Classification Of Lung Cancer Subtypes Based On Cytopathologic Images

Lung cancer is one of the most common cancers worldwide,and its incidence and mortality rates remain high all year round,seriously affecting human health.Therefore,early diagnosis and treatment are necessary to improve the survival rate of lung cancer patients.However,it is crucial to clarify the lung cancer subtypes since the treatment and prognosis options are different for different lung cancer subtypes.Currently,in traditional lung cancer cytopathology diagnosis,it mainly relies on microscopic reading with pathologists,which is a difficult and time-consuming process,and its diagnostic results are easily influenced by physicians’ subjective factors.Therefore,this paper investigates the classification method of lung cancer subtypes by using computer-related technology with lung cancer cytopathological images as the research object.The main research contents of this thesis are as follows:In the preprocessing stage of lung cancer cytopathology images,firstly,the brightness adjustment of the images was realized based on the adaptive Gamma correction method,which improved the contrast between the cellular region and the background region in the images;secondly,the data enhancement algorithm of SAGAN was used to solve the problem of uneven image categories;finally,the data were expanded using image processing methods such as rotation,flip and scaling.The experimental results show that the classification accuracy of the dataset obtained based on the SAGAN data enhancement algorithm on the Resnet50 and EfficientNet B0 classification models is improved by 10.92% and 7.15%,respectively,compared with the previous ones.In the lung cancer subtype classification stage,six different classification models,including ResNet34,ResNet50,DenseNet161,DenseNet169,MobileNet v2 and EfficientNet B0,were used to classify three lung cancer subtypes,including squamous lung cancer,lung adenocarcinoma and small cell carcinoma,and the model performance was also analyzed and evaluated by combining five evaluation indexes.The experimental results showed that the classification model based on EfficientNet B0 showed good performance in classifying lung cancer image subtypes,and its accuracy,precision,recall rate,specificity and F1-Score indexes were 83.3%,84.0%,83.3%,91.7%,83.0%,respectively.In order to further improve the model classification performance,the main module MBConv in the EfficientNet B0 model was optimized.Firstly,to address the problem that it is difficult to distinguish lung squamous carcinoma from lung adenocarcinoma,the CBAM module is used in the feature extraction part.Secondly,to address the slow convergence of the model,the h-swish activation function is introduced.Then,to address the poor characterization of traditional convolutional local sensory field features,the group convolution and channel confusion operations are introduced.Finally,an improved model EfficientNet-HCS combining these three is proposed.The experimental results show that the model achieves 88.7% accuracy in lung cancer subtype classification,which is 6.48% higher than the original EfficientNet B0 network,and 5.60%,6.48%,2.84%,and 6.63% higher than the original EfficientNet B0 network in terms of precision,recall,specificity,and F1-Score indexes,respectively.