Research On Key Techniques Of Pelvic Nuclear Medicine Imaging Lesion Depth Segmentation

SPECT bone imaging captures the abnormal structure,morphology and functional state of bone tissue through the metabolic information of radiolabeled drugs in the human body to achieve the purpose of disease detection.It is a common disease diagnosis method in the field of nuclear medicine and has been widely used in malignant bone metastases diagnosis and efficacy evaluation.In view of the fact that the traditional manual reading method is easily affected by personal experience and cumbersome and time-consuming clinical diagnosis,the accurate segmentation of the lesion area of SPECT bone imaging has broad application prospects in the field of nuclear medicine.This thesis mainly studies the depth segmentation method of bone metastases in the pelvic region.This thesis has carried out the following four research work.(1)Dataset construction.In order to focus on the detection of lesions in the pelvic region and reduce the consumption of computing resources,the pelvic region was segmented from the whole-body bone scan image by the polynomial kernel function fitting method;In order to avoid the influence of overfitting caused by the small number of nuclear medicine image samples on the depth segmentation method,the traditional geometric transformation method is used to realize the data expansion;In view of the low resolution of SPECT images and the blurred lesion boundary,the view fusion of the front and rear body position images is used to enhance the characteristics of the lesion area,thereby improving the recognition ability of the segmentation model for the lesion area.Through the image processing of the above methods,an experimental dataset suitable for the deep segmentation model is constructed.(2)Segmentation of bone metastases in pelvic nuclear medicine based on fully convolutional neural networks.A segmentation method based on the FCN model was constructed,and the cross-entropy loss function was introduced as the loss function to improve the feature extraction ability of the model and the accuracy of lesion segmentation,thereby constructing a deep segmentation model of bone metastases.Finally,the bone metastases are segmented based on the experimental data of pelvic nuclear medicine,and the differences in feature extraction of different VGG networks are compared.The experimental results show that the IoU index of the segmentation model with VGG16 as the feature extraction network reaches 0.485.(3)Segmentation of bone metastases based on improved U-Net network.This thesis further proposes an automatic segmentation model Att_R2_U-Net that integrates attention mechanism and cyclic residual convolution.This model introduces cyclic residual convolution blocks instead of ordinary convolution layers,which is beneficial to the model to extract lower-level features and perform stronger and better feature representation;add Attention Gate in the network decoder stage.Then,on the pelvic nuclear medicine experimental data,the lesions were segmented.The Att_R2_U-Net network improved the segmentation results of pelvic bone metastases,and the IoU index reached 0.602,which was 6%higher than the U-Net model.(4)Segmentation of bone metastases based on dense convolutional networks.Due to the low quality of nuclear medicine images and the uneven brightness of the lesion area in the image,this thesis uses the view fusion operation to add pixel values point by point to the anterior and posterior images of the patient’s pelvis area to enhance the patient’s lesion area features.This thesis further proposes a lesion segmentation model incorporating a hybrid attention mechanism,which enables the model to automatically focus on the lesion area with salient features during the training process,thereby improving the segmentation accuracy of the model.Through two sets of comparative experiments,the improvement of the lesion segmentation effect of the view fusion and the improved model is verified.The experimental results show that the improved model fused with the hybrid attention mechanism has the best segmentation effect on the view fusion dataset,and the IoU index value reaches 0.614.