Diagnostic accuracy of texture analysis applied to T- and T-Relaxation maps for liver fibrosis classification via machine-learning algorithms with liver histology as reference standard.

OBJECTIVES

To explore texture analysis' ability on T and T relaxation maps to classify liver fibrosis into no-to-mild liver fibrosis (nmF) versus severe fibrosis (sF) group using machine learning algorithms and histology as reference standard.

MATERIALS AND METHODS

In this single-center study, patients undergoing 3 T MRI who also had histology examination were retrospectively enrolled. SNAPSHOT-FLASH sequence for T1 mapping, radial turbo-spin-echo sequence for T2 mapping and spin-echo echo-planar-imaging magnetic resonance elastography (MRE) sequences were analyzed. Grey-level co-occurrence matrix texture analysis features were extracted from T (TA-T) and T2 (TA-T) maps from single-slice whole-liver region-of-interest. The extracted features were evaluated as predictors for nmF and sF group classification separately using support-vector-machine algorithm combined with principal component analysis in case of texture features. Recursive Feature Elimination with Cross-Validation (RFECV) was used to identify the most significant features and the importance of selected features was assessed with permutation importance algorithm. A combined model was identified and evaluated. Area under the receiver operating characteristic curve (AUC) was used for scoring and model comparison.

RESULTS

46 patients (mean age 52.8 ± 16.1 years, 23 males) were evaluated. TA-T performed comparably to MRE (0.748 vs 0.759, p = 0.905) and T performed slightly worse compared to MRE which was not statistically significant (0.692 vs 0.759, p = 0.396). MRE outperformed T (0.759 vs 0.552) and TA-T (0.759 vs 0.515). RFECV algorithm identified four features: MRE, T and 1st two TA-T principal components, constituting the first combined model. The permutation importance identified T as feature of very low importance, therefore a second combined model was constructed, omitting T from the first combined model. Even though both combined models performed higher than MRE (0.759 vs 0.797, p = 0.597 for MRE vs MRE + T + TA-T, and 0.759 vs 0.817, p = 0.373 for MRE vs MRE + TA-T), it was not statistically significant.

CONCLUSIONS

TA-T performed comparably to MRE in liver fibrosis classification to nmF and sF groups, and even though not statistically significant, combining those with MRE increased the performance, suggesting their complementary nature. Given the broad availability, robustness and short scanning times of T mapping, we would advocate for the inclusion of T mapping in every clinical and research liver examination.