Abstract 147P
Background
Radiomics emerged as a non-invasive tool able to decode tumor tissue characteristics. We aimed to develop radiomics-based multivariate classifiers to non-invasively decipher TIME (tumor immune microenvironment) features in resected NSCLC patients (pts).
Methods
The density of Tumor-Infiltrating Lymphocytes (TILs) subpopulations and PD-L1 were histologically assessed. Radiomic features (RFs) were extracted from CT scans using the pyRadiomics tool. Amachine learning-based radiomic model was developed to predict each TIME endpoint: M1 (PD-L1), M2 (CD8/CD3 ratio), M3 (PD1/CD8 ratio), M4 (total CD3+ TILs), M5 (CD4+ TILs), M6 (CD8+ TILs), M7 (predefined hot and cold TIME). Pts were stratified into high vs low groups using median values as cut-offs for respective outcomes, except for M1 (PD-L1 positive vs negative) and M7, hot vs cold TIME. For each radiomic model, the Random Forest Classifier algorithm with sequential feature selection and Stratified-5-fold cross-validation and Leave-One-Out Cross Validation (LOOCV) splits was employed.
Results
We collected clinicopathological and radiological data from 98 stage I-IIIA resected NSCLC pts. A total of 1702 RFs were initially extracted for each patient, with 851 from tumor and 851 from peritumoral region. After removing highly correlated features, 116 RFs from tumor and 115 from peritumoral region remained. The final models selected between 3 and 7 RFs from both regions, with maximum depths ranging from 7 to 10 and the number of estimators between 100 and 400. Accuracy, precision, sensitivity, receiver operating curve (ROC) AUC and precision-recall curve (PRC) AUC values ranged from 70% to 80%, except for M2 (PRC AUC=64%). Specificity was slightly lower, with the highest value shown by M6 (77%) and the lowest by M2 (55%). Selected RFs varied in the different models predicting distinct outcomes, except for peritumor HLL-wavelet GLCM Correlation feature, which was shared by four models (M1, M2, M4, M5).
Conclusions
Our results determined that CT-derived RFs, from tumoral and peritumoral regions, might effectively predict distinct TIME features, thus embodying a valid non-invasive tool to manage NSCLC pts.
Legal entity responsible for the study
University Hospital of Parma, Parma, Italy.
Funding
AIRC-Associazione Italiana per la ricerca sul cancro.
Disclosure
M. Tiseo: Financial Interests, Personal, Advisory Board: AstraZeneca, Boehringer Ingelheim, Roche, Amgen, Takeda, MSD, Merck, BMS, Pfizer, Eli Lilly, Novartis, Janssen, Daiichi Sankyo; Financial Interests, Institutional, Research Grant: AstraZeneca, Boehringer Ingelheim, Roche. All other authors have declared no conflicts of interest.