Research On Pathological Image Assisted Diagnosis Method Based On Deep Learning

Pathological images are high pixel,full information,and standardized images.Pathological examination has always been the gold standard for cancer identification and diagnosis.During clinical diagnosis,doctors provide detailed diagnostic results by analyzing pathological images of human tissues or cells.The diagnosis of a patient usually requires a few to a dozen pathological sections,which typically contain about 15 mm × 15 mm tissue area,and the pathological images obtained by scanning each pathological section usually reach billions of pixels.In order to facilitate doctors’ observation,each pathological image is divided into tens of thousands of small images,and pathologists need to carefully observe these images.Therefore,the diagnosis of a patient usually takes a pathologist several days.Computer aided diagnosis is an important research field in digital intelligent medicine.It has been used to help medical professionals identify and highlight suspicious areas in medical images.Based on existing hospital cooperation projects,this article mainly conducts research and exploration on pathological image analysis in the field of medical imaging.The research direction includes two directions: pathological image classification and segmentation,and applies relevant algorithms to domestic embedded systems,while improving the detection rate of diseased areas and reducing the long working hours of pathologists.The main research contents and contributions of this article are as follows:(1)Regarding the perception of cancerous regions,considering the high labeling cost and multiple types of cancerous regions,a scheme is proposed to quickly locate candidate cancerous regions.Based on the labeling results,this paper constructs a “Multi-strategies and lightweight classification network” scheme,which is tested on multiple classification neural networks,and selects the network with the best effect for classification.This set of schemes has achieved good results in clinical classification tasks.(2)For the nuclear segmentation task of cervical cells,this article improves the encoder\decoder\skip connections based on the U-Net network framework.In order to increase the recall rate of the target cell nuclei,a multi-scale feature fusion module is added to the network,and a target network model is constructed by combining the existing Squeeze and Excitation modules and attention modules.Good results are obtained on both public and private datasets.(3)The combination of smart healthcare and edge embedded devices has enormous industrial potential.In order to verify the engineering practicality of the algorithm,this paper quantitatively deploys and verifies the proposed segmentation algorithm on the Horizon embedded device Sunrise X3 Pi.Through experiments,it is concluded that the quantified model can achieve good results on the development version,and the reasoning speed can be effectively improved.