3D-CNN to Predict Non-Small Cell Lung Cancer Survival Outcome

Lung cancer is one of the leading causes of cancer-related deaths world- wide [7]. Non-small cell lung cancer (NSCLC) results in approximately 80-85% of lung cancer diagnosis, with a 5-year survival rate of 24%. Improving dis- ease outcomes of NSCLC relies largely on implementing personalized treatment plans early in diagnosis [13]. Prognosis of NSCLC is key in determining this personalized approach to treatment, as a more-informed decision on treatment can be made after assessing the risk of the patient [8].

Evaluation of prognosis is typically performed by physicians and on- cologists, and consists of a largely qualitative approach that may be prone to human error. However, machine learning offers an accurate alternative to quan- titatively evaluate prognosis of a patient. The widespread use of ML techniques in disease research has led to an increase in the use of deep learning for cancer prognosis, resulting in a 15%-20% increase of prediction accuracy from human accuracy to deep learning in the past few years [8]. Current techniques for can- cer survival prediction involving machine learning largely consist of analysis of genomic, clinical, and histological information surrounding a patient [6]. How- ever, tumor environments include many factors located not only within cancer cells, but also stromal and immune cells. This can often lead to the inclusion of unnecessary information that limits the predictive accuracy of ML algorithms [16]. Such approaches also typically require extensive tests and tissue sampling, and therefore are fairly invasive.

Radiomics for cancer prognosis offers a non-invasive and straightfor- ward approach to cancer survival prediction. Connections between image-related biomarkers and NSCLC progression have largely been established, as CT scans are widely used for TNM staging [10]. Medical imaging can provide a picture of tumor phenotype and the surrounding environment, offering a promising av- enue for prediction of survival through ML. In recent years, radiomics has been extensively investigated for the purpose of outcome prediction, along with other aspects of cancer research (e.g. diagnosis, classifying subtypes, nodule detec- tion) [6]. A number of radiomic features have prognostic power in datasets of lung cancer patients, many of which were not previously noted [1, 6].

Here, we developed an end-to-end model for NSCLC survival outcome prediction. We trained and evaluated the model on pretreatment CT-scans of 422 patients from The Cancer Imaging Archive. Our 2D approach methodology results in an AUC of 0.59 in predicting survival outcome. The 3D approach is currently in implementation, but is anticipated to have a higher AUC. Our framework presents a novel application of a recently developed approach, and demonstrates avenues for further improvement.