INTRODUCTION

Chronic liver diseases are a common source of morbidity and mortality in both children and adults in the United States and around the world. ^1,2 Detection and progression of such liver diseases is typically assessed using a combination of clinical history, physical examination, laboratory testing, biopsy with histopathologic assessment, and imaging. ³ MR elastography (MRE), a non-invasive method of assessing liver stiffness, uses an active-passive driver system (with the passive paddle placed over the right upper quadrant of the abdomen at the level of the costal margin) to create transverse (shear) waves in the liver. ⁴ Although MRE may obviate the need for liver biopsy in some patients and allows more frequent longitudinal assessment of liver health, it has associated drawbacks related to additional patient time in the scanner, patient discomfort, and added costs (e.g., infrastructure and patient charge-related). ⁵ In the current project, we aim to develop a deep learning approach to categorically classify liver stiffness-determined by MRE-using both T2-weighted imaging and clinical data from pediatric and young adult patients.

METHODS

In this retrospective study, we collected 178 MRE examinations from a GE scanner for model development and internal validation and 95 MRE examinations from a Philips scanner for external validation. For each subject, the mean liver stiffness value in kPa (shear modulus) and 27 clinical features were retrieved from the electronic health record (Epic Systems Corporation; Verona, WI). Based on pre-defined liver stiffness cutoff, ⁶ patients were divided into two groups (<3 kPa=no/mild vs. ≥3 kPa=moderate/severe liver stiffening). Axial two-dimensional T2-weighted fast spin-echo fat-suppressed images were extracted from our clinical Picture Archiving and Communicating System.

Our deep model contains two separate input channels for imaging and clinical data, respectively (Figure 1). The imaging channel is comprised of an image input layer, a transfer learning block, and an adaptive learning block. First, the image input layer contains S parallel input sub-channels, taking S individual slices of fixed-size axial T2-weighted MR images. Next, to extract liver image features, we designed a transfer learning block by reusing the weights of a VGG-19 model ⁷ (from 1st to 21st layers) that was trained based on ~1.2 million color images from ImageNet database. ⁸ Then, we designed an adaptive learning block that contains S parallel sub-channels (two convolutional layers with [8, 16] neurons and 3×3 filters, and a fully-connected layer with 8 neurons) corresponding to the input sub-channels for learning the individual latent features of S liver slices, respectively. In the end, those sub-channels in the adaptive learning block are integrated by a fully-connected layer with 8 neurons. In the current study, we used four axial T2-weighted images of the liver from the same anatomic levels as MRE images (i.e., S=4). For the clinical channel, a fully-connected layer with 8 neurons is directly applied. After the feature extraction, a fusion block (8 neurons) is applied to integrate the latent features from both imaging and clinical data. A two-way softmax classifier was utilized to classify the severity of liver stiffness.

Considering the imbalanced subject ratio (e.g., <3 vs. ≥3 kPa = ~2:1 in the current study), a rotation and shift-based data augmentation scheme ⁹ is used to balance the subjects between two groups. The diagnostic performance of the model is assessed with the metrics of accuracy, sensitivity, specificity, and area under the receiver operating characteristic curve (AuROC). We used 10-fold cross-validation in the internal validation with the internal cohort. In the external validation, we trained the model with the internal cohort and tested the model with an external cohort.

RESULTS

Demographics of two cohorts were listed in Table 1.

Internal Validation. We first set to determine the performance of DeepLiverNet using only non-stiffness T2-weighted imaging data, including liver volume and chemical shift-encoded fat fraction. As shown in Table 2, the DeepLiverNet was able to correctly classify patients with regard to categorical MRE liver stiffness with an AuROC of 0.80. Using only clinical data, the model classified patients with an AuROC of 0.83, achieving a significantly greater AuROC (p=0.003) compared to the one using only imaging data. The DeepLiverNet combining both T2-weighted MR imaging and clinical data was able to correctly classify patients with an AuROC of 0.86. This was significantly greater than imaging data alone (p<0.0001) or clinical data alone (p<0.0001).

External Validation. The internally-validated DeepLiverNet using both clinical and imaging features achieved an accuracy of 80.0%, with a sensitivity of 61.1%, a specificity of 91.5%, an AuROC of 0.77 on independent external subjects.

DISCUSSIONS and CONCLUSIONS

In this work, we applied both transfer learning and data augmentation strategies to avoid deep learning model overfitting. Our model demonstrated good generalizability when externally validated. Our proposed deep learning model that incorporated clinical features and T2-weighted MR images demonstrated a means of classifying patients into normal/minimally elevated versus moderately/severely elevated liver stiffness with an accuracy up to 88%. Further studies are needed to continue to refine the model as well as validate it in other patient groups, including older adults and cohorts with very specific liver diseases (e.g., sclerosing cholangitis, viral hepatitis, non-alcoholic fatty liver disease).

Acknowledgements

We highly appreciated the internal research grant of the Department of Radiology, Cincinnati Children's Hospital Medical Center.