MENU

Click here to

×

Are you sure ?

Yes, do it No, cancel

Prognosis Prediction with Homology-Based Radiomic Features Quantifying the Lung Tumor Malignancy in CT-Based Radiomics

S Tanaka1*, N Kadoya1 , T Kajikawa1 , K Abe1, 2, S Dobashi3 , K Takeda3 , K Nakane4 , K Jingu1 , (1) Department of Radiation Oncology, Tohoku University Graduate School of Medicine, Sendai, Japan, (2) Department of Radiology, Tokyo Metropolitan Cancer and Infectious Diseases Center Komagome Hospital, Tokyo, Japan, (3) Course of Radiological Technology, Health Sciences, Tohoku University Graduate School of Medicine, Sendai, Japan, (4) Department of Medicine, Osaka University Graduate School of Medicine, Osaka, Japan

Presentations

(Tuesday, 7/16/2019) 4:30 PM - 6:00 PM

Room: 304ABC

Purpose: We developed homology-based radiomic features and evaluated the accuracy of prognosis prediction for non-small cell lung cancer (NSCLC) patients.

Methods: CT images of 277 NSCLC patients were recruited from The Cancer Imaging Archive. We extracted three axial images from GTV. Each axial image was converted to binarized images in which each pixel could have two values: 0 or 1. In binarized images, we calculated the two Betti numbers: b0, which is the number of isolated components, and b1, which is the number of “circular� holes. We created homology-based histograms for both b0 and b1 using the binarized images by changing the threshold of CT value (range: -150 HU to 300 HU). We summed up three histograms into a histogram for both b0 and b1. We defined five homology-based radiomic features from summed histogram; ratio of b1 to b0, maximum number of b0 and that of b1, area under histogram of b0 and that of b1. To clarify the prognostic power, the relationship between homology-based radiomic features and overall survival was evaluated by Kaplan-Meier method, univariate analysis and multivariate analysis with random forest machine learning to predict 2-year survival state.

Results: Significant differences in Kaplan-Meier curves were observed between two groups according to the median value of the five homology-based radiomic features (log-rank p value < 0.01). The highest AUCs for univariate analysis and multivariate analysis with random forest were 0.70 and 0.69 for homology-based radiomic features, respectively, whereas they were 0.68 and 0.66 for standard radiomic features, respectively, showing that the prediction performance with homology-based radiomic features had higher AUC than standard radiomic features.

Conclusion: We evaluated the performance of prognosis prediction using homology-based radiomic features. Our result showed homology-based radiomic features had the great potential for improving the accuracy of prognosis prediction in CT-based radiomics.

Keywords

Radiation Therapy, Lung

Taxonomy

IM- CT: Machine learning, computer vision

Contact Email