Evaluation of MRI-derived radiomics features of hepatic fat as biomarkers of type 2 diabetes mellitus and the metabolic syndrome

Abstract

Objective: To assess MRI derived radiomics features of liver fat collected from a cohort of individuals without prior cardiovascular events as imaging biomarkers of type 2 diabetes mellitus (T2DM) and metabolic syndrome (MetS).

Material and methods: 400 participants of the KORA MRI substudy underwent comprehensive whole body MRI imaging protocols including T1-weighted dual-echo Dixon (T1-DED), T1-weighted multi-echo Dixon (T1-MED) and magnetic resonance spectroscopy (MRS) sequences. A total of 684 radiomics features were extracted on T1-DED relative fat water content (rfwc) maps of 310 artefact free manually contoured liver volumes of interest (VOI). The corresponding individuals (n = 310, T2DM 12.6 %, MetS 34.5 %) were assigned to stratified training (n = 232, 75 %) and validation (n = 78, 25 %) sets. To assess feature stability, test-retest and inter-rater variance was approximated by generating noise augmented rfwc maps and computationally deformed VOIs, respectively. Feature stability was assessed in terms of the intraclass correlation coefficient (ICC) for test-retest reliability (ICC(1,1)) and inter-rater agreement (ICC(3,k)) on training set data. Stable features (ICC ≥ 0.85) were assessed as imaging biomarkers of T2DM and MetS in random forest (RF) models. For benchmarking, RF models were trained on the participants’ hepatic proton density fat fraction (PDFF) quantified previously on the T1-MED and MRS images as well as the body mass index (BMI). All RF models were evaluated on the validation set using the area under the curve of the receiver operating characteristic (AUROC) and the balanced accuracy (AccuracyB) as performance metrics.

Results: The epidemiological characteristics in training and validation sets were not statistically significantly different (p < 0.001). Furthermore, training and validation sets showed strong associations for both test-retest reliability (β 1.027 [1.015 – 1.040]) and inter-rater agreement (β 1.072 [1.040 – 1.103]). A total of 171 features (25.0 %) met the stability threshold (ICC ≥ 0.85). In subsequent RF modelling radiomics features predicted T2DM with AUROC 0.835 and AccuracyB 0.822 and MetS with AUROC 0.838 and AccuracyB 0.787, thereby outperforming all benchmark RF models in both metric categories.

Conclusion: In this single-center study radiomics features of MRI-derived hepatic fat were superior biomarkers of T2DM and MetS than hepatic PDFF and the BMI. In the future, hepatic radiomics features deserve further exploration and development as potential biomarkers in metabolic disease.