Enhancing the Prediction of Lung Cancer Survival Rates Using 2D Features from 3D Scans

Abstract

The survival rate of cancer patients depends on the type of cancer, the treatments that the patient has undergone, and the severity of the cancer when the treatment was initiated. In this study, we consider adenocarcinoma, a type of lung cancer detected in chest Computed Tomography (CT) scans on the entire lung, and images that are "sliced" versions of the scans as one progresses along the thoracic region. Typically, one extracts 2D features from the "sliced" images to achieve various types of classification. In this paper, we show that the 2D features, in and of themselves, can be used to also yield fairly reasonable predictions of the patients' survival rates if the underlying problem is treated as a regression problem. However, the fundamental contribution of this paper is that we have discovered that there is a strong correlation between the shapes of the 2D images at successive layers of the scans and these survival rates. One can extract features from these successive images and augment the basic features used in a 2D classification system. These features involve the area at the level, and the mean area along the z-axis. By incorporating additional shape-based features, the error involved in the prediction decreases drastically -by almost an order of magnitude. The results we have obtained deal with the cancer treatments done on 60 patients 1 at varying levels of severity, and with a spectrum of survival rates. For patients who survived up to 24 months, the average relative error is as low as 9%, which is very significant.

Novel Block Diagonalization for Reducing Features and Computations in Medical Diagnosis

“…-Although a lot of work has been done when it concerns the diagnosis of lung cancer, the work related to the survival times and their correlation to the size/shape of the tumor is relatively unexplored. In an earlier paper [10], we had shown that by a regression analysis, we can predict the survival times based on various features of the tumor. This paper builds on those results to use ensemble machines and block-diagonal phenomena.…”

“…In our application domain [9], [10], the dimensionality of the feature vector is 110. Computing the eigenvalues and eigenvectors of such a large matrix is, certainly, time consuming.…”

“…However, if we can push this envelope one step further and are able to judge the severity of a nodule, the prognosis and determination of treatment plans can be adjusted to yield a greater chance of success. The results we have obtained [10] deal with the cancer treatments done on 60 patients 2 at varying levels of severity and with a spectrum of survival rates. For patients who survived up to 24 months, the average relative error is as low as 9%, which we believe is very significant.…”

“…-First Order Statistics: This feature class [9], [10], focused on computing features that described the histogram of the nodule image, i.e., the grey-level intensity distribution etc., amounting to a total of 19 features. -3D Shape-based: These 16 features [9], [10], unlike the rest of the feature classes, computed the values from the mask of the given nodule as they were independent of the grey level intensities. The Pyradiomics library built a triangle mesh from the provided mask using the marching cubes algorithm.…”

“…The Pyradiomics library built a triangle mesh from the provided mask using the marching cubes algorithm. -Grey Level Cooccurrence Matrix (GLCM): This matrix was computed based on the probability function where the (i, j) th element represents the number of times grey levels i and j appeared next to each other in the image, and led to a total of 24 features [9], [10].…”