(Sunday, 7/29/2018) 2:05 PM - 3:00 PM
Room: Davidson Ballroom B
Purpose: First-order shape features such as volume, surface area and sphericity have been reported to predict cancer survival. However, they cannot reflect local information, while high-order shape features such as surface curvature are often sensitive to noises or variation of structure contouring. This study tested whether median geodesic distance (MGD) can serve as a novel and robust high-order feature to better predict survival.
Methods: We first tested the robustness of MGD by adding 5% and 10% noises to a mesh, evaluating the Relative Percent Difference (RPD) of MGD before and after adding the noises, and comparing that with other high-order shapes, namely, the mean curvature and Gaussian curvature. We then tested its capability of predicting survival in 238 non-small cell lung cancer (NSCLC) stage-III patients. For each patient, MGD was calculated at each vertex of the mesh extracted from GTV contours in planning CT. Principle component analysis was used for dimension reduction, and the first 15 components were used as the MGD features. These features were applied to a random forest tree classifier to predict the two-year survival of a patient. The accuracy was compared to the classifier using the combination of volume, surface area and sphericity as features with 5-fold cross validation.
Results: The RPDs were 141.6%, 163.3% and 1.9% for the mean curvature, Gaussian curvature and MGD respectively after adding 5% noises, and were 159.7%, 177.8% and 8.3%, respectively, after adding 10% noises. The classifier accuracy was boosted from 54.8% for the combination of volume, surface area and sphericity to 59.7% for MGD.
Conclusion: MGD can serve as a high-order shape biomarker to predict survival. It may have the benefit of combining with high-order texture feature in future study to further improve prediction.
Keywords
Not Applicable / None Entered.
Taxonomy
Not Applicable / None Entered.
Contact Email
