On utilizing 2D features from 3D scans to enhance the prediction of lung cancer survival rates

Abstract

Accounting for 40% of diagnosed lung cancers, Adenocarcinoma is detected through Computed Tomography (CT) scans which provide images as “sliced” versions of the scans, as one progresses down the thoracic region. Typically, one extracts 2D features from the “sliced” images, and this been utilized to achieve various types of classification. In the same vein, the diagnosis and detection of Adenocarcinoma from such scans using image processing techniques, has been studied in the literature. However, as with any cancer, the survival rate of Adenocarcinoma patients depends on the spread of the cancer, the treatments the patient has undergone, and the severity of the cancer at the juncture when the treatment was initiated. This paper presents the ground reality whereby 2D features extracted from the images, or slices, can yield reasonable survival rate predictions of patients by changing the context goal from classification to a regression problem. Furthermore, the emphasis depicted in this paper revolves around the discovery of a strong correlation between the shapes of the 2D images at successive layers of the scans and the corresponding survival rates. Augmenting a classification feature set with aggregated shape-based feature computations of the tumour throughout the consecutive slices, such as the mean area along the z-axis, displays a significant improvement in the forecasting of the patients’ survival rates. We are not aware of any prior research in this area, and while these results have been proven to be relevant for lung cancer scenario, our position is that these principles are also valid for other tumour-based cancers. The dataset consists of scans from 60 different patients, with varying levels of severity, for which a wide range of survival rates was also available. A particular observation is that for patients that had a survival rate of up to 24 months (i.e., short term predictions), we achieved an error as low as 9%.