(Monday, 7/30/2018) 9:30 AM - 10:00 AM
Room: Exhibit Hall | Forum 2
Purpose: Feature extraction is crucial for tumor segmentation in PET. Though being carefully designed, some hand-crafted features, such as HOG, LBP, Gabor features and so on, have only limited discriminative ability. It is natural to find a way to enhance these hand-crafted features. We assumed that a hand-crafted feature can be moved to a correct position (with a higher discriminative ability) in the feture space by adding an offset and the offset can be learned. We proposed a novel method to promote the discriminative power of the hand-crafted features by learning their Feature Offset via convolutional neural networks (CNNs) for PET tumor segmentation.
Methods: The Feature Offset can be learned by a feedforward neural network and the feature itself can be added to its offset by a shortcut connection in the network. Gray-level, Gradient, HOG, Gray-level Histogram features were examined. After the Feature offset learning model was learned, the original hand-crafted features were sent to the model to get their enhanced versions. SVM classifier was adopted for PET segmentation with features enhanced or not. Dice similarity index (DSI), Classification error (CE) and Volume error (VE) were used for performance evaluation. 14 non-small cell lung cancer (NSCLC) PET images were used as training samples and other 26 NSCLC images were used as testing samples. To further validate the effectiveness of the Feature Offset Learning framework, CIFAR-10 dataset and Caltech dataset were also tested.
Results: When using the original feature for tumor segmentation, the mean DSI of the 26 testing samples was 0.8068. After features were enhanced by offset, the mean DSI reached 0.8373.
Conclusion: The proposed Feature Offset Learning network showed its strong ability to promote traditional features’ representation capacity and their discriminative power, and can improve the accuracy of PET tumor segmentation.
Funding Support, Disclosures, and Conflict of Interest: This work was supported in part by National Natural Science Foundation of China (NNSFC), under Grant Nos. 61375018 and 61672253.
Keywords
Not Applicable / None Entered.
Taxonomy
Not Applicable / None Entered.
Contact Email
