(Sunday, 7/29/2018) 4:30 PM - 5:00 PM
Room: Exhibit Hall | Forum 8
Purpose: This study is an attempt to classify medical images using topological features in the data. By characterizing tumors using quantified features obtained from the persistent homology, we tried to classify hepatic tumor images that are difficult to distinguish with the naked eye into hepatocellular carcinoma (HCC) and benign hepatic tumor (BT).
Methods: Topological data analysis has been developed from theoretical aspects to applications in the last decade. Persistent homology is one of the topological theory which is used widely for capturing multiscale topological features in data. In this theory, persistence diagrams are calculated as compact descriptors for characterizing those features in topological data analysis. For this retrospective study, non-contrast enhanced fat-suppressed three-dimensional T1-weighted MR images of tumor regions of 100 cases with HCC (n = 50) and BT (n = 50) were acquired during November 2009 – August 2017. For classification, persistence diagrams of three types (zero-, one- and two-dimensional) were generated for each lesion from the MR images. The persistence diagrams were vectorized to apply machine learning models. We used two models, a logistic classifier model with the elastic net penalty and extreme gradient boosting (XGBoost), to classify the feature vectors of HCC and BT and evaluate the classification abilities.
Results: The highest classification accuracy, sensitivity, specificity, and the area under the receiver operating characteristic curve were, respectively, 82%, 80%, 84% and 0.90.
Conclusion: Our method using persistent homology allows for classification of hepatocellular carcinoma and benign hepatic tumor only from non-contrast enhanced T1-weighted MR images with considerable accuracy. This method may be useful when utilized in the computer-aided diagnosis of hepatic tumors with MR imaging.
Keywords
Taxonomy
Contact Email
