Lung cancer malignancy predication with recurrent neural networks

Abstract

Lung cancer has the highest mortality rate among all cancer types in the United States, comprising almost 25% of all cancer deaths. Existing work in computer-aided diagnosis (CAD) has applied convolutional neural networks (CNNs) to detect and classify nodules in CT scans, with the goal of assisting radiologists diagnose lung cancer. In the past decade, new screening pro- tocols have been enacted that advise high-risk patients to get annual CT screenings to monitor any suspicious lesions found in the lungs. This change increases the availability of CT scans and the number of scans per patient for computational models to learn from. In this thesis, we present bounding box annotations for a subset of patients from the National Lung Screening Trial (NLST) over three years time and provide baseline results on the benchmark task of malignancy prediction using this time-series data. We analyze the use of longitudinal models to capture the progression of nodule malignancy and see that recurrent neural networks (RNNs) outperform standard CNNs by 4.58% in accuracy, 5.03% in precision for a fixed sensitivity of 95.06%, and 6.61% in area under the curve (AUC).