Program Information
Focal Liver Lesion Detection and Classification in Non-Enhanced T2-Weighted by Means of Wavelet-Based Hierarchical Cluster Analysis and Probabilistic Neural Network
I Gatos , S Tsantis , M Karamessini , G Kagadis*, University of Patras, Rion, Ahaia
Presentations
MO-F-CAMPUS-IT-3 (Monday, July 31, 2017) 4:30 PM - 5:30 PM Room: Imaging ePoster Theater
Purpose: To segment and characterize focal liver lesions(FLLs) on nonenhanced T2-weighted magnetic resonance imaging(MRI) scans using a computer-aided diagnosis(CAD) algorithm.
Methods: Clinical dataset includes 71 FLLs (30–benign FLLs, 19–Hepatocellular Carcinomas(HCC) and 22–liver metastases). All three classes’ diagnosis was established by means of typical signs on dynamic contrast-enhanced MRI and liver biopsies. A novel lesion detection algorithm is introduced within this study that comprise at first a preprocessing step towards edge information extraction that provides a rough approximation of the lesion contour. Subsequently, the mean pixel values of regions between and outside contour positions are computed and employed as input into an unsupervised Hierarchical Cluster Analysis(HCA) for final FLL Segmentation. HCA analysis optimizes clusters selection by employing the maximum dissimilarity distance between the lower linkage level and the upper linkage level of each cluster group.For each segmented lesion, texture information is derived using 42 features from the gray-level histogram, co-occurrence and run-length matrices, whereas shape information employing 12 morphological features. Feature selection was performed utilizing Stepwise Regression Analysis(SRA) leading to a reduced feature subset. A multiclass PNN classifier is then designed towards lesion classification. PNN model evaluation was performed by means of leave-one-out method for all possible feature combinations from the selected subset.
Results: Segmentation accuracy of FLLs employing overlap degree between automatic contour and manual segmentation performed by radiologists was 0.92+/-0.14. Maximum classification accuracy for all three classes (90.1%) was obtained using three textural features (inverse different moment, sum variance, and long run emphasis) which describe a lesion’s textural contrast and variability. Sensitivity/specificity values for all three classes were of 90%/97.3%, 89.5%/92.2%, and 90.9%/95.6%, respectively.
Conclusion: The proposed CAD system employing a sophisticated segmentation algorithm and a powerful classification model exhibited promising results, and can supplement radiologists decision making, towards patients’ number decrease undergo invasive procedures.
Contact Email: