RSNA 2016

Abstract Archives of the RSNA, 2016


SSM20-05

Quantitative Texture Classification for Differentiating Fat-poor Angiomyolipoma from Renal Cell Carcinoma: Development of Feature Classification in MDCT Images

Wednesday, Nov. 30 3:40PM - 3:50PM Room: S404CD



Han Sang Lee, MS, Daejeon, Korea, Republic Of (Presenter) Nothing to Disclose
Helen Hong, PhD, Seoul, Korea, Republic Of (Abstract Co-Author) Nothing to Disclose
Dae Chul Jung, MD, Seoul, Korea, Republic Of (Abstract Co-Author) Nothing to Disclose
Seunghyun Park, Seoul, Korea, Republic Of (Abstract Co-Author) Nothing to Disclose
Junmo Kim, Seoul, Korea, Republic Of (Abstract Co-Author) Nothing to Disclose
CONCLUSION

Our method can be applied to the differentiation of various types of RCC, e.g. clear cell-type, papillary, and oncocytomas, and can be used in early diagnosis of small renal masses.

Background

Differentiating benign fat-poor angiomyolipoma (fp-AML) from malignant renal cell carcinoma (RCC) is an important task for early diagnosis of renal cancer. Since fp-AML has similar intensity distribution and heterogeneity with RCC, classifying them is considered to be a challenging problem. In this research, we propose a texture-based classification method for differentiating fp-AML from clear cell-type RCC in contrast-enhanced CT images.

Evaluation

Our method was tested on a dataset consisting of 30 volumetric renal CT scans from thirty patients from ten with AML without gross fat, and twenty with RCC. CT examinations were performed on MDCT scanners at 100s to 120s delay after contrast injection, to acquire axial plane images with a slice thickness of 1.0-3.0 mm, and a resolution between 0.66 x 0.66 mm to 0.77 x 0.77 mm. For each scan, a region of interest (ROI) for a renal mass was marked by a radiologist. From the tumor ROI of training set, 117 features consisting of 22 with gray-level histogram, 14 with gray-level co-occurrence matrix (GLCM), 22 with gray-level run-length matrix (GLRLM), and 59 local binary patterns (LBP), were extracted. Then, feature selection was performed to select useful features with high separability with the ReliefF method. Throughout the feature selection process, 70 features with 16 gray-level histogram, 7 GLCM, 12 GLRLM, and 35 LBP features were selected. Finally, a support vector machine (SVM) was trained with the training features and labels to classify the unseen test features. In evaluation, 5-fold cross validation was performed and our results were quantitatively evaluated by average accuracy rates, sensitivity, specificity, positive predictive values (PPV), and negative predictive values (NPV) of 84.0%, 55.2%, 98.5%, 95.1%, and 81.5%, respectively.

Discussion

Our feature extraction can provide complementary information for separating fp-AML and RCC in MDCT images. Our feature selection can improve the classification performance by enhancing separability of features.     This work was supported by the National Research Foundation of Korea grant funded by the Korean Government (MEST) (NRF-2015R1A2A2A04003460)