• 2019-07
  • 2019-08
  • 2019-09
  • 2019-10
  • 2019-11
  • 2020-03
  • 2020-07
  • 2020-08
  • Aztreonam br Materials and methods br Results


    Materials and methods
    Discussion The clinical characteristics are the most easily available for NSCLC patients, which are usually expected to be predictors of EGFR mutation. There have been cumulative studies reporting that several clinical factors such as female, non-smoker, adenocarcinoma histology, and East Asian population were associated with EGFR mutation [29,30]. Therefore, in clinical practice, these are the primary considerations for EGFR-TKI treatment. Consistent with most of previous findings, we found EGFR mutation rates were higher in female and never-smoker patients, despite no significant difference in smoking status. Furthermore, the rate of EGFR mutations in early-stage NSCLC patients was significantly higher than that in advanced stage patients, similar to a recent study [22]. In the morphological features of NSCLC lesions, maximum diameter, location, density, lymphadenopathy, vacuole sign, and air bronchograms were associated with EGFR mutation status. Maximum diameter of EGFR mutant group was significantly smaller than that of EGFR wild type. Hence, smaller tumors suggest EGFR mutation, which is consistent with other studies [15,16]. The rate of EGFR mutations in peripheral lung cancer was higher than that in central lung cancer, which may be related to the predominance of adenocarcinoma in peripheral location. Prior studies [30,31] showed EGFR mutation was more common in adenocarcinoma than in squamous cell carcinoma. We also found tumors with GGO were more frequent in the EGFR mutation group than those without GGO, similar to previous studies [16,32]. This may be explained by that adenocarcinomas often manifest as ground glass nodules and have a relatively higher mutation rate of EGFR. Tumors without lymphadenopathy, with vacuole sign and air bronchograms were more likely to be EGFR mutants, which is consistent with previous studies [15,32]. A few other features were also found to be associated with EGFR mutations, such as pleural retraction, absence of Aztreonam and spiculation [15,29,32]. Morphological features depend on the radiologist’s experience and subjective interpretation of the signs, which may explain the low AUCs for the morphological features. Due to over-fitting problem, clinical model 1 is not applicable to other datasets and cannot be recommended in clinical practice. In clinical model 2, sex, density and location were important predictors of EGFR mutations. In this study, unsupervised consensus clustering was used to reduce the redundancy of features and select representative medoid features to build radiomics signature. Consensus clustering analysis of radiomics features has been used to establish imaging subtypes of cancer, which are associated with specific molecular pathways and prognosis, such as breast cancer [33] and glioblastoma [34]. Two textural features were involved in the radiomics signature, i.e. X0_GLRLM_RLN and X0_GLCM_homogeneity1, which could reflect the heterogeneity of tumors. The numerator of GLRLM_RLN calculates the squared sum of the run-length values of each run-length, while its denominator serves as a normalizing factor. Therefore, the value of GLRLM_RLN tends to emphasize the non-uniform distribution of the run-length. Meanwhile, GLCM_homogeneity1 assumes larger values for smaller gray-level differences in the co-occurrence pairs. Therefore, these features could measure the textural distribution in the ROIs from different aspects, and reflect the heterogeneity of tumors. X0_GLRLM_RLN was the most important feature because of the significant contribution to the radiomics signature (with the highest regression coefficient). We additionally found that GLRLM_RLN was strongly correlated with volume, and the Spearman correlation coefficient reached 0.89, similar to the study by Welch et al. [35]. The volume was also associated with EGFR mutations (P < 0.05, AUC = 0.603). On the other hand, considering that maximum diameter was an independent predictor in the integrated model, we believe that tumor size is an effective predictor for EGFR mutation and should pay more attention on this finding in future research and clinical applications. Comparing with the morphological and clinical features, radiomics signature showed the best predictive performance to differentiate EGFR mutation group. Thus, radiomics signature may be served as a surrogate for genetic tests, or as additional information to monitor response to therapy.