metastases than the improved Mask R-CNN model.(3)A deep learning-based semi-supervised segmentation model for lesions in SPECT images is proposed.Due to the difficulty of nuclear medicine image labeling,especially for SPECT bone imaging,which is a typical ultra-low resolution and largescale imaging,large-scale SPECT data labeling is time-consuming and laborious.segmentation model.Specifically,first of all,the model introduces the hole convolution residual module and the Inception module to replace the ordinary convolution layer,which can learn the information of different scales of the image while deepening the network depth,and introduces the famous Chan and Vese models as models.The unsupervised loss of image segmentation of;then,lesion segmentation is performed on a set of real SPECT bone scan images,and the semi-supervised segmentation model has Dice,CPA and Io U values of 0.683,0.715,and 0.601 on bone metastases lesion segmentation,respectively.Experimental results demonstrate the feasibility of the semi-supervised model proposed in this paper to segment a large number of unlabeled bone metastases.(4)Quantitative evaluation was performed according to the lesion results obtained by segmentation.Specifically,first,the lesion information of bone metastases is extracted,and the region of interest in the thoracic cavity is extracted at the same time,and the skeletal region of the thoracic cavity is segmented by an unsupervised model;Distinguish mildly concentrated,moderately concentrated,and heavily concentrated lesion areas from the histogram of the lesion area,and count the number of pixel values in each lesion area;finally,calculate the size ratio of each lesion area according to the number of pixels in different areas,so as to determine the lesion area severity.