Pathological Analysis Of Thyroid Puncture Cells Based On Spectral Imaging And Multi-link Dual Network

In recent years,over 200000 people nationwide have been diagnosed with thyroid cancer every year,and this number is still increasing year by year.The detection and subtype diagnosis of thyroid cancer have gradually become the most urgent issues to be solved in the national and even global medical field.At present,the results of thyroid fine needle aspiration examination are still the "gold standard" for the diagnosis of thyroid cancer.The process of observing and verifying smears under a microscope by pathological experts is time-consuming and requires a lot of effort from doctors.The rich experience of doctors can have a huge impact on the diagnostic results.Currently,there is very little work that can provide doctors with qualitative and quantitative auxiliary examination results.The image analysis method based on deep learning,by summarizing the common features of pathological samples,saves doctors a lot of energy and reduces the impact of inconsistent subjective opinions on patients.Therefore,this method is thriving.Conventional medical image pathology analysis methods usually analyze microscopic color images of samples,while microscopic hyperspectral images not only contain spatial information but also spectral information,providing a new solution for qualitative and quantitative auxiliary detection tasks of thyroid fine needle puncture.On this basis,this article focuses on the segmentation and classification of thyroid cancer fine needle aspiration cells using deep learning and microscopy hyperspectral imaging techniques.The main content of this article includes the following aspects: Firstly,a preprocessing technique based on Lambert Beer’s law is used to perform spectral correction on the microscopy hyperspectral images of thyroid fine needle aspiration cells.Subsequently,a thyroid fine needle aspiration microscopy hyperspectral dataset is established and feature matching is performed using the SIFT algorithm.The above two technologies can reduce the impact of system noise on the image.Secondly,based on the characteristics of microscopy hyperspectral images of thyroid fine needle aspiration cells,this paper proposes a multi-link dual convolution microscopy hyperspectral image segmentation method and a multi-link twin microscopy hyperspectral image classification method to assist doctors in segmenting and classifying different cells.The segmentation algorithm combines the advantages of multi-link dual convolution structure,KL divergence,and full advance residual convolution module,which can combine the prediction results with multimodal information during network training.The classification network combines multi-link twin structures,Pierce loss,pre background branches,and SE modules based on sample characteristics,enabling the network to learn more effective features during training.Finally,this article investigates the global attention module based on attention mechanism and the pre training model of microscopic hyperspectral images based on Byol to enhance the global perception and stability of the network.The experimental results show that the multi-link double convolutional integral cutting algorithm studied in this paper can achieve automated segmentation of different cells in microscopy hyperspectral images.The overall accuracy,accuracy,recall,Dice Score,and Io U of the segmentation can reach 0.8803,0.8784,0.8330,0.8642,and 0.7891,respectively.The multi-link twin classification algorithm proposed in this article can also effectively complete the classification task of thyroid fine needle puncture cells,with an overall accuracy of 0.928.The relevant results can assist doctors in thyroid cancer cell recognition analysis and pathological diagnosis.