Introduction
Despite improvements in global health [
1], incidence of liver disease continues to rise, with deaths due to hepatic conditions increasing by 400% since the 1970s (British Liver Trust -
https://britishlivertrust.org.uk/), making it the leading cause of death in those aged 35–49 years in the UK (ONS 2019 -
https://www.ons.gov.uk/). Significant progress has been made in recent years in the use of non-invasive imaging methods to measure the pathological changes that are features of increasingly common liver conditions. This includes non-alcoholic fatty liver disease (NAFLD) [
2,
3], fibro-inflammation [
4,
5] and fibrosis [
6]. The prevalence of these conditions, associated with obesity, insulin resistance and type-2 diabetes (T2D), are likely only to increase further given the current obesogenic environment. New approaches are needed to differentiate between those with mild disease, compared with those at risk of more significant conditions (cirrhosis/end stage liver disease), and particularly those who may experience accelerated disease processes [
7]. One potential approach to address these issues is the implementation of novel morphometric methods to gain a deeper understanding of the processes underpinning the development and progression of many clinical conditions [
8]. For instance, investigating whether changes beyond simple volume or fat measurements, such as liver shape, are associated with particular environmental risk factors, or whether they can be differentially related to the aetiology of a particular condition. These methods may potentially provide insight into different mechanisms of disease development and enable optimised treatment strategies to be developed.
Automated segmentation of the liver to produce image-derived phenotypes (IDPs) such as volume or fat deposition measurements are becoming more commonplace at scale as deep learning methods gain traction [
9]. While these methods enhance our understanding of the liver at a population level, they are limited when it comes to providing additional knowledge regarding morphological, functional and regional variation in response to a particular condition.
Mapping organ segmentations to a standardised three-dimensional (3D) surface mesh, enables many thousands of measurements relating to variation in organ shape to be performed using statistical parametric maps (SPMs). A similar widely applied technique is statistical shape analysis, which transforms the 3D surface mesh measurements into a smaller number of principal components, known as shape parameters, has been used to characterise variations in organ shape across a population. These approaches have been successfully applied in neuroimaging [
10,
11], abdominal computer tomography (CT) images [
12,
13], and cardiac imaging [
14,
15] and they have shown to be useful in identifying genetic interactions with cardiac pathology [
16] and brain ageing [
17]. However, they have been less frequently applied to abdominal organs, where morphological changes are known to take place in a variety of clinical conditions [
18,
19].
In the current study we have applied SPM methods to determine morphological variations in the liver and their potential association with anthropometric traits and clinical conditions. We further investigated whether the emerging 3D liver mesh-derived phenotype can add value to the prediction of disease outcomes. Our study made three main contributions. The first contribution is that we investigate the impact of the population size and the robustness of the liver template construction. Specifically, we investigated how the template image and statistical parametric mapping are affected, providing valuable insights into determining the optimal number of subjects for the liver template to represent the broader cohort. We also examined the relevance of different participant samples in the template construction process. The second contribution and also a novelty of our work is that we extend the SPM method to the domain of liver image analysis. Here, we delve deeper into the application of SPM in liver image analysis and applied it to the UK Biobank dataset, which comprises a large-scale population-based cohort, resulting in increased statistical power. Through the linear regression model, we examined the impact of anthropometric, phenotypic and clinical conditions on regional geometry of the liver and visualised these findings on the template surface mesh. The third contribution is that we extracted shape features derived from the 3D mesh-derived phenotype by dimensionality reduction and evaluated whether these shape features were better predictors of disease outcomes than the conventional measurement of liver volume.
Discussion
In this study, we mapped local shape variations across the liver and determined how these changes were associated with anthropometric, phenotypic and health traits. To achieve this we constructed surface meshes from liver segmentations of 33,434 participants from the UK Biobank. Previous studies using similar SPMs have suggested that this is a useful technique in neuroimaging [
10] and cardiac imaging [
14], enabling the associations between phenotypic and genetic variation in specific anatomical regions to be mapped [
16].
We constructed a representative liver template, and showed that a 200-participant template was sufficient to represent the broader cohort. Indeed, the number of participants included in the template construction did not impact the power of the statistical analysis across a 500-participant test cohort, or a second cohort of 479 participants with liver disease. This is in line with previous studies that found a cohort with 100 participants was sufficient to construct a representative cardiac template to investigate the shape of the left ventricle [
29].
Liver size has been explored extensively using a variety of approaches from autopsy measurements [
39], CT [
40], ultrasound [
41], and MRI [
19], as well as regression-based algorithms designed to predict liver size based on body surface area [
42]. Given accurate assessment of liver volume is essential for many aspects of hepatic surgery and determining disease progression [
43], suitable methods are needed. However, until recently, the manual annotation required to make true volumetric measurements of the liver from CT and MRI images has been extremely time consuming. Imaging studies tended to rely on more easily measured metrics, such as liver span or diameter [
44,
45], or calculation of volume indices from the measurement of multiple diameters [
46]. Consequently, these approaches limit in depth morphometric assessment and only provide information associated with overall changes to liver size or volume. The SPM method implemented in the current study demonstrates significant regional variations in liver shape associated with anthropometric variables and disease status, including simultaneous inwards and outwards adaptations. These novel phenotypic variables may be useful in longitudinal population studies, as well as determining trajectories of progression in aggressive clinical conditions, including monitoring liver cirrhosis and hepatic oncology.
While studies of liver volume have generally focussed on patient populations, there is increasing interest in understanding how hepatic volume and form is influenced by age, anthropometry and metabolic markers in the wider population [
9,
46]. Despite this, few studies employ methods that enable precise measurements of these parameters, particularly with regard to regional variation in liver shape and size. In the present study we observed that decline in the liver S2S distances were associated with increasing age. This is in agreement with previous observations, by ourselves and others, that overall liver volume decreases with age [
9,
41,
47]. However, there are some ultrasound reports suggesting liver size increases with age [
44]. This discrepancy may relate to variations in methodology since ultrasound measurements of liver diameter may not reflect overall changes in liver volume. This clearly reinforces the importance of absolute volumetric measurements, which, when combined with statistical parametric mapping, enables simultaneous extraction of global and local changes.
Additionally, we found a strong and distinct regionality in liver morphometry which was associated with body composition and liver PDFF. Specifically, we found that higher BMI and WHR were strongly associated with positive S2S distance, in line with others who have reported a positive correlation between liver size and anthropometric variables [
45,
46]. We also found that higher liver PDFF was significantly associated with positive S2S distances, suggesting that hepatic fat is associated with both liver size and shape, with some clear regional variations. We further explored whether the time of day the participants were scanned was associated with S2S distances, given we have previously shown this to be associated with fluctuations in liver volume [
9]. However, we did not find a measurable effect.
We investigated whether conditions with known involvement of hepatic function had discernible effects on our S2S measurements. For this we selected T2D, commonly associated with increased deposition of liver PDFF, and subjects with known liver conditions, which we expected to be associated with a more adverse phenotype. We found that T2D was associated with outward shape variations in the liver after adjusting for PDFF, suggesting that T2D affects liver morphology. It is well recognised that T2D is associated with a range of liver conditions, with the prevalence of NAFLD in patients with T2D reported to be 55% and NASH 37.3% [
48], substantially higher than the proportion of individuals in the general population with NAFLD (19.9%) [
3] or NASH (2.2%) [
49]. Given the clinical heterogeneity of our current T2D cohort, in terms of time of diagnosis and medication, as well as the possibility of collider bias or reverse confounding, it is impossible to identify causal mechanisms for the observed variation in S2S distances. Interestingly when we considered the interaction between age and disease, we found no statistically significant interaction for liver disease, but there was a significant interaction between age and the presence of T2D. We also considered whether the interaction between disease and the presence of liver PDFF was associated with S2S distances. Moreover, the variations covered a larger proportion of the liver in T2D compared with liver disease. This may suggest that the hepatic tissue in T2D retains its overall relative plasticity (i.e. less fibrotic-cirrhotic tissue), while in liver disease there may be regions that have reduced capacity to accumulate fat or lost their plasticity and thus be less responsive to geometrical changes. Further work is needed to determine how these changes may be utilised to improve diagnosis or monitoring of disease progression. Future work in patients with biopsy-characterised hepatic tissue should help to shed light on the heterogeneity of response to the interaction between liver fat accumulation and liver health status.
We further identified regional variations in liver morphometry that are associated with liver disease. Specifically, we observed an inward shape variation at the anterior part of the right lobe, and posterior parts of the left and right lobes accompanied by an outward increase in liver S2S distances in the anterior part of the left lobe in participants diagnosed with liver disease. Previous studies have suggested that statistical shape modelling is a viable approach for enhancing the understanding of the liver shape variations linked to the stage fibrosis and even predicting it [
13,
50]. With limited outcome and longitudinal data in the current study, the clinical significance of these changes, particularly the simultaneous regional inward and outward deformations in S2S distances are unclear. However, histological and radiological studies of the liver in patients with cirrhosis have shown that the degree of volume reduction and fibrosis is greater in the right lobe compared to the caudate lobe (which reportedly expands) [
51]. This suggests regional changes in S2S distances may reflect physiological processes in the liver. It is well established that many diseases do not progress uniformly across the liver, with differences reported within different zones (periportal, mid-lobular and pericentral) of the liver lobule, which may reflect populations, different cell types, metabolic function and differences in blood flow [
52]. Whilst it is premature to adjudicate a mechanism responsible for the changes described in the current study, the regional shape differences associated to both AST:ALT and FIB-4, hinting at hepatocellular changes underpinning the variation in S2S distances.
We assessed the predictive performance using shape features derived from the S2S distances on the case-control cohorts with liver disease and T2D. We aim to determine whether these shape features can add to prediction of disease beyond those obtained using conventional volumetric measurements. We demonstrated that the model using the shape features of the S2S distances improved the prediction of liver disease, however, there was no improvement in T2D compared to the model with liver volume. Our methods using the shape features, particularly in which histology is available, may provide additional information to confirm the utility of our approach in monitoring disease and potentially predicting outcomes. This in turn would open up the possibility of applying this methodology, in conjunction with other techniques to determine and predict the overall trajectory of progression of disease and identify those subjects requiring closer monitoring and more aggressive forms of treatment. Future work is also needed to explore variations in liver morphometry by condensing the entire coordinate matrix or deformation matrix into most distinct principal component modes to categorise population variations, which could be used in genetic association studies to enhance our understanding of chronic liver disease [
17,
53].
Our study was not without limitations. To ensure sufficient numbers of participants in the liver disease group, we included all participants in the imaging cohort who had a diagnosis of liver disease, regardless of aetiology (alcoholic, toxic and inflammatory liver disease, hepatitis, fibrosis and cirrhosis). This precludes us from a more in-depth granular analysis, although our data does suggest that hepatocellular damage, particularly in more advanced disease stages, resulted in significant S2S changes across the liver. Variation in disease aetiology, the point of disease progression and the impact of on-going treatment may further confound the interpretation of our observations in the liver disease cohort. Furthermore, this study has only 3,088 follow-up data since the imaging visit, which limits the identification of more severe cases and may limit the predictive power. Additional longitudinal measurements will need to be required to assess age-related changes in disease cohorts.
Publisher’s Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.