(Thursday, 8/2/2018) 7:30 AM - 9:30 AM
Room: Karl Dean Ballroom B1
Purpose: To predict lung nodule malignancy by combining handcrafted features (HFs) and natural features learned at the output layer by deep convolutional neural network (CNN).
Methods: The dataset includes 431 malignant nodules and 795 benign nodules extracted from LIDC/IDRI database with malignancy ratings and contours given by four radiologists. We first extracted 29 handcrafted features from contoured nodules, including nine intensity features, eight geometric features and 12 texture features based on grey-level co-occurrence matrix (GLCM) averaged from thirteen directions. Then 3D CNNs modified from three state-of-the-art 2D CNN architectures (AlexNet, VGG-16 Net and Multi-crop Net) were trained to extract CNN-based features. For each 3D CNN, features at the output layer combined with the 29 handcrafted features were used as the input for support vector machine (SVM) coupled with a sequential forward feature selection (SFS) method to construct classifiers. The combining classifiers takes advantage of both handcrafted features and CNN. On one hand, it can overcome the limitation of handcrafted features that may not fully reflect unique characteristics of a particular nodule by combining CNN-extracted intrinsic features. On the other hand, it may solve the challenge of CNN that requires a large scale annotation dataset.
Results: The areas under receiver operating characteristic curve (AUCs) for combined HF-CNN models based on AlexNet, VGC-16 Net and Multi-crop Net are 0.92, 0.93 and 0.92, respectively. The AUC of SVM based on HFs alone is 0.91. The average AUC of CNN without HFs for three aforementioned architectures is 0.90.
Conclusion: We proposed an algorithm combining the CNN features learned at the output layer and the handcrafted features (HF-CNN) for predicting the lung nodule malignancy. The combined HF-CNN models outperform models based on handcrafted features or CNN alone.
Keywords
Taxonomy
Contact Email
