Automating Patient-Level Lung Cancer Diagnosis in Different Data Regimes

Adam Pardyl, Dawid Rymarczyk, Zbislaw Tabor, Bartosz Zielinski

Abstract: As the leading cause of cancer-related mortality, lung cancer is responsible for more deaths than colon, breast, and prostate cancer put together. Screening with low-dose computed tomography detects cancer at an early stage and reduces mortality. However, it requires the tedious work of radiologists to obtain malignancy scores, which additionally are very subjective. That is why many researchers worked on methods automating lung cancer classification, usually using the publicly available LIDC-IDRI dataset for training. However, most of those methods consider only node-level classification and provide poor results for patient-level diagnosis. In this paper, we fill this gap by introducing an end-to-end methods with a CT scan on the input and the patient-level diagnosis on the output. We consider three approaches for three different data regimes to examine how stronger and weaker supervision influences the model performance.