Whole-orbit radiomics: machine learning-based multi- and fused- region radiomics signatures for intravenous glucocorticoid response prediction in thyroid eye disease

Abstract

Background: Radiomics analysis of orbital magnetic resonance imaging (MRI) shows preliminary potential for intravenous glucocorticoid (IVGC) response prediction of thyroid eye disease (TED). The current region of interest segmentation contains only a single organ as extraocular muscles (EOMs). It would be of great value to consider all orbital soft tissues and construct a better prediction model.

Methods: In this retrospective study, we enrolled 127 patients with TED that received 4·5 g IVGC therapy and had complete follow-up examinations. Pre-treatment orbital T2-weighted imaging (T2WI) was acquired for all subjects. Using multi-organ segmentation (MOS) strategy, we contoured the EOMs, lacrimal gland (LG), orbital fat (OF), and optic nerve (ON), respectively. By fused-organ segmentation (FOS), we contoured the aforementioned structures as a cohesive unit. Whole-orbit radiomics (WOR) models consisting of a multi-regional radiomics (MRR) model and a fused-regional radiomics (FRR) model were further constructed using six machine learning (ML) algorithms.

Results: The support vector machine (SVM) classifier had the best performance on the MRR model (AUC = 0·961). The MRR model outperformed the single-regional radiomics (SRR) models (highest AUC = 0·766, XGBoost on EOMs, or LR on OF) and conventional semiquantitative imaging model (highest AUC = 0·760, NaiveBayes). The application of different ML algorithms for the comparison between the MRR model and the FRR model (highest AUC = 0·916, LR) led to different conclusions.

Conclusions: The WOR models achieved a satisfactory result in IVGC response prediction of TED. It would be beneficial to include more orbital structures and implement ML algorithms while constructing radiomics models. The selection of separate or overall segmentation of orbital soft tissues has not yet attained its final optimal result.

Keywords: Intravenous glucocorticoid; MRI; Multi-organ segmentation; Radiomics analysis; Response prediction; Thyroid eye disease.

Fig. 1

Radiomics workflow

Fig. 2

Illustration of the two segmentation strategies on the T2WI. The MOS strategy for the construction of multiple SRR and MRR models is presented in plot a, and the FOS strategy for the construction of FRR model is depicted in plot b

Fig. 3

Feature screening for MRR model. Plot shows the coefficients a and MSE b of LASSO regression model and features selected for model construction c

Fig. 4

Color maps demonstrating the diagnostic performance of different SRR models (EOM, OF, LG, or ON radiomic models) when utilizing different ML algorithms a–f. Colors depicted on each structure represent the AUC of corresponding SRR model based on a specific ML algorithm

Fig. 5

Predictive performance of the MRR (Multi-regional radiomics) and FRR (Fused-regional radiomics) models in the training and test cohorts. The ROC curves of MRR model in training cohort a and test cohort b; the ROC curves of FRR model in training cohort c and test cohort d

Fig. 6

The result and evaluation of prediction models in the test cohort. a The ROC curves of different radiomics models and SIR model based on the machine learning algorithms that achieved the highest AUC value. b DeLong’s test comparing the diagnostic performance (AUC) of different models. Calibration curves c and DCA d of different models. MRR Multi-regional radiomics, FRR Fused-regional radiomics

Fig. 7

Radar chart of the performance of MRR (Multi-regional radiomics) models and FRR (Fused-regional radiomics) models by using different machine learning algorithms (a–f)