Radiomics is a promising new field in oncology that utilizes a high-throughput extraction method to transform medical images into high-dimensional, mineable data to obtain disease information to guide medical decisions in a non-invasive manner [30, 31]. Previous studies have demonstrated that MRI-based radiomics features were applied for the diagnosis, prediction of treatment response, prognosis estimation or other phenotypes [32–36]. Specifically, several studies have investigated MRI radiomics-based machine learning algorithms for differentiation of glioblastoma and primary central nervous system lymphoma[37], squamous cell carcinoma and lymphoma of the oropharynx [38], ocular adnexal lymphoma and idiopathic orbital inflammation [39] and so on. However, there is a lack of radiomics-based models for predicting the prognosis of extranodal natural killer/T-cell lymphoma (ENKTL), with only Wang et al. proving that radiomics features derived from PET images can predict patient outcomes [25]. Nonetheless, PET examination is expensive and its images have relatively low spatial resolution and high noise, which may influence lesion identification [40]. In contrast, MRI has emerged as o a vital tool for clinical diagnosis and therapeutic evaluation of nasal ENKTL due to its superior soft tissue contrast compared to CT scans. Therefore, in this study, we innovatively explore the MRI-based radiomics features performance for predicting OS of nasal ENKTL. To the best of our knowledge, this is the first study to incorporate prognostic features based on MRI radiomics to enhance the prognosis prediction for ENKTL, and significantly, our study was validated in an independent validation cohort (79 cases).
In this study, we have introduced multiparametric MRI-based radiomics signature defined Rad-score as a new approach to evaluate the OS before treatment in nasal ENKTL individually. Furthermore, we developed and validated a radiomics nomogram that integrated Rad-score and clinical variables, which exhibited superior performances compared with the radiomics signature, clinical model and provides an easily approachable, nonintrusive method for risk assessment for patients with ENKTL. To begin, we extracted comprehensive features from T2-w and CET1-w images, which were decreased to a set of only 20 potential descriptors by using the two sample t-test selection method and LASSO logistic regression model. These 20 radiomics features can reflect different aspects of structural and potential tumor biological information. 4 features were derived from the GCLM, which quantifies the probability that the adjacent gray scales with the highest frequency appear in pairs in ROI. A smaller probability indicates a more complex texture pattern. 2 features were derived from the GLSZM, reflecting the volume uniformity in ROI. The larger the value is, the better the regional uniformity is. 1 feature was derived from first-order metrics of the histogram and shape, which refer to gray level statistical information in ROI (including average value, entropy and homogeneity) and describe the overall distribution of image intensity information [18].1 feature was derived from the NGTDM, which characterizes texture consistency, aperiodic or spotty texture, has better performance than particle size, travel matrix and co-occurrence matrix in terms of nucleus, dermis, road quality (asphalt road) and PET image texture [41]. Finally, 11 features were derived from the wavelet features. The undecimated three-dimensional (3D) wavelet transform was used to decompose the original image, which can be regarded as preprocessing prior to feature extraction. By changing the ratio of high-frequency to low-frequency signal in images, wavelet transform increases the information of low-frequency signal [18].
The radiomics signature demonstrates satisfactory discrimination in both the training cohort (C-index = 0.733) and the validation cohort (C-index = 0.824). One possible reason for this is the remarkable spatio-temporal heterogeneous of tumors at various levels: genes, proteins, cells, microenvironments, tissues and organs. This heterogeneity provides promising prospects for tumor imaging evaluation [18, 42]. Radiomics features are typically extracted from intratumor regions, reflecting biological characteristics related to tumor heterogeneity [43]. Imaging techniques can quantify the spatial variation in the structure and function of individual tumors, measure blood flow, hypoxia, metabolism, cell death and other phenotypic characteristics, as well as map the spatial distribution of biochemical pathways and cell signal networks by quantifying basic biophysical parameters such as MRI signal relaxation rate [44, 45].
According to the results of the univariate and multivariate analyses, we included age, Ann Arbor stage, distant lymph node involvement in the clinical model. These factors have already been identified and evaluated retrospectively in several previous multicenter studies [15–17]. The C-indexes of the clinical model were 0.707 and 0.635 respectively. Some recent studies attempted to construct radiomics nomogram by combining the radiomics signature with TNM stage or other parameters, which showed a significant improvement in progression-free survival (PFS) or OS prediction [23, 46]. Accordingly, to enhance the power of the decision support model and provide a clinically applicable method, we further combined the Rad-score with significant clinical risk factors to develop an effective radiomics nomogram, which achieved better predictive values (C-index = 0.849 and 0.931) than that of the Rad-score and clinical model. These findings provided new insight into the future delivery of treatment for high-risk patients, who should receive more intense treatment. Future studies are needed to validate our findings.
With the development and application of new imaging devices, contrast agents, standardized scanning protocols and multimodal imaging techniques, quantitative imaging of tumors can be performed in imaging [30]. Radiomics, a field combining medical imaging, genetic analysis, and clinical data through artificial intelligence methods, provides a quantitative analysis of the uniformity of pixels in images and offers more information for clinical applications than traditional imaging methods [19]. accurately classifying tumor areas and conducting radiomics analysis pose as major challenges [47]. Further research should address this problem.
The limitations of the present study are as follows. First, as a single-center, retrospective study with a small sample size, it may have affected patient selection and the radiomics quantification results. Future prospective studies with a large sample are necessary to confirm the conclusion of the current study. Secondly, the tumor three-dimensional segmentation was accomplished manually, which is time consuming and complicated. An automatic segmentation method with favorable reliability and reproducibility is needed. Finally, the stability of radiomics nomogram needs further improvement by employing a larger training set with multicenter enrollment using different MRI protocols.