CT-Based Radiomics For Ki67 Prediction In Lung Adenocarcinoma

Lung adenocarcinoma is a more common subtype of lung cancer with high morbidity and mortality.Frequently,the first-time patients miss the best time for treatment due to its early unconspicuous symptoms.Therefore,Rapid and accurate determination of tumor differentiation status is of important clinical significance for the treatment and evaluation of patients with lung adenocarcinoma.As an antigenic protein that can detect tumor differentiation status.With the development of personalized precision medicine,tumor biomarkers have become increasingly important in clinical management.Ki67 plays a particularly important role in the clinical treatment of lung adenocarcinoma.At present,Ki67 detection in lung adenocarcinoma is mainly based on immunohistochemistry,which is invasive for patients,and complicated and time-consuming in operation.It may also cause cancer cell metastasis during sampling.Therefore,based on the CT imaging data of patients with lung adenocarcinoma,this study investigated a noninvasive prediction method for Ki67 in lung adenocarcinoma with the help of radiomics and deep learning,so as to provide an auxiliary means for clinical treatment.The main research contents and contributions of this study are as follows:(1)A prediction method for Ki67 in lung adenocarcinoma based on CT-based radiomics and SHAP framework was proposed.Firstly,the lesion area of lung adenocarcinoma was segmented,and high-throughput radiomics features were extracted from the lesion area.Secondly,to reduce feature redundancy and prevent the curse of dimensionality,dimensionality reduction was conducted on high-dimensional features using the t test,variance selection method and LASSO regression algorithm successively,so as to determine the optimal subset of radiomics features.Additionally,the Ki67 prediction performance of multiple machine learning classification models was compared.Moreover,to improve the interpretability of the models,the important radiomics features affecting the prediction results of the models were quantified and analyzed by interpretability analysis using SHapley Additive ex Planation(SHAP).Finally,the method proposed in this study was trained and evaluated in the CT imaging data set of 661 patients with lung adenocarcinoma.The results showed that the feature subset selected by the feature selection method presented the optimal classification performance in the XGBoost model,and its AUC,accuracy,specificity and sensitivity were 0.773,0.751,0.816 and 0.630,respectively.Compared with other noninvasive prediction methods for Ki67,the proposed method in this study had better prediction performance.(2)A prediction method for Ki67 in lung adenocarcinoma based on multi-source feature fusion was proposed.Firstly,the clinical medical features,radiomics features and deep learning features of the patients were extracted.In addition,multi-source features were fused through the tensor fusion network,to retain the features of single-source features and realize the interaction between multi-source features.Finally,the fused features were input into the neural network classifier composed of full-connection layers to realize the noninvasive prediction of Ki67 in lung adenocarcinoma.The experimental results of this method in the above data set demonstrated that its AUC,accuracy,specificity and sensitivity were 0.800,0.774,0.793 and 0.739,respectively.Compared with the Ki67 prediction method of lung adenocarcinoma based on traditional radiomics,the AUC and accuracy of this method were improved by 2.7% and 2.3%,respectively.