Towards an automatic imaging biopsy of non-small cell lung cancer

Because of the high aggressiveness and lethality of lung cancer, its early detection and accurate characterization are among the most investigated challenges in the last years. Biomedical imaging is an important technology in lung cancer assessment, strongly impacting decision making in clinical practice, also employed as a provider of predictive imaging biomarkers. In this context, radiomics approach, which consists of mining vast arrays of quantitative features derived from digital images, has shown promising application perspectives, but suboptimal standardization and controversial results emerged. This work presents the design of a completely automated pipeline for the non invasive in-vivo characterization of Non-Small Cell Lung Cancer (NSCLC), devised to be a support for radiologists and physicians, and to speed up the diagnostic process. Our pipeline exploits data from routinely acquired PET and CT images in order to automatically obtain a reliable segmentation of the tumor lesion: Accurate textural features are computed in the detected Volume of Interest (VOI), thus providing data for the characterization of lung lesion through machine learning algorithms. We evaluated our pipeline on real datasets supplied by a private hospital. Our approach reached a mean accuracy of 94.2pm 5.0% for the VOI segmentation, and it showed the potential of PET/CT features in differentiating both between primary and metastatic lung lesions and between primary lung cancer subtypes.