Study On Chest Medical Image Segmentation Algorithm Based On Deep Learning

Medical images have different imaging methods,there are many organs involved in medical images and the background is very complicated,these factors have brought great challenges to the research of medical image segmentation algorithms.In this thesis,for lung computed tomography(CT)image,an automatic lung lobe segmentation algorithm based on a fully convolutional network is proposed.For breast dynamic contrast-enhanced magnetic resonance image(DCE-MRI),a semi-automatic segmentation algorithm based on graph cuts with automatic suspected lesion labelling and a fully-automatic convolution recurrent network based on spatio-temporal features for breast lesions are proposed.The proposed segmentation methods for two imaging modality are described as follows:CT is a common method for examining lung diseases.Accurate segmentation of the lung lobes helps in the diagnosis of chronic obstructive pulmonary disease and the planning of pulmonary nodule resection.Lung fissure is the boundary between different lobes and is an important tissue in the lung.However,because the lung fissure may be incomplete in CT and the contrast of the lung fissure is very low,it is difficult to segment the lung lobe directly based on the fissure.Therefore,this thesis proposes an automatic lung lobe segmentation algorithm based on the fully convolutional network.Because the position information is very important for lung lobe segmentation,so this algorithm introduces the global position information in the network.In order to improve the accuracy of the segmentation of the lung lobe boundary,a lung fissure aware module was added during the training process.At the same time,according to the fact that the lung fissure is a three-dimensional plane,the lung fissure was enhanced based on the Hessian matrix,which improves the lobe segmentation results.The proposed segmentation algorithm achieves satisfactory performance in a public dataset.Breast DCE-MRI is a highly sensitive imaging diagnostic method in breast disease diagnosis.The accurate segmentation of breast lesions can help the precise calculation of the medical characteristics of the lesions,which can help to diagnose benign and malignant lesions.For the region given by the doctor,the semi-automatic segmentation method is proposed in this thesis,the algorithm first gives the suspected lesion region,and then models the breast lesion based on the suspected lesion to segment the lesion accurately.In addition,based on that the dynamic enhancement sequence in breast DCE-MRI is very important for judging breast lesions,this thesis proposes a fully-automatic breast lesion segmentation algorithm based on spatio-temporal features extraction with convolutional recurrent network,which can analyse the spatio-temporal feature in breast DCE-MRI.The algorithm can segment breast lesions automatically without manually selecting the region containing the lesions.The algorithm can produce accurate segmentation results.