Using a model based on clinical and PET radiomics appears to be a promising strategy for screening at diagnosis, lung adenocarcinoma patients who may have a targetable molecular alteration. This could certainly (i) reduce unnecessary costs by avoiding having to test patients for whom we could know that there is little chance of finding molecular conditions but also (ii) speed up the management of these patients by avoiding waiting for the results of an unnecessary test. For instance, in the present study PET/CT was usually performed more than one month before the availability of genetic analyses results (median = 40 days). At our institution, somatic mutation detection, by means of immunochemistry and NGS Panel CLv3 and single tests for ROS1 and ALK mutations, is performed for nearly all lung adenocarcinoma patients at diagnosis and finally almost half of them had negative findings (46%). It is worth noting that frequencies of molecular alterations observed in our database are representative of those previously observed in a Western population of lung cancer . Using the model equation proposed would therefore have avoided 29 molecular tests (21 in the training and 8 in the validation datasets). In other words, applying this strategy, 1/3 would not have led to molecular testing at the cost of two false negative tests.
The model included was a mix of clinical characteristics (age, sex and AJCC stage) and PET characteristics (correlation_GCLM and GLNU_GLZLM). In sight of the logistic regression analysis, the strongest predictive variable was the sex with an increasing risk of molecular alterations in females. Besides, it was the only variable found to be significantly different between patients with and without molecular alterations on univariable analysis. In our study, more than 80% of female patients had at least one targetable molecular alteration. This is somewhat concordant with previous evidence of a higher rate of EGFR mutations [20–22], KRAS mutations [23, 24], ROS1 gene fusion [25, 26] and STK11  expression in women. Correlation_GLCM represents the linear dependency of grey-levels in GLCM and GLNU_GLZLM the non-uniformity of the grey-levels or the length of the homogeneous zones. Of note, these two parameters were amongst those that needed to be harmonized between the two PET systems, in line with previous studies demonstrating the impact of reconstruction parameters on conventional PET metrics and texture features in NSCLC [2, 28] and therefore the need for harmonizing standards. Moreover, when it comes to the comparison of analogic and digital PET quantitative variables analysis, it is important to point out that conventional and histogram parameters were not found to be different between our systems in the actual reconstruction configurations. A harmonisation process was needed only for some second and third order textural features and only one failure of the process was observed (LZHGE_GLZLM). These findings demonstrated for the first time that textural features extracted from digital and analogic PET systems can be pooled using harmonisation strategies currently under development [15, 29]. However, it has been shown that applying a smoothing filter with a large kernel as per EARL procedure  or using larger voxel size can lead to the loss of accuracy of radiomics metrics for tumor characterisation purposes .
The study has some limitations. First of all, these encouraging results will need to be confirmed by a larger multicentre study and their extrapolation to other populations for which the repartition of histological subtypes and mutational status could be different, will need to be investigated [32–34]. Moreover, to ensure its translation into clinical practice, a worldwide harmonisation strategy is needed and the development of dedicated software for an automatic computation of the model equation seems mandatory. Given the flourishing number of models of this type in the oncology literature for a multitude of hypotheses, this axis of development should be carried out in the short term. Secondly, in view of the great emulation in the framework of precision oncology and genetics, our knowledge will surely be in constant evolution and a strategy which works today will certainly have to be constantly adapted according to new discoveries in the field. Thirdly, the molecular analysis was performed in most cases on biopsies with the risk of spatial tumoral heterogeneity and therefore to miss some tumoral molecular alterations. For example, Swanton et al. noted that many oncogenic alterations were only identified in specific tumor locations generating tumor heterogeneity . Similarly, Pelosi et al. micro-dissected several tumoral regions of different architectures from 20 adenocarcinomas and revealed that 60% of these tumors had intratumoral molecular heterogeneity . We can therefore wonder if some tests considered as false positive results in our study were not linked to a biospsy sampling error. Indeed, metabolic tumor characterization takes into account the entire tumor volume and not just a sample. This can be seen as a strength together with its non-invasiveness. Finally, spatio-temporal heterogeneity could also be considered. In this study the strategy was explored at diagnosis but another time point in patients’ management could be investigated as no consensus has yet been reached . The few patients with false negative tests at diagnosis might benefit from a molecular analysis at another time of their treatment.