Jonas Teuwen | Predicting survival in lung cancer with neural networks

Non-small lung cancer (NSCLC) is a common and very deadly disease. Medical imaging techniques have been a major contributor in personalising the treatment based on individual characteristics of the patient and their disease. On these images, lung cancer tumours exhibit strong visual prognostic features related to their shape, intensity, or texture. However, in clinical practice, the appearance of the tumour is often described qualitatively such as: very spiculated, has an ill-defined border, or having a large necrotic core.

In the past years, there has been a great interest in developing quantitative descriptors of these tumour features. Up to now, most of these features are hand-crafted and might not represent all prognostic information available in the image.

In this talk, after briefly explaining the CT scan, we will look at how NSCLC patients are typically treated using radiation and why the CT scan provides important information. I will present the current state-of-the art in predicting survival in NSCLC using hand-crafted features. We will compare this to our work-in-progress on learning these features from data collected at the radiotherapy department of the NKI-AvL in Amsterdam using neural networks. Finally, we will look into the technical and theoretical problems of applying such models to heterogeneous tumour sizes and censored data.