Cancers associated with smoking and tobacco use can occur in any part of the body, and non-lung primary or secondary cancers can be as harmful as lung cancers. Early detection of these cancers poses an opportunity to inform treatment plans and even prevent the onset of cancers, directing the study in this paper. The model developed in this paper involves reviewing single low-dose chest computed tomography (LDCT) scan images to predict future smoking-related cancer risk using an extension of the deep learning algorithm Sybil. Sybil predicts future risk of developing lung cancers instead of identifying an existing tumor, unlike other machine leaning (ML) models used today. Our extended model masters in predicting non-lung, smoking-related cancers with over 70% accuracy, cancers which are harder to predict than just lung cancers due to the sole availability of lung screening data in smokers for training models. Our model also utilizes diagnostic metadata to predict the grade and stage of potential lung cancer with over 80% accuracy. Ultimately, such a model will aid doctors in better treating their patients, making the treatment plan cater well to individual needs. We also propose a plan for clinical studies using our model so that it can be implemented in hospitals to work alongside onco-radiologists.

Participating in superUROP gives me the ability to have long term ownership on a project, something that I cannot have normally in class or via a UROP, something that I yearn to have and am fortunate to have in my undergrad. I have had numerous research experiences in computational and biological labs at MIT. I get excited by computational techniques applied to biology, and this project allows me to do that in my favorite direction, oncological diagnosis through radiology.