The Capability of Deep-radiomics to Predict Pathological Response to Neoadjuvant Immunochemotherapy in Non–small Cell Lung Cancer: A Retrospective Multicenter Study

Yuanxin Ye¹Yuchi Tian²Lingling Wang¹Zihan Xi¹Yangfan Zhang³Tong Zhou⁴Zhenhua Zhao⁴Yifeng Zheng³Xiaoyun Liang^2*Haitao Jiang^1*

• ¹Zhejiang Cancer Hospital, Hangzhou, China
• ²Neusoft Medical Systems Co Ltd, Shenyang, China
• ³Huzhou Central Hospital, Huzhou, China
• ⁴Shaoxing People's Hospital, Shaoxing, China

Background To establish a predictive model that combines radiomics, deep learning and clinical features for predicting the pathological complete response (pCR) of non-small cell lung cancer (NSCLC) patients after neoadjuvant immunochemotherapy (NIT). Methods We retrospectively collected patients from three centers(split into training, internal testing and external testing cohorts). In this study, tumor segmentation was performed on chest CT images before (pre-NIT) and after (post-NIT) neoadjuvant therapy. The radiomics features were extracted from pre-NIT and post-NIT image. Deep learning (DL) features were extracted from the post-NIT images. The most meaningful features were selected using the mRMR and LASSO. A logistic regression classifier was then applied to create a classification model to predict pCR or non-pCR. The predicted probability were referred to as the Rad-scores and Deep-scores. Finally, Rad-scores, Deep-scores, and meaningful clinical features were fused to build a combined model. Results A total of 178 patients were enrolled in the current study. In conventional radiomics, the efficacy of post-NIT model was better than the pre-NIT. In delta radiomics model, delta1 had the best efficacy. Subsequently, the post-NIT and delta1 features were further constructed as the combined model 1 with AUCs of 0.939 and 0.849, respectively. iRECIST was combined with the radiomics and the DL features to establish the combined model 2, which achieved the best performance among all the models, with AUCs of 0.955(training), 0.882(In-testing), and 0.839(Ex-testing). 2 Conclusions Our results demonstrated that combination of three dimensional features can provide complementary information to predict pCR more accurately.