Deep Learning-based Assistant Diagnosis Of Benign And Malignant Tumors Of Eyelid

All kinds of skin tumors are the most common tumors.Eyelids are a small part of skin,however,about 5% to 9.2% skin cancers taking place on the eyelids.Since eyelids develop from different germinal layers,a wide variety of benign and malignant tumors may develop.The benign and malignant tumors of eyelids are similar and might be misdiagnosed.Most benign tumors only affect facial appearance,whereas,some malignant types might be lethal.Pathological histological examination is the golden standard of clinical diagnosis.Unfortunately,the pathologists in China are so rare comparing with huge number of patients.Ophthalmology,as a small specialty,has a bigger gap in the corresponding direction of pathologists.Pathologists are overloaded and some patients with malignant tumors cannot get timely and accurate diagnosis,delaying valuable early treatment.In recent years,along with the perfection of digital pathological imaging technology and the rapid development of deep learning,computer-aided diagnosis of pathological slides has become a research hotspot.Computer-aided diagnosis can overcome the shortcomings of pathologist's diagnosis,and has great potential for improving the efficiency and accuracy of pathological diagnosis.Based on the digital pathological slides of benign and malignant tumors of the eyelid,this paper constructs a computer-aided diagnosis system by deep learning.Firstly,it realizes the binary classification system of malignant melanoma(MM)and non-malignant melanoma(NMM),which is the basic classification.Then,a system for classifying nine types of benign and malignant tumors is realized,which is an exploration of the system to improve the classification accuracy of difficult to distinguish tumors.The main research work and contributions of this paper are as follows:(1)Aiming at the task of binary classification of MM and NMM of eyelid,a deep learning tandem machine learning framework is proposed to realize the binary classification system of computer-aided diagnosis of pathological slides of the eyelid.The convolutional neural network is used to realize the patch level classification,and the tandem random forest classifier is used to achieve the pathological image level classification.Both classifications achieve more than 90% accuracy.The system also realizes the visualization of probabilistic heatmap,marking MM and NMM areas on digital pathological slides,and assisting pathologists in diagnosis quickly.(2)For the problem of poor generalization of the model,the edge extraction mapping method based on the Sobel operator(EMBS)is proposed,which can slow down over-fitting and improve the accuracy by 24.4%.The color constancy preprocessing method solves the problem of different color distribution and improves the accuracy by at least 2%.(3)Aiming at the classification task of nine types of benign and malignant tumors for the eyelid,for the problem of low classification accuracy,the random bounding box and zooming preprocessing method is proposed to improve the accuracy by 25.7%.The center loss limited loss function is used to reduce the intra-class distance and improve the classification accuracy.The model ensemble method is used to further improve the performance.The accuracy of patch level classification reaches 75.41%.