Abstract 3971
Background
Radiomics delivers multifaceted tumor characterization with complex quantitative features extraction from medical imaging related with prognostic clinical outcomes in cancer decision support. Distinct from texture and histogram analysis, advanced 2D and 3D shape parameters could reflect tumor local invasion and outcome as well as histological invasion patterns in LA HNSCC pts.
Methods
Tumor contours were defined from semi-automatic delineation of baseline contrast-enhanced CT scan. We trained a radiomic signature made of 27 complex parameters reflecting tumor convolution and shape complexity from 120 LA HNSCC pts retrospectively evaluated at Gustave Roussy Institute. We validated this particular shape signature on 86 LA HNSCC pts from TCGA database with assessable pre-treatment CT. Theses 3D silhouette parameters (n = 27) comprised fraction of convex on concave edges, distance to arithmetic mean average position, and Minkowski dimension. We connected it to PFS, OS, and outcome events. Concordance Index (CI) and Kaplan Meier estimations assessed our signature's competence. Unsupervised bi-clustering methods linked radiomics features and clinical endpoints.
Results
We successively trained a powerful shape signature based on 9 complex parameters in these populations, significant through OS, PFS and local control predictions. Combining our shape radiomic signature with clinical factors (age, tumor location …) significantly improved results over clinical evaluation alone. Particularly, an unsupervised analysis of radiomics shape factors alone linked with tumor location (p
Conclusions
We created a shape-based radiomic signature presenting a potential prediction of invasive histological characters, correlated recurrence and prognosis in LA HNSCC pts, independently from TNM stage, tumor location, surgery, chemoradiation or further treatment, which could change the initial treatment plan for LA HNSCC pts and improve patient stratification.
Clinical trial identification
Legal entity responsible for the study
N/A
Funding
N/A
Disclosure
All authors have declared no conflicts of interest.