Start Your Search
O05 - Cancer Control (ID 130)
- Event: WCLC 2013
- Type: Oral Abstract Session
- Track: Prevention & Epidemiology
- Presentations: 1
- Moderators:N. Van Zandwijk, A. McWilliams
- Coordinates: 10/28/2013, 10:30 - 12:00, Bayside Auditorium A, Level 1
O05.02 - Local, Surrounding and Global Features for Improved Computer Aided Diagnosis of Lung Cancer (ID 318)
10:30 - 12:00 | Author(s): J.C. Sieren
The National Lung Screening Trial reported a 20% reduction in lung cancer mortality achieved through low dose computed tomography (CT) screening of the at risk population, compared to screening with chest x-ray. Challenges with clinical implementation of CT screening for lung cancer include the high number of lesions detected that require further follow-up, approximately 97% of which are ultimately diagnosed as benign. A computer-aided diagnosis (CAD) tool can be designed to determine the probability of malignancy of a lung nodule based on objective measurements. While current CAD tools examine the pulmonary nodule’s shape, density, and border, analyzing the lung parenchyma surrounding the nodule is an area that has been minimally explored. By quantifying characteristics, or features, of the surrounding tissue, this study explores the hypothesis that textural differences in both the nodule and surrounding parenchyma exist between malignant and benign cases, which can be utilized to improve CAD performance.
From CT data, several novel feature extraction techniques were developed, including a three-dimensional application of Laws’ Texture Energy Measures to quantify the textures of the parenchyma as well as the nodule. In addition, the densities of the nodule and parenchyma were summarized through metrics such as mean, variance, and entropy of the intensities. The margins of the nodule were characterized following ray casting and rubber-band straightening to analyze mean and variance of border irregularity. Basic demographics and risk factor data were also included. The large feature set was reduced by statistical testing and stepwise forward selection to a few independent features that best summarize the dataset. A neural network was used to classify the cases in a leave-one-out method.
To illustrate proof of concept, the CAD tool was applied to 27 lung nodule cases: 10 malignant and 17 benign. These data were diverse with regards to data acquisition protocol, reconstruction kernel and slice thickness – all of which can pose challenges to CAD. Through statistical testing, 36 features were found to be significant predictors of malignancy (p < 0.05), including many textural and parenchymal features. Two of these significant features, selected through stepwise forward selection, were utilized to classify the data: nodule variance (p = 0.0003) and parenchyma median intensity (p = 0.0028). CAD performance achieved a sensitivity of 90%, specificity of 100%, and an accuracy of 96.3%.
Preliminary findings indicate features from both the nodule and the surrounding parenchyma have value in distinguishing benign and malignant lesions. This is particularly valuable in the analysis of early detected, small pulmonary lesions (<10mm). In these small lesions, standard CAD approaches are hindered by few CT data voxels contained within the lesion. By incorporating local, surrounding and global features, more information is included and augmented CAD performance may be achieved. Finally, many significant features were identified despite diversity in the CT data acquisition parameters which indicates the suitability of the approach to broad clinical application. We are currently working on applying the CAD tool to a larger dataset.
Only Active Members that have purchased this event or have registered via an access code will be able to view this content. To view this presentation, please login or select "Add to Cart" and proceed to checkout.