| The curvilinear structure in an image refers to a collection of line-like structures that are criss-crossed and interconnected in a two-dimensional image,and widely exist in the fields of medicine,remote sensing,microscopy,etc.Clinical medicine has shown that curvilinear structures such as retinal blood vessels and corneal nerves in medical images are often closely related to diseases such as age-related macular degeneration,diabetes,glaucoma,hypertension,arteriosclerosis,and multiple sclerosis.In addition,retinal blood vessel segmentation is also the basis for subsequent classification of arteriovenous blood vessels,which plays a great auxiliary role in the diagnosis of ophthalmic diseases.There are two major difficulties in the study of curvilinear structures: 1)the detection of curvilinear structures;2)the reconstruction of fracture curvilinear structures.The complete curvilinear structures provide more accurate auxiliary information for clinical treatment,and its connectivity is crucial for establishing the topology of the curve structure,arteriovenous classification,etc.Combining the excellent feature extraction capabilities of neural networks with the combined advantages of group tendency and individual randomness of random walk algorithms,this paper proposes a method for detecting curvilinear structures and reconnecting fracture structures in medical images.The main contributions of this paper are as follow:1)This paper proposes a curvilinear structure segmentation method based on dense dilated convolutional neural networks.Utilizing the excellent feature extraction capabilities of neural networks,a curvilinear structure segmentation network based on convolutional neural networks is designed to segment curvilinear structures such as retinal blood vessels in fundus.Compared with the traditional method,the proposed method avoids a lot of manual design feature extraction parameters,and adopts a learning-based method to let the network autonomously extract features.The proposed method can fully extract the curvilinear structures in medical images and the proposed method reached a higher accuracy compared to the state-of-the-art methods.2)This paper proposed a probability-regularized walk algorithm by taking full advantage of the feature extraction of neural network.Inspired by the partial gravitational random walk algorithm,the constraints of the walkers are strengthened,and the walkers are restored to more detailed information during the fracture reconstruction process.The probability of each pixel is then integrated into the random walk algorithm,which is adopted to constrain the movement of the walker.Taking full advantage of neural network to extract feature advantages and group advantages of random walk algorithm,a probabilistic regularized walk algorithm is proposed to reconstruct the fracture curve structure.The neural network needs to classify each pixel in the image when segmenting the curvilinear structures and output the probability value of each pixel being judged as a curve structure.The probabilities are integrated into the random walk algorithm,and the probability of the pixel itself is used to constrain the movement path of the walker.The proposed algorithm breaks the uncertainty of the walker’s movement,thereby reconstructing the fracture region more accurately.3)A neural network segmentation algorithm based on attention mechanism is proposed.In order to add more attention weights to curvilinear structures in ophthalmic images with low signal-to-noise ratio,this paper proposes a segmentation algorithm based on the attention mechanism,which makes the segmentation model tend to extract curvilinear structures in training and prediction.The learned feature contains more attention weights of curvilinear structures.Compared with the dense dilated convolutional neural network proposed in this paper,this method has more segmentation types of images,wider modalities,and better segmentation and connectivity of the curvilinear structure. |
PDF Full Text Request