Pre- and post-operative lung cancer recurrence prediction following curative surgery: A retrospective study using European radiomics and clinical data.

person Ann Valter North Estonia Medical Center, Tallinn, Estonia info_outline Ann Valter, Tanel Kordemets, Aydan Gasimova, Noah Waterfield Price, Lutz Freitag, Anil Vachani, David Paul Carbone, Kersti Oselin

Authors person Ann Valter North Estonia Medical Center, Tallinn, Estonia info_outline Ann Valter, Tanel Kordemets, Aydan Gasimova, Noah Waterfield Price, Lutz Freitag, Anil Vachani, David Paul Carbone, Kersti Oselin Organizations North Estonia Medical Center, Tallinn, Estonia, North Estonia Medical Centre, Tallinn, Estonia, Optellum Ltd., Oxford, United Kingdom, University of Pennsylvania, Merion Station, PA, The Ohio State University Comprehensive Cancer Center and the Pelotonia Institute for Immuno-Oncology, Columbus, OH Abstract Disclosures Research Funding Optellum Ltd Background: Accurately estimating the risk of lung cancer recurrence is vital for making informed treatment choices, both before surgery, such as the prescription of neo-adjuvant/peri-operative therapies and the extent of lung resection, and after surgery, such as the prescription of adjuvant therapies and planning follow-up strategies. In this study, we build on previous work demonstrating the predictive power of pre-operative patient demographic and imaging features and develop a machine-learning model using a combination of clinical and tumour radiomic features in the pre- and post-operative setting. Methods: We collected a dataset of 323 clinical stage I-IIIA lung cancer patients who underwent surgical treatment for lung cancer, of which 94 had lung cancer recurrence. This includes retrospectively collected CT images and associated patient demographic and diagnostic data from North Estonia Medical Centre (NEMC). The models were trained to predict the likelihood of recurrence on a diverse set of features, including radiomic features extracted from CT images and relevant clinical variables. Post-operative clinical factors such as pathological staging and tumor histology were only included in the post-operative model. An 8-fold cross-validation strategy was used, where, in each fold, 6 (of the 8 equally sized) subsets were used for training, one for model tuning, and one for validation. As a baseline, we compare the pre- and post-operative models to clinical and pathological staging, respectively. Performance was evaluated using the Area-Under-the-ROC-Curve (AUC) and sensitivity (at a fixed specificity of 90%). Results: Lung cancer recurrence classification results are tabulated below for the pre-operative models. We find that the pre-operative model (AUC=74.1%, Sens=35.29%) performs significantly better than clinical staging alone (AUC=65.4%, Sens=14.71%, p=0.025), and that the post-operative model (AUC=74.4%, Sens=26.47%) performs significantly better than pathological staging alone (AUC=67.3, Sens=22.4%, p=0.045). Top predictive features for both the pre-operative and post-operative models included clinical staging, tumor shape and textural radiomics, tumor size and tumor location. The post-operative model additionally included pathological staging and tumor histology. Conclusions: Based on this retrospective analysis, we find that the model outperforms staging prediction of lung cancer recurrence in pre- and post-operative settings. With further development, these algorithms could prove a valuable tool to aid the management of lung cancer patients. Predictors AUC Sensitivity Specificity cTNM 65.4 (59.0, 71.6) 14.7 (8.9, 43.5) 90.0 preop ML model 74.1 (68.0, 79.8) 35.3 (24.5, 49.0) 90.0 pTNM 67.3 (61.5, 73.3) 22.4 (13.5, 35.8) 90.0 postop ML model 74.4 (68.8, 79.9) 26.5 (16.2, 41.9) 90.0