MRI-Based Radiomic Models Predict Tumor Grading in Soft-Tissue Sarcoma Patients

Abstract

The treatment decision for soft tissue sarcomas (STS) patients vastly depends on the tumor grading determined by invasive biopsies. In this work, we used quantitative imaging features ("radiomics") of pre-therapeutic magnetic resonance imaging (MRI) scans to develop a non-invasive grading model for the differentiation of low-grade from high-grade STS. MRI scans (T1-weight with fat saturation and contrast enhancement (T1FSGd) and T2-weight with fat saturation (T2FS) sequences), clinical information and French Fédération Nationale des Centres de Lutte Contre le Cancer tumor grading were determined from two retrospective cohorts obtained from two independent institutions (122 and 103 patients). After manual segmentation, preprocessing steps including isotropic resampling, intensity discretization, and N4ITK bias field correction were applied. 1394 radiomic features including volume, intensity, texture, wavelet and local binary pattern were extracted using a software program package in python. Features unstable to segmentation variances by three independent experts with an intra-class coefficient below 0.8 were excluded. The ComBatHarmonization was used to compensate for MRI scanner types. Feature reduction and model generation was performed using the least absolute shrinkage and selection operator (LASSO) algorithm. Performance stability was determined by 1000 fold bootstrapping. All statistical methods were performed in R. Low-grade and high-grade STS were found in 48 (39%) and 74 (61%) patients in the training set, and in 20 (19%) and 83 (81%) patients in the validation set, respectively. T1FSGd and T2FS showed similar performances with area under the curve (AUC) values of 0.94 and 0.90 on the training set (see Table 1 for 95% confidence intervals). On the external validation set, the T2FS-based model showed a better reproducibility with an AUC of 0.76 compared to the T1FSGd model (AUC: 0.68). A clinical model (TNM-stages and age) and a tumor volume model did not predict better than random. Neither a combined radiomic model nor a combination with clinical features improved the predictive capacity. In contrast to T2FS (p=0.811), the T1FSGd radiomic models allowed for significant patient stratification for overall survival in the validation set (p=0.048). Decision curve analysis revealed a similar net benefit as tumor grading for the T1FSGd radiomic model. We first describe MRI-based radiomic models for non-invasive tumor grading of STS. Future studies should aim to further improve the reproducibility of MRI-based studies.Abstract 2323; Table 1Performances for differentiation of low-grade and high-grade STS.Training AUC (95%CI)Independent Validation AUC (95%CI)Radiomics-T1FSGd0.94 (0.892-0.976)0.68 (0.555-0.798)Radiomics-T2FS0.90 (0.829-0.949)0.76 (0.636-0.863)Clinical0.56 (0.455-0.663)0.55 (0.414-0.686)Volume0.56 (0.446-0.668)0.62 (0.464-0.759) Open table in a new tab