Publication | Closed Access
Computer extracted features of cancer nuclei from H&E stained tissues of tumor predicts response to nivolumab in non-small cell lung cancer.
12
Citations
0
References
2018
Year
Oncologic ImagingEarly Stage NsclcPathologyCancer BiologyLigand Pd-l1Tumor BiologyCancer DetectionMolecular DiagnosticsTumor Predicts ResponseMolecular OncologyCancer ResearchRadiologyMedicineResponse RateCell BiologyTumor MicroenvironmentLung CancerCancer NucleiMultiple Pulmonary NoduleCancer GenomicsBronchial NeoplasmImmune Checkpoint InhibitorOncology
12061 Background: Immune checkpoint inhibitors have recently been FDA-approved for use in advanced stage non-small cell lung cancer (NSCLC). These drugs target the PD-1 receptor or its ligand PD-L1, but treated patients only have a response rate of about 20%. It is thus crucial to identify which patients will derive maximal benefit from such treatments, especially since the current gold standard biomarker, detection of tissue-based PD-L1 expression, has been shown to be inadequate. Previous studies have shown that computer extracted features of nuclear shape and texture are predictive of recurrence in early stage NSCLC. The goal of this work is to evaluate the role of features of nuclei shape and arrangement in the tumor in predicting response to Nivolumab for NSCLC. Methods: The study included 56 patients with NSCLC from two different institutions who had had pre-treatment tumor biopsies and were treated with Nivolumab. The patients were split into two categories, responders and non-responders, that were determined by clinical improvement and radiologic assessment through RECIST criteria. The 245 features from tumor nuclei included standard measures used to characterize shape and texture of nuclei as well graph based features that capture the distinct spatial arrangement of the nuclei. Features were extracted from tumor regions manually annotated on digitized H&E images by two expert pathologists. Results: A statistical feature selection method determined the top five tumor nuclear features from the training set. These features included the spatial arrangement of nuclei and variance in nuclear shape and chromatin structure. A machine learning classifier trained with these top five features yielded an AUC = 0.65 on the training set (n = 32) and an AUC = 0.6 on the independent validation set from a separate institution. (n = 24). Conclusions: Computer extracted features of cancer nuclei were found to distinguish between patients who did and did not respond to Nivolumab immunotherapy. Validation is needed on larger cohorts from multiple different sites.