SAT0569 “IMAGES ARE MORE THAN PICTURES, THEY ARE DATA” [1] – EXPLORATION OF RADIOMICS ANALYSIS FOR SYSTEMIC SCLEROSIS-ASSOCIATED INTERSTITIAL LUNG DISEASE

Abstract

Background:Interstitial lung disease (ILD) affects 60% of patients with systemic sclerosis (SSc) and is the primary cause of death. Medical imaging is an integral part of the routine work-up for diagnosis and monitoring of SSc-ILD and includes high-resolution computed tomography (HRCT). Radiomics is a novel research area that describes the in-depth analysis of tissue phenotypes in medical images with computational retrieval of quantitative, mineable metadata appropriate for statistical analyses.Objectives:To explore the performance of HRCT-derived radiomic features for the assessment of SSc-associated ILD (i.e. diagnosis, staging, and lung function).Methods:Radiomics analysis was performed on HRCT scans from 98 SSc patients, including n=33 SSc patients without ILD, n=33 with limited and n=32 with extensive ILD as defined by 0%, <20% and ≥20% visual extent of fibrosis on HRCT, respectively. Following semi-automated segmentation of lung tissue on 3D reconstructed HRCT scans, 1386 radiomic features, including 17 intensity, 137 texture, and 1232 wavelet features were extracted using the in-house developed software Z-Rad (Python 2.7). In order to identify robust features, we conducted intra- and inter-reader correlation analysis (ICC) in a subgroup of patients. Only features with good reproducibility (ICC ≥ 0.75) entered subsequent analyses. We applied the Wilcoxon test, followed by Receiver Operating Characteristic ROC) curve analyses, to identify features significantly different between a) ILD and non-ILD and b) limited vs. extensive ILD patients. Spearman rank correlation was performed to reveal significant associations of radiomic features from a) and b) with lung function as measured by percentage of predicted forced vital capacity (FVC% predicted).Results:In total, 1355/1386 radiomic features passed the test of robustness and were eligible for further, exploratory analyses. Radiomic features with good performance (area under the ROC curve (AUC) ≥ 0.7 and p-value ≤ 0.05) were considered as potential candidate discriminators. Under this criterion, we identified 288/1355 (21.3%) radiomic features that were significantly different between ILD and non-ILD patients and 409/1355 (30.2%) features that significantly discriminated between limited and extensive ILD (Fig. 1). For diagnosis, the texture featuredependence count entropywas the top parameter to distinguish ILD patients from healthy controls (AUC = 0.89, p = 1.83x10-10), whereas for staging the wavelet featureHHH long run high grey level emphasisproved to be best suited to separate limited from extensive ILD (AUC = 0.88, p = 7.76x10-9).Fig 1.Correlation analysis of the most significant (best performing) discriminative radiomic features with lung function revealed a significant negative correlation ofdependence count entropy(rho = -0.51, p = 9.89x10-8) andHHH long run high grey level emphasis(rho = -0.51, p = 1.73x10-5) with FVC% predicted.Conclusion:Our study adds novelty to the field of SSc-ILD showing that radiomic features have great potential as quantitative imaging biomarkers for diagnosis and staging of SSc-ILD and that they may reflect lung function. As the next step, we are planning to build predictive models, using machine learning, for diagnosis, staging, and lung function and validate them in external patient cohorts. If validated such models will pave the way for computer-aided management in SSc-ILD and thus improve patients’ outcome.