Automatic And Accurate Labeling Of Multi-scale Blood Vessels Of Dorsal Skin-fold Window Chamber In Mice

Vascular Targeted Photodynamic Therapy(V-PDT)has been widely used in the treatment of various microvascular diseases because of its good targeting and selective properties.The biological response of target vessel dilatation and contraction during the course of treatment is an important basis for making individualized V-PDT treatment plan and evaluating the treatment effect.In the animal experiment of V-PDT,after real-time monitoring of the morphology of the blood vessels in the Dorsal skin-fold chamber(DSWC)of mice based on optical imaging technology,it is necessary to segment the blood vessels automatically using image processing algorithm,and the accuracy and sensitivity of vascular segmentation will directly affect the quantitative evaluation of vascular morphology.At present,deep learning algorithm has been widely used in vascular segmentation,but for DSWC vascular images,the segmentation of small vessels depends on the characteristic migration of retinal vessels,and the accuracy of segmentation still needs to be improved.In order to improve the precision of blood vessel segmentation and realize the real-time monitoring and quantitative evaluation of multi-scale blood vessel morphology and structure in V-PDT,in this thesis,combined with the traditional vascular segmentation methods,an automatic and accurate labeling method for vascular images of mouse DSWC model is proposed,and the vascular segmentation data set of DSWC model is constructed,and then the deep learning algorithm is used to achieve fine segmentation of multi-scale blood vessels.The main research contents and innovations of this thesis are as follows:(1)A dual-scale fusion method is proposed to segment different-scale blood vessels.The OTSU algorithm combined with morphological filtering and edge detection was used to extract the binarization map for large-scale blood vessels.The small-scale blood vessels were enhanced using gauss matched filter combined with Hessian matrix eigenvalue.Finally,the small-scale vessels were binarized by iterative filling after extracting vascular contour based on Canny operator.The obtained results were fused with the large-scale vessel binarization map.The experimental results show that this method can extract the vascular network more effectively than other methods.(2)An automatic and accurate multi-scale vascular labeling method based on image registration is proposed.Firstly,the global and local vascular images of DSWC model are obtained by changing the optical magnification of the Zoom Lens Group in Stereo microscope,then the image registration algorithm is used to map the local vascular features to the global vascular location in order to correct the small-scale vascular information in the global large field of view.The results show that the method of automatic and accurate vessel labeling is feasible and can be used to construct multi-scale vessel segmentation dataset.(3)The blood vessel image of DSWC is segmented by using the constructed data set and the deep learning algorithm.The experimental results show that the model trained by the data set can improve the precision of multi-scale vascular segmentation.It can not only segment the whole vascular network completely,but also keep the continuity of the segment for the small-scale blood vessels,and it does not need the hybrid fundus data set for transfer learning.