Texture features from computed tomography correlate with markers of severity in acute alcohol-associated hepatitis

Abstract

The aim of this study was to use texture analysis to establish quantitative CT-based imaging features to predict clinical severity in patients with acute alcohol-associated hepatitis (AAH). A secondary aim was to compare the performance of texture analysis to deep learning. In this study, mathematical texture features were extracted from CT slices of the liver for 34 patients with a diagnosis of AAH and 35 control patients. Recursive feature elimination using random forest (RFE-RF) was used to identify the best combination of features to distinguish AAH from controls. These features were subsequently used as predictors to determine associated clinical values. To compare machine learning with deep learning approaches, a 2D dense convolutional neural network (CNN) was implemented and trained for the classification task of AAH. RFE-RF identified 23 top features used to classify AAH images, and the subsequent model demonstrated an accuracy of 82.4% in the test set. The deep learning CNN demonstrated an accuracy of 70% in the test set. We show that texture features of the liver are unique in AAH and are candidate quantitative biomarkers that can be used in prospective studies to predict the severity and outcomes of patients with AAH.

Figure 1

Results from the Recursive Feature Elimination using Random Forest (RFE-RF) algorithm, used to identify key CT liver texture features to differentiate between AAH and control patients. A total of 178 texture features, and a combination of 23 of these features were associated with the best model performance (85.4% accuracy in distinguishing AAH from control patients). These 23 texture features are listed in Table 3. The 23-feature model applied to the left-out test set is described in the bottom-right table and resulted in an accuracy of 82.4% (14 of 17 patients), sensitivity of 100%, specificity of 75%, NPV of 100%, and PPV of 62.75%.

Figure 2

Examples of elastic-net regression outputs for three texture features. For each texture feature, 13 clinical variables were used as predictors. The top elastic-net clinical predictor for each texture feature was identified for testing in a simple linear regression model.

Figure 3

Results for deep learning accuracy and loss over 200 epochs for both the training and validation datasets. The model accuracy in the training set reached 95% accuracy, with overall accuracy of 70%. The validation set reached 100% accuracy. Loss improved as expected over 200 epochs, with the lowest loss of the model’s performance on the test set of 0.56.

Figure 4

Flowchart demonstrating the inclusion and exclusion criteria used to create our cohorts of AAH and control patients.

Figure 5

Example of the texture analysis pipeline, which is identical for both AAH (left) and control (right) patients. The chosen CT axial slice was selected to be at the level of the right portal vein bifurcation for each patient. The liver was manually segmented from the axial slice. Texture features were then extracted from liver segments. Maps of representative texture features (variance and homogeneity) are shown.

Figure 6

Illustration of the DenseNet architecture used to create a 2D convolutional neural network (CNN) to automatically detect acute AAH. The image of the liver segment undergoes successive layers of 2D convolution, batch normalization, rectified linear unit (ReLU) activation and 2D max pooling. The feature maps for each layer were concatenated with those of the previous layer. The last layer was fully connected with a ReLU activation function, followed by random dropout and prediction.