Persistent Homology of Tumor CT Scans Predicts Survival In Lung Cancer
ABSTRACTRadiomics, the objective study of non-visual features in clinical imaging, has been useful in informing decisions in clinical oncology. However, radiomics currently lacks the ability to characterize the overall structure of the data. This field may benefit by incorporating persistent homology, a popular new algorithm that analyzes whole data structure. We hypothesized that persistent homology could be applied to lung tumor scans and predict clinical variables. We obtained computed tomography lung scans (n = 565) from the NSCLC-Radiomics and NSCLC-Radiogenomics datasets in The Cancer Imaging Archive. Segmentation data was used to create a cubical region centered on the primary tumor in each scan. For each scan, a cubical complex filtration based on Hounsfield units was generated. We created a feature curve that plotted the number of 0 dimensional topological features against each Hounsfield unit. The curve’s first moment of the distribution was utilized as a summary statistic to predict survival in a Cox proportional hazards model. The first moment of the distribution is equivalent to the area under the curve of our topological feature curves (AUC). The Kruskal-Wallis H Test and a post-hoc Dunn’s test with Bonferroni correction were used to test AUC differences among survival quartiles. After controlling for tumor image size, age, and stage, AUC, was associated with poorer survival (HR = 1.118; 95% CI = 1.026-1.218; p = 0.01). AUC was significantly higher for patients in the lowest survival quartile compared to the highest survival quartile (p < 0.001). We have shown that persistent homology can generate useful clinical correlates from tumor CT scans. Our 0-dimensional topological feature curve statistic predicts survival in lung cancer patients. This novel statistic may be used in tandem with standard radiomics variables to better inform clinical oncology decisions.