Research On Recognition Method Of Pathological Image Of Basal Cell Carcinoma Based On Deep Learning

Skin cancer has a high incidence and poses a greater threat to people’s life and health.Pathological images are the gold standard for diagnosing cancer,but pathologists often need to spend a lot of time repeatedly observing pathological images to make a diagnosis.This subjective judgment method is prone to misdiagnosis,missed diagnosis and low efficiency.The uneven distribution of lesion areas and huge amount of information in skin cancer pathological images make it impossible to use traditional algorithms to extract pathological image features.As the deep learning technology gains more and more attention,more and more excellent models have been proposed,but the existing models cannot well solve the problem of skin cancer pathological image recognition.Therefore,in view of the problem of basal cell carcinoma pathological image recognition,combined with the existing theoretical methods,two models are designed to achieve the purpose of identifying basal cell carcinoma.In the preprocessing stage of skin cancer pathological images,the SPCN algorithm is used to standardize the image for staining differences in the image;for the problem of a large number of useless areas in the image,the Grab Cut algorithm is used to remove the background of the stained standardized image;for the original data For problems with a small number,data enhancement is performed by extracting image patches and image transformations.Aiming at the problem of unclear pathological image features of basal cell carcinoma and complicated pathological features that are difficult to identify,a VGG-16-based basal cell carcinoma recognition model is designed: that is,to determine whether the image to be recognized is basal cell carcinoma.Based on the simple structure of the VGG-16 model but the bloated fully-connected layer,the fully-connected structure of the original model is removed,the original model is improved by adding GAP,Dropout and Softmax layers,and the impact of different transfer learning methods on the improved model is explored.Experiments show that the recognition accuracy of this model for basal cell carcinoma reaches 98.8%,which is 7.4% higher than that of the original model.Aiming at the problem that the pathological image of basal cell carcinoma and the pathological image of piloblastoma and hair epithelioma have similar pathological features and are difficult to distinguish,a basal cell carcinoma recognition model based on Inception-v3 is designed: firstly,it is judged whether the image to be recognized is Basal cell carcinoma,when judged as No,continue to be identified as piloblastoma or trichomoepithelioma.According to the characteristics of deep network learning features that are stronger but easier to overfit,the fully connected structure of the original model is removed,and GAP,BN,Dropout,Dense,Dropout and Softmax layers are sequentially added to complete the improvement of the original model.According to the characteristics of inconspicuous cell characteristics and structural characteristics in pathological images,a multi-size sampling strategy is proposed.Experiments show that after using the multi-size sampling strategy,the accuracy of the model test is increased by 2.7%,and the accuracy of the recognition of basal cell carcinoma reaches 99.3%.