T Shiinoki*, K Fujimoto, Y Kawazoe, Y Yuasa, M Kajima, Y Manabe, T Hirano, K Matsunaga, H Tanaka, Yamaguchi University

• 
  T Shiinoki

• 
  

MO-B-TRACK 6-3 (Monday, 7/26/2021) 11:30 AM - 12:30 PM [Eastern Time (GMT-4)]

Purpose: Immune checkpoint therapy has been the research hotspot in clinical practice since new cancer therapy was discovered. Immune inhibitors which can hinder the combination between programmed death 1 (PD-1) and programmed death ligand 1 (PD-L1) may extend patients’ progression-free survival or overall survival with lower side effects. The purpose of this study is to construct the model to predict the PD-L1 expression level in Non-small cell lung cancer (NSCLC) using CT images and machine learning.

Methods: 203 patients who underwent CT scan were enrolled in this study. PD-L1 expression level was tested through the immunohistochemistry (IHC). Patients’ data were divided into two parts (training and test data) by random selection during feature selection procedure. For each Tumor, Tumor-5 mm and Ring structures, 1130 radiomics features were calculated. Then, these features and 7 clinical features (age, sex, stage, histology, location, smoking, CEA) were used to construct the prediction model. Feature selections with Null importance and Lasso algorithms were performed. The prediction models were constructed using three machine learning algorithms (LightGBM, SVM: Support vector machine and LR: Logistic regression) with five-fold cross validation. For training and test data, AUCs were calculated to evaluate the accuracy of prediction model for each algorithm.

Results: For training data of PD-L1≥50%, prediction models for LightGBM, SVM and LR resulted in an AUC of 0.95, 0.50, and 0.71, respectively. For test data, AUC were 0.79, 0.63, and 0.67, respectively. For training data of PD-L1≥1%, those resulted in AUC of 0.98, 0.69, and 0.68, respectively. For test data of PD-L1≥1%, AUC were 0.76, 0.59, and 0.58, respectively.

Conclusion: The radiomic-based predictive approach may predict PD-L1 expression status in NSCLC patients relatively accurate. It may be helpful in guiding immunotherapy in clinical practice and deserves further analysis.

Funding Support, Disclosures, and Conflict of Interest: Japan society for the promotion of science (JSPS) KAKENHI Grant number 18K07753(TS)

Handouts

Keywords

Radioimmunotherapy, Image Analysis, Lung

Taxonomy

IM- CT: Machine learning, computer vision

Contact Email