Image Mining and ctDNA to Improve Risk Stratification and Outcome Prediction in NSCLC Applying Artificial Intelligence.

Lung cancer is the leading cause of cancer-related death in Europe. Pathological staging is the gold standard, but it can be influenced by neo-adjuvant treatment and number of sampled lymph nodes; it is not feasible in advanced stages and in patients with high-risk comorbidities. Therefore, patients with tumors of the same stage can experience variations in the incidence of recurrence and survival since suboptimal staging leads to inappropriate treatment that result in poorer outcomes. It is still undetermined what are the tumor characteristics that can accurately assess tumor burden and predict patient outcome.Our central hypothesis is that image-derived and genetic characteristics are consistent with disease stage and patient outcome. Combining through artificial intelligence techniques data coming from imaging and circulating cell-free tumor DNA (ctDNA) can provide accurate staging and predict outcome. This hypothesis has been formulated based on preliminary data and on the evidence that image-derived biomarkers by means of image mining (radiomics and deep learning algorithms) are able to provide "phenotype" and prognostic information. On the other hand, the analysis of ctDNA isolated from the plasma of patients has been proposed as an alternative method to assess the disease in the different phases, in particular, at diagnosis and after surgery, for detection of residual disease.

Our central hypothesis is that peculiar image-derived and genetic characteristics are consistent with disease stage and patient outcome. Therefore, they are biomarkers of disease burden and relapse, which can be non- invasively assessed. The combination through artificial intelligence methods of data coming from medical imaging and circulating cell-free tumor DNA (ctDNA) can provide accurate staging and outcome prediction. This hypothesis has been formulated based on the evidence that medical images are able to provide meanable data reflecting tumor characteristics, capturing intrinsic tumor heterogeneity, non-invasively, using a whole- body and whole-lesion assessment. In fact, in recent years, advanced analysis of medical imaging using radiomics, machine learning or in combination - image mining, has been explored. Image- derived biomarkers, by means of texture feature extraction and convolutional neural network application, have been tested to provide "phenotype" information (malignant vs benign, and histotype identification, and T or N staging. Moreover, correlations between image-derived quantitative features with tissue gene-expression patterns have been shown, linking the imaging phenotypes to the genotype as also demonstrated in our preliminary data. Secondly, image mining approach has been proposed to provide prognostic information at baseline evaluation, as also shown in our previous work. Still, few prospective studies with robust methodological approach have been published. On the other hand, the analysis of circulating cell-free tumor DNA (ctDNA) isolated from the plasma of lung cancer patients has been proposed as an alternative method to assess the disease in the different phases. In particular, at diagnosis, the post-surgical detection of residual disease, the identification of mutations in the metastatic setting for treatment guidance and monitoring treatment response. Even if, ctDNA has been detected in patients with all stages of NSCLC with levels increasing with stage and tumor burden ctDNA information has not been explored yet for the purpose of staging. The possibility to detect a tumor in the early phase of its development or the recurrence has to face the issue of the low amount of cfDNA in patients with minimal disease burden. Moreover, the presence of a para-physiological ctDNA background particularly in aged people affects the specificity. In this respect, the investigators expect that the combination of different biomarkers will allow to solve this problem.Artificial Intelligence analytics are increasingly described in healthcare applications. In recent years, supervised, semi-supervised, and unsupervised machine learning methods have been applied to analyze genomic, proteomic, clinical data and radiographical characteristics. Deep learning methods offer opportunities for comprehensive analysis of multi-dimensional data for improved prognosis prediction. The rationale for the proposed project is that, once it is known which imaging features and ctDNA-derived information is linked to the tumor stage and post-operative risk of relapse, the developed algorithm will be an effective and innovative approach for both staging and follow-up of patients affected by lung cancer, with implications on decision-making in clinical practice.