Radiomics in liver diseases: Current progress and future opportunities

Abstract Liver diseases, a wide spectrum of pathologies from inflammation to neoplasm, have become an increasingly significant health problem worldwide. Noninvasive imaging plays a critical role in the clinical workflow of liver diseases, but conventional imaging assessment may provide limited information. Accurate detection, characterization and monitoring remain challenging. With progress in quantitative imaging analysis techniques, radiomics emerged as an efficient tool that shows promise to aid in personalized diagnosis and treatment decision‐making. Radiomics could reflect the heterogeneity of liver lesions via extracting high‐throughput and high‐dimensional features from multi‐modality imaging. Machine learning algorithms are then used to construct clinical target‐oriented imaging biomarkers to assist disease management. Here, we review the methodological process in liver disease radiomics studies in a stepwise fashion from data acquisition and curation, region of interest segmentation, liver‐specific feature extraction, to task‐oriented modelling. Furthermore, the applications of radiomics in liver diseases are outlined in aspects of diagnosis and staging, evaluation of liver tumour biological behaviours, and prognosis according to different disease type. Finally, we discuss the current limitations of radiomics in liver disease studies and explore its future opportunities.


| INTRODUC TI ON
Liver diseases, a wide spectrum of pathologies from inflammation to neoplasm, have become a major health problem worldwide.
Noninvasive imaging plays a critical role in the characterization and monitoring of liver diseases. Conventional ultrasound, computed tomography (CT) and magnetic resonance imaging (MRI) are widely used for qualitative evaluation of liver morphology and blood supply. [1][2][3] Tremendous progress is still being made in liver imaging with introduction of advanced techniques, including metabolic imaging, molecular imaging, and multi-parametric functional MRI, etc, allowing improved evaluation of liver diseases and assisting personalized medical decision making. [4][5][6] With accumulation of scalable liver imaging data, radiomics emerges as a novel radiological technique that comprehensively utilizes large-scale medical imaging into the process of liver disease management via artificial intelligence techniques. 7,8 It enables extraction of high-throughput quantitative imaging features beyond inspections of naked human eyes and converting encrypted medical imaging into minable numerical data. 8 Combined with clinical, pathological, or genetic information, radiomics would assist in lesion characterization, preoperative diagnosis, treatment efficacy evaluation, as well as prognosis prediction in various clinical settings. [9][10][11] Quantitative imaging traits were proved to be associated with global gene expression programmes, and could reconstruct 78% of the global gene expression profiles in liver cancer. 12 This groundbreaking result laid a foundation and greatly encouraged researchers to explore the potential of quantitative imaging tool in preoperative genetic/pathological outcome prediction. Hence, a great deal of radiomics studies have been conducted using multi-parametric and multi-modality imaging in terms of liver disease diagnosis and treatment decision making.  In certain scenarios, this artificial intelligence-based technique could even compete pathological gold standard, providing new ways for unsolved clinical problems in the paradigm of liver disease management. 16 Nevertheless, it still requires further multi-centre and prospective validation for the validity of radiomics. The interpretability and the correlation with biological/pathological underpinnings also represent substantial obstacles for the translation of artificial intelligence into real clinical practice.
Here, we review the basic concepts of radiomics methodologies specific for liver studies from data acquisition, liver/lesion segmentation, feature design, to model construction ( Figure 1). Meanwhile, representative clinical applications of radiomics in liver diseases regarding diagnosis, staging, evaluation of liver tumour biological behaviours, and prognosis are also within the scope of this study.
Finally, we summarize the current challenges and limitation of radiomics, and explore its future directions in liver diseases.

| Data acquisition and curation
Data used in radiomics studies can be single-centre or multi-centre, and retrospective or prospective. Here, we searched PubMed (8 October 2019) for radiomics studies on liver diseases using terms (liver diseases AND radiomics), and found 36 clinical target-oriented published work.  Most (33 out of 36) studies were performed on single-centre with retrospective cohort, while only two studies were performed on multi-centre and prospective cohort (Table 1). And the most commonly used imaging modality was CT (18 studies), followed by MRI (12 studies), positron emission tomography (PET) (two studies) and ultrasonography (US) (four studies) ( Table 1).
Considering the effect of inconsistent imaging acquisition protocol and reconstruction procedure in multi-centres via multi brand manufactories, preprocessing of the collected imaging data is required. Currently, the most commonly used methods conclude resampling and intensity normalization. Image resampling is used to improve image quality and eliminate bias introduced by non-uniform imaging resolution. 49,50 Image intensity normalization is utilized to correct inter-subject intensity variation by transforming all images from original greyscale into a standard greyscale. 51,52 Park et al normalized liver signal intensity according to the spleen signal on hepatobiliary phase (HBP) images to extract high-order textural features and revealed the improved diagnostic value as compared with non-normalized data. 29 In addition to imaging data, clinical factors were also involved in radiomics analysis, including patient age, gender, Child-Pugh stage, histologic grading, BCLC stage, cirrhosis and its cause, etc. 13

Key points
• Radiomics as an emerging technique based on medical imaging analysis is more commonly used in liver disease studies.
• Inter-personal heterogeneity could be revealed via extracting high-dimensional quantitative imaging features and analysed by artificial intelligence algorithms.
• Radiomics can be applied in the diagnosis, treatment effect evaluation and prognosis prediction in liver diseases.

| Region of interest segmentation
Segmentation of region of interest (ROI) could be divided into manual segmentation and semiautomatic/automatic segmentation. Most radiomics studies on liver disease applied manual segmentation. Only six studies performed semiautomatic/automatic segmentation. 17,30,39,46,53,54 Manual segmentation is performed by radiologists to annotate the location and precise boundary of the lesion. Another way of manual segmentation is realized by placing a rectangular/circle box via deep learning analysis. Wang et al conducted a squared ROI segmentation as the input of convolution neural network (CNN) and achieved satisfying performance in liver fibrosis stage prediction. 16 Naganawa et al applied similar segmentation approach with a 2-cm diameter circular ROI covering the lesion while excluding intrahepatic vessels. 15 Considering the discrepancy of subjective judgement in manual segmentation, segmentations by multi-clinicians, of multi-time point, and using computer perturbation are required to decrease the intra-and inter-reader variability. 32 Feature reproducibility and robustness are generally evaluated through calculation of intra-class correlation coefficient and concordance correlation coefficient. 36,56,57 Automatic segmentation aims to annotate ROIs by computer automatically, whereas semiautomatic segmentation still needs partial manual intervention to mark the centre of the lesion before automatic segmentation. Several classic segmentation algorithms showed good performance in liver lesion annotation. 58  In addition, filtered features are extracted from ROI preprocessed by wavelet, Laplacian and Gaussian filters from multiple dimensions. 62 Commonly used manual engineered features are shown in Table 2

| Task-oriented modelling
Generally, the methods for feature selection conclude filter-based, Among the aforementioned methods, filter-based methods require less computation time than the other two methods but with lower prediction accuracy. Thus, they are most commonly used as a primary selection method to initially reduce features. 23,55 Regarding modelling strategy, radiomics studies on liver disease mostly utilized supervised learning modelling. LASSO logistic regressing modelling was commonly used, demonstrating satisfying performance particularly in small sample size based studies. 22,31,72 Support vector machine and random forest were also used in published liver disease radiomics studies. 19 Their result indicated that adaptive boosting, random forest and support vector machine stood out as superior modelling methods with improved accuracy for fibrosis prediction.

| R AD I OMI C S IN THE D IAG NOS IS AND S TAG ING OF LIVER D IS E A S E S
For clinical application, radiomics plays a pivotal role in the diagnosis, staging and grading of several liver diseases, of which most efforts focused on hepatic malignancies and liver diffuse diseases (Figure 2).

| Liver diffuse diseases
Besides hepatic malignancies, radiomics also showed potential in characterization of liver diffuse diseases including fatty liver diseases and liver fibrosis. The first study evaluating the performance

| R AD I OMI C S IN THE E VALUATI ON OF LIVER TUMOUR B IOLOG IC AL B EHAVIOUR S AND PROG NOS IS
Beyond diagnosis and staging, radiomics enables quantitative assessment of liver tumour biological behaviours, as well as prediction of prognosis and antitumoral treatment effect (Figure 2).

| Measurement of tumour differentiation and proliferation
Histologic grade was one of the most important risk factors for postoperative recurrence in HCC. 80

| Assessment of tumour vascular invasion
Preoperative discrimination between neoplastic and bland portal vein thrombosis and detection of microvascular invasion in HCC is critically important. 86

| Prediction of treatment efficacy and prognosis
Radiomics analysis permits accurate prediction of prognosis and effective diverse therapy evaluation. 73,90 Several studies were conducted for hepatic resection evaluation, and one study was for liver transplantation evaluation. 13

| ICC
ICC is an aggressive primary hepatic cancer arising from the bile duct epithelium. 97 However, unlike HCC, surgical resection is currently the only curative treatment for ICC patients. 98

| Metastatic hepatic malignancies
In addition to primary liver cancers, radiomics also showed promise in the evaluation of several metastatic hepatic malignancies.

| FUTURE CHALLENG E S AND OPP ORTUNITIE S
Current published studies revealed the potential of radiomics analysis in liver disease diagnosis, tumour biological property profiling, and F I G U R E 2 Illustration of clinical application of radiomics on liver diseases prognosis estimation. However, although MR imaging can provide the multi-parametric information regarding hepatic function and microenvironment with higher tissue resolution, most studies to date have focused on radiomics analyses of CT. [103][104][105][106] In addition, a large number of studies were retrospective in design and lack independent external validation across different geographical areas and races, which may limit the generalizability and applicability of the current findings.
Different prevalence of disease may also influence the accuracy of the algorithm (eg positive and negative predictive values). Moreover radiomics results are extremely sensitive to the various technical acquisition parameters, especially among different vendors. Therefore, more large scale multi-centre prospective studies with standardized acquisition, segmentation and imaging postprocessing are needed to ensure further development of radiomics in liver diseases.

| CON CLUS IONS
Radiomics as a newly emerged quantitative technique is burgeoning in liver disease management with consistently developing methodology.
Previous studies, although mainly retrospective in design and based on single imaging modality, have revealed its potential in diagnosis, treatment evaluation and prognosis prediction of several liver diseases. Nevertheless, further multi-centre and prospective validation is still needed to valid its clinical usefulness, especially in prognosis-related targets.
Current main obstacles for the application of radiomics in liver disease rely on high-quality data collection and mechanism explanation on the biological basis. Multi-institutional data sharing and intensive collaborations on data cleansing and labelling offer appeal in filling this gap. Artificial intelligence algorithms with improved accuracy and interpretability meanwhile need to be developed to facilitate broader translation and clinical adoption.

ACK N OWLED G EM ENTS
The authors appreciate the study participants, as well as researchers and staff.