Supervoxel-based image-to-biomarker conversions - An initial study on morphological age prediction from whole-body MRI and its clinical relevance in the UK Biobank

Authors

Abstract

Biological aging remains a central focus of research, from the scale of sub-cellular processes to whole-organism tissue morphology and function. In this work, we developed a novel and quantitatively interpretable method for the prediction of variables, such as age, from tomographic medical images. The method uses supervoxels (whose granularity is selected by the user), standardized through inter-subject image registration, and tissue-specific feature extractions from each supervoxel to convert all image data collected into a set of well-defined imaging biomarkers. We applied the method to age prediction, using linear modelling and whole-body water-fat MRI data of 38,235 subjects in the multicentre UK Biobank study, resulting in predictions representing morphological age (MA). We observed state-of-the-art whole-body age prediction performance on a held-out test set with a mean absolute error of 1.951/2.057 years, and R2 of 0.884/0.892 for females/males, respectively. The method was observed to outperform both previously reported CNN-based results from the UK Biobank and predictions from explicit biomarkers from a multi-organ/tissue segmentation approach, in a direct comparison. The interpretability of the method enabled a detailed analysis of the body-wide associations with age. Volumes of the aorta, regional muscle, bone marrow, and adipose tissue depots, and lean tissue fat content were of particular importance. The predicted MAs were of clinical relevance as they were significantly related to both type 2 diabetes and all-cause mortality. A key finding was an accuracy/utility trade-off where the more parsimonious models showed lower chronological age (CA) predictive performance but higher clinical relevance and interpretability. The proposed method facilitates automated image-to-biomarker conversions and predictions based on subsets of anatomies, tissues, and image features, for potential application in numerous future medical studies. This study was funded by the Swedish Heart-Lung Foundation, the Swedish Research Council, EXODIAB, and Uppsala Diabetes Center. Research in contextO_ST_ABSEvidence before this studyC_ST_ABSWe searched PubMed and Google Scholar with search terms: "uk biobank age prediction" and "whole-body mr age prediction". Two studies involving age prediction from whole-body MR images were found, primarily using black-box AI methods, and both achieved good performance on chronological age (CA) prediction. For other imaging sequences, such as brain images, high-performance AI models have been developed. For other imaging sequences and other cohorts, age prediction using features derived from image segmentation has also been explored, commonly exhibiting lower CA prediction performance but higher interpretability. Added value of this studyThe study introduces a novel general-purpose image-to-biomarker method: tissue-specific standardized supervoxel-based prediction (TS-SSP), relying on the definition of well-defined imaging-biomarkers through inter-subject image registration, tissue-specific supervoxel-based feature extraction, and using linear models to derive an interpretable morphological age (MA) prediction from a large dataset of whole-body MR images. The study evaluates the proposed method in comparison to segmentation-based methods, and indirectly with deep learning-based methods, and explores the interplay between model complexity, CA prediction performance, and clinical relevance. Interpretability analysis reveals spatially-resolved tissue- and organ-level associations with age and both fat content and tissue volume. Implications of all the available evidenceThe study shows that it is feasible to achieve age prediction with the interpretability of segmentation-based approaches with higher spatial resolution and the high CA prediction performance of deep learning-based approaches (or even higher performance) simultaneously through the use of tissue-specific standardized supervoxel-based prediction relying on image registration and linear models. The study uncovers a trade-off between the CA prediction performance and relevance to diabetes and mortality (and interpretability), underscoring the need for rigorous evaluation of age prediction methods against clinically relevant outcomes. Previously known associations with age were found, such as aorta volume and muscle volume, in addition to detailed tissue-level associations.

Tags

Topics