Research On Classification Method Of Lung Cancer Subtypes Of Unbalanced PET/CT Samples

Lung cancer is one of the most common cancers,and its morbidity and mortality are increasing rapidly.For different lung cancer subtypes,the treatment options are very different.For example,squamous cell carcinoma is mainly treated with radiotherapy;adenocarcinoma is mainly treated with chemotherapy.Currently,lung cancer typing mainly relies on manual diagnosis,resulting in low efficiency and poor accuracy.This article uses the patient’s CT and PET images as the data set,and uses the deep learning method to train the model to complete the automatic classification of lung cancer types.In the paper,ResNet50 is used as a feature extraction network to realize classification using only CT images or PET images and using CT and PET images as input.Observing the results,it is found that the combined use of CT and PET is better.The model used by CT and PET is used as the classification benchmark model.In order to alleviate the extremely imbalanced problem of the existing lung cancer data sets,we propose a variety of mitigation methods.Specifically,it can be divided into four directions,namely the method based on resampling,the method based on loss function optimization,the method based on hybrid strategy,and the optimization of network structure.We have conducted perfect experiments on these methods.Experimental results show that for a variety of methods to alleviate category imbalance,except for downsampling methods,others can efectively improve the classification accuracy of minority categories.Among them,the model combining Remix strategy and threedimensional convolutional neural network performs best,in each category The accuracy is higher than that of the doctor’s manual diagnosis,and at the same time it effectively alleviates the problem of category imbalance.Considering that people suffering from adenocarcinoma and squamous cell carcinoma account for more than 80%of lung cancer patients,we built a binary classification model to distinguish between adenocarcinoma and squamous cell carcinoma and used the ROC curve for evaluation.The experimental results show that both the Xgboost model and the three-dimensional convolutional neural network model exceed the accuracy of doctors’ manual diagnosis.