Room: Track 2
Purpose: The identification of epidermal growth factor receptor (EGFR) mutations, which lead to aberrant proliferation or differentiation in cancer cells, is a crucial step to decide treatment strategies to non-small cell lung cancer patients, because EGFR-tyrosine kinase inhibitors (TKIs) extend median progression-free survivals of those patients compared with standard chemotherapies. We have investigated a phenotyping potential of topologically invariant Betti numbers (BNs) for EGFR mutations in lung cancer patients.
Methods: One hundred patients (EGFR mutation status: mutant (40), wildtype (60); gender: male (59), female (41); age range: 27–89 y (median 65 y); stage: I(4), II(1), III(12), IV(84)) with non-small cell lung cancer were chosen for this study. A total of 41,475 radiomic features based on BNs were obtained from computed tomography images by applying 14 histogram- and 40 texture-based feature calculations to BN maps that phenotype topologically invariant heterogeneous characteristics of lung cancer. A BN-based signature was constructed using an elastic-net-regularized logistic regression (LR) model. LR models were utilized to build an identification model of EGFR mutants using the BN-based signature. The identification model was evaluated in a five-fold cross validation. The identification performance of the BN-based signature was compared with two conventional identification approaches including wavelet-decomposition (WD)-based radiomic features and a deep learning (DL) using area under the receiver operating characteristics curves (AUCs).
Results: The AUCs for BN-, WD- and DL-based approaches were 0.89 (95 % confidence interval (CI): 0.81–0.94), 0.75 (CI: 0.62–0.83) and 0.57 (CI: 0.45–0.67), respectively. The p-values (Delong’s test, significance threshold p < 0.05) between two AUCs were 1.8 × 10?² (BN v.s. WD), 1.8 × 10?6 (BN v.s. DL) and 1.9 × 10?² (WD v.s. DL).
Conclusion: Topologically invariant Betti numbers could have a potential to phenotype mutations of EGFR in lung cancer patients.
IM/TH- Image Analysis (Single Modality or Multi-Modality): Quantitative imaging