Research On Recognition And Analysis Method Of Lymphoma Microscopic Hyperspectral Image Based On Deep Learning

In recent years,the incidence and mortality of lymphoma have gradually increased,which has seriously affected the physical and mental health of human beings.Because the early symptoms of lymphoma are not obvious,patients often have missed the best time for treatment when they are diagnosed.At present,the "gold standard" for the diagnosis of lymphoma is still histopathological examination.This process requires experienced doctors to operate under a microscope,which is time-consuming,laborious and is easily affected by the doctor’s subjective judgment.Therefore,the pathological diagnosis method that combines image processing technology with traditional color pathological images has attracted wide attention.However,the amount of information in traditional color images is limited,which limits the accuracy of the diagnosis results.Compared with traditional color images,hyperspectral images contain rich spatial and spectral information,which provides a new solution for the task of identifying and segmenting lymphoma.Based on this,this thesis applies the microscopic hyperspectral imaging technology to the identification and analysis of lymphoma,and studies the mosaic and segmentation algorithms of lymphoma microscopic hyperspectral images.Aiming at the stitching of lymphoma microscopic hyperspectral images,this thesis proposes a SURF algorithm based on PCA and a priori information.The method first uses PCA to realize the band selection of hyperspectral images,then uses the SURF algorithm combined with the prior information in the acquisition system to realize the detection,description and matching of feature points,and finally uses the linear weighted image fusion strategy to realize the post-splicing microscopic hyperspectral smooth transition of brightness in the image.Aiming at the identification and segmentation of cancerous regions in lymph node tissues,in order to make full use of the spatial and spectral information in the microscopic hyperspectral image,this thesis proposes a 3D SE-U-Net model.In this model,the spatial and spectral features of the lymphoma microscopic hyperspectral image are simultaneously obtained through 3D convolution,and the SENet module is used to assign weights to different features of the highest dimension,so that the network can learn more important high-dimensional features and obtain better segmentation results.Experimental results show that the PCA and SURF algorithm with prior information can accurately achieve the registration of hyperspectral images,and the linearly weighted image fusion method can also eliminate the brightness jump in the stitched image.The 3D SE-U-Net proposed in this thesis can more accurately achieve the segmentation of the lymphoma area.The overall segmentation accuracy can reach0.939,and the accuracy and recall rates of the cancerous area are higher than 0.8.The deep learning-based lymphoma microscopic hyperspectral image recognition and analysis method studied in this thesis can realize automatic segmentation of cancerous regions in lymph nodes,provide a new method for the diagnosis of lymphoma,and provide doctors to a certain extent diagnosis provides support and assistance.