Radiomics and Deep Learning: Hepatic Applications

Abstract

Radiomics and deep learning have recently gained attention in the imaging assessment of various liver diseases. Recent research has demonstrated the potential utility of radiomics and deep learning in staging liver fibroses, detecting portal hypertension, characterizing focal hepatic lesions, prognosticating malignant hepatic tumors, and segmenting the liver and liver tumors. In this review, we outline the basic technical aspects of radiomics and deep learning and summarize recent investigations of the application of these techniques in liver disease.

Keywords: Artificial intelligence; Computer-assisted; Deep learning; Liver; Radiomics.

Fig. 1. Schematic description of morphologic features.

Area and perimeter are calculated from ROI drawn on image. Ellipse fitted to given ROI is obtained. Then, major and minor axes of ellipse and convex area are calculated. Based on these values, morphologic features (circularity, roundness, aspect ratio, solidity, compactness, and others) are calculated according to equations shown in Figure 1. Values of morphologic features for angular and rod shapes are compared with those for complete circle, which has value of 1 for all morphologic features. ROI = region of interest

Fig. 2. Schematic description of histogram features.

From ROI drawn on image, histogram of gray-scale pixel values is obtained. Then, multiple features are calculated from histogram to describe pattern of distribution of gray-level values within ROI. CV = coefficient of variation, ENT= entropy, SD = standard deviation

Fig. 3. Schematic description of textural feature extraction assuming 3 × 3-pixel image with three different gray-scale levels.

GLCM describes frequency of two neighboring pixels having certain gray-level pixel values, while GLRLM describes length of continuous pixel having certain gray-level pixel value. After aggregating different directional matrices, secondary features are calculated from matrices to describe textural pattern of given image, including CON, ENT, CORR, and HOM, and others from GLCM and SRE, LRE, LGRE, and HGRE, and others for GLRLM. CON = contrast, CORR = correlation, ENT= entropy, GLCM = gray-level co-occurrence matrix, GLRLM = gray-level run-length matrix, HGRE = high gray-level run emphasis, HOM = homogeneity, LGRE = low gray-level run emphasis, LRE = long run emphasis, SRE = short run emphasis

Fig. 4. Example images depicting effects of image filters.

Portal venous phase CT image was transformed by using Gaussian, Laplacian, and LOG filters and using wavelet transformation of high-frequency and low-frequency parts. Higher-order features are histogram and textural features extracted from these transformed images. LOG = Laplacian of Gaussian

Fig. 5. Schematic description of development process for radiomics classification model.

Model for staging liver fibrosis using gadoxetic acid-enhanced hepatobiliary phase magnetic resonance images is assumed for demonstration purposes. Input images undergo preprocessing, including normalization of gray-scale pixel values and image resampling to standardize image resolution. Radiomics features are then extracted, which may include shape, histogram, texture, and high-order features. Feature selection is performed to reduce feature dimension, and classification model is then developed using selected radiomics features. Final radiomics model is used for classification of new input images. LASSO = least absolute shrinkage and selection operator

Fig. 6. Schematic depiction of training CNN.

From input images, Conv layer extracts feature maps, and pooling layer downsizes feature maps. ReLU is usually followed by Conv layer as activation function. High-level features are extracted through multiple Conv and pooling layers, and then fed into fully connected layer. Fully connected layers integrate all features to perform classification task. In this schematic diagram, liver fibrosis staging using CT images is presented as classification task for demonstration purposes. Using softmax function, fully connected layer returns probability of each class as output. During training phase, output of CNN is compared with ground truth to calculate errors using loss function. Error is then back propagated, and weights of network are adjusted to decrease loss and thereby maximize accuracy of CNN for given classification task. CNN = convolutional neural network, Conv = convolution, ReLU = rectified linear unit