Background
Non-small cell lung cancer (NSCLC) is one of the deadliest subtypes of cancer. Recent breakthroughs in the development of immunotherapy have added a very impactful weapon to the arsenal of therapies to combat NSCLC, in some cases doubling the overall survival rate of late-stage NSCLC patients. However, only a fraction of patients responds to these drugs, as low as twenty percent, while the vast majority is exposed to drug toxicity without therapeutic effect.
The current therapy eligibility criteria for immunotherapy are based on visual histological assessment; this assessment is subject to all the limitations that a human reader brings to the table, including reader intervariability and incomplete analysis of the tissue on the slide. Moreover, we hypothesize that the current eligibility criteria fail to capture the whole biological picture, incorporating too few factors of a multi-faceted story.
In this project, we intend to develop deep learning algorithms that a) help reproduce the biomarkers needed for the eligibility criteria so that pathologists may do this analysis in a faster and more reproducible way; b) surpass the current criteria in accuracy of therapy-response prediction by extracting novel biomarkers and combining existing ones in an automated fashion.
