Automatic Segmentation Based On Fuzzy Connectedness And Deep Learning For Thyroid Ultrasound Images

The incidence rate of thyroid related diseases in the domestic population is increasing year by year.The computer-aided diagnosis of thyroid diseases will greatly reduce the workload of doctors.Ultrasound imaging technology has become the most popular way to diagnose the thyroid diseases due to its real-time,low price,non-invasive and non-radioactive characteristics.However,the inherent characteristics of ultrasound image,such as speckle noise,signal loss and low contrast,increase the difficulty of thyroid region segmentation for ultrasound images.In order to solve the above problems,this thesis focuses on the research of fuzzy connectedness segmentation model and deep learning model.Fuzzy connectedness model has been widely used in the segmentation of magnetic resonance imaging(MRI)and computed tomography(CT).However,it is rarely used in ultrasound images.Through experiments,we find that the existing theory of fuzzy connectedness is not enough to deal with fuzzy boundary between adjacent tissues in ultrasound images,and the membership function of fuzzy affinity in the original model is not enough to measure the spatial structure similarity among pixels.In order to solve this problem effectively,the definition of fuzzy affinity is extended by combining the local entropy of the image,and a segmentation model of fuzzy connectedness based on the local information entropy is proposed.Through comparative experiments,we verify the effectiveness of the fuzzy affinity membership function which bases on local entropy.It is also proved that the segmentation result of the proposed model is better than that of the original model in thyroid ultrasound image.The fuzzy connectedness model needs to initialize the sequence of seed points manually,and the choice of the location of seed points has a great influence on whether the desired segmentation results can be obtained.In order to achieve automatic segmentation under this model,we proposed an automatic fuzzy connectedness segmentation model based on deep learning,taking advantage of which the deep learning model can provide more accurate presegmentation results than other segmentation models.Through experiments,firstly we find that the segmentation result of the automatic segmentation model is better than that of the nonautomatic segmentation model.Secondly,using the deep neural network with better structure can improve the segmentation effect of the whole automatic model to some extent.Thirdly,when the expected segmentation results are the same,the training data used by the automatic fuzzy connectedness model is significantly lower than that of the deep learning model,which has certain practical value in the field of medical image segmentation where the cost of obtaining labeled data is high.Loss function is very important to train an effective deep neural network model.Recently,the most commonly used loss function in training medical image segmentation model is the per-pixel loss function such as cross entropy loss.It is simpler and more effective to improve the segmentation result of the whole model by improving the loss function than to design a new network structure.In this thesis,the concept of relative fuzzy connectedness is used to add the relative position of thyroid and trachea in ultrasound images as a prior knowledge into the loss,and a deep learning segmentation model based on the relative fuzzy connectivity loss function is designed.The training model based on the relative fuzzy connectedness loss function can accelerate the convergence rate of the model and is better than the training model based on cross entropy loss function in segmentation result.Compared with the automatic fuzzy connectedness segmentation model,it is an end-to-end segmentation model.