Research On The Method Of Automatic Discrimination Of Tumor Histology Types Of Non-small Cell Lung Cancer Based On PET-CT

Different Tumor histology of non-small-cell lung cancer(NSCLC)have distinctly different oncogenic mutations and divergent therapeutic responses.Therefore,the classification of tumor tissue subtypes has an important impact on patients’ choice of the most appropriate treatment plan.It can help doctors to make treatment methods and reduce the pain of patients to fully mine the imaging information of tumor from PET-CT images and use these information to accurately judge the tissue type of the tumor.Mining information from medical images to assist doctors in their work is a hot research,but there are still some problems.On the one hand,the instability of radiomics features makes a large number of features face the problem of underutilization.On the other hand,most studies only classify adenocarcinoma(ADC)and squamous cell carcinoma(Sq CC)and ignore other tumor histology types.In view of the above problems,this paper proposes an automatic grouping and ensemble model of PET-CT image features.Firstly,ADC and Sq CC of non-small cell lung cancer tissue types are classified,and then ADC,Sq CC and other tumor tissue types are classified.The research includes the following two parts.The first part is about the prediction of ADC and Sq CC which solves the problem of instability of image group features and insufficient use of features.In the experiment,PET-CT image features were integrated torandomly group the features in pairs and establish the ensemble model,which solves problems that the contribution of PET radiomics feature are much higher than that of CT radiomics feature and insufficient use of features in traditional methods.The accuracy of the final experimental results is 0.72 and AUC is 0.69.The accuracy is improved by 0.02 and AUC by 0.04 comparing with the feature selection method.In the second part,the research content of the experiment was to predict ADC,Sq CC and other histology.To some extent,it solves the problem that other tumor histology types except ADC and Sq CC are ignored.The three classifications of tumor histology types extend the ensemble method of the two classifications.At the same time,the random forest model will be used in the three classification experiments to sort the features to compare the impact of low contribution rate features on the experimental results.The precision of the best result is 0.48 and the standard deviation is 0.05.The recall rate was 0.59 and the standard deviation was 0.03.Our study showed that quantitative radiomic imaging features of lung tumor extracted from PET/CT images are associated with histology subtypes pulmonary ADC and Sq CC.Radiomics features of PET-CT may be used for histology subtype classification before pathological diagnosis.