Quantification Of Photodynamic Skinfold Vascular Damage By New Image Segmentation Method

Vascular-targeted Photodynamic Therapy(V-PDT),which achieve targeted therapy with destroying lesion vessels,has been widely used for treatment of malignant tumor such as prostate cancer and benign vascular diseases such as senile macular degeneration and port wine stains.Vascular biological response is one of V-PDT quantification assessment methods,and its vascular biometric parameters include vascular area,vascular diameter,vascular density,etc.This dissertation presents a new method for quantifying vascular area constriction,and proposes a vascular segmentation framework based on deep learning used in nude mice dorsal skinfold chamber models.The proposed method can effectively extract vascular characteristic during V-PDT,and quantify vascular biological response based on vascular area constriction.The main research contents and innovation points in this dissertation are presented as follows:(1)A new method of circular region segmentation based on Unet and Hough transformation was proposed for identifying nude mice dorsal skinfold chamber area.This strong robustness method depends on less adjustment parameters,and can improve the efficient of skinfold vascular segmentation.(2)A new method was used for labeling automatically blood vessel.Supervised deep learning for image segmentation usually needs large number of labeled samples.In proposed method,Blood vessels were classified according to vascular edge gradient values.Multi-scale dynamic threshold segmentation method was used to label strong edge gradient vessels.Based on transfer learning,open datasets of retinal images were used to characterize weak edge gradient vessels.The results presented that the segmentation accuracy,sensitivity and specificity using automatic labeled blood vessel to train Unet model are 96.08%,84.28% and 96.58%,respectively.(3)A new method of quantification vasoconstriction without manual intervention was proposed.This method required vascular segmenting and registering in pre and post-V-PDT images.Then vasoconstriction map was labeled by using search method instead of image subtraction.Meanwhile,vasoconstriction was defined as ratio between vasoconstriction pixels and initial vascular pixels.The results showed that a margin of average vasoconstriction error for control group was 3.3% with this method.(4)The GA-Xnet method was presented for extracting the blood vessel in nude mice skinfold chamber.GA-Xnet with little manual labeled training samples can achieve95.86% accuracy and 87.59% sensitivity of image segmentation.In addition,GA-Xnet can also restore the skinfold chamber vessels under tissue membrane,and thus improves efficiency of vascular skeleton segmentation.GA-Xnet provides a theoretical basis which used vascular diameter to quantificationally evaluate V-PDT.