Soft tissue sarcomas (STSs) are histologically heterogeneous and account for less than 1% of all malignant tumors (1). Radical resection is the standard treatment for patient with localized disease. Even after resection, the prognosis is poor, as reported recurrence rates range from 33% to 50% (2, 3) and the rate of distant metastasis is approximately 46% (4). Early identification of patients with a high risk of recurrence or metastasis after surgical resection to enable optimal use of standard or intensified neoadjuvant chemoradiotherapy might improve outcomes. This would require formulation of an accurate model for risk stratification in STS patients.

Despite their limitations, the TNM, Fédération Nationale des Centres de Lutte Contre Le Cancer (FNCLCC), and National Cancer Institute (NCI) staging systems are commonly used to guide STS prognostication and treatment (5). Several statistical models based on clinical and pathological data have been constructed and examined in previous studies to predict outcomes in STS patients (5–7). However, these models were based on low-dimensional clinical information and overlooked massive high-dimensional imaging characteristics. Therefore, their performance and generalized applicability are controversial (5, 8).

Radiomics extracts more detailed imaging features than traditional visual interpretation and can provide more data for clinical decision making (9). Radiomics-based models constructed using carefully screened features have the potential to predict STS outcomes (8, 10–13). In previous studies, radiomics data of tumor regions were analyzed as a whole and neglected intratumoral subregions (tumor habitats) with similar radiological phenotypes (14). Aggressive habitats might be crucial for tumor prognosis determination (14, 15). Several studies have demonstrated that tumor habitat analysis has high value in predicting tumor outcomes, both alone (16, 17) and in combination with radiomics analysis (18). Previous studies mainly concentrated on evaluation of the primary tumor and overlooked subtle changes in the peritumoral microenvironment (19, 20). However, the peritumoral microenvironment can explain aggressive biological behavior (21). Therefore, both tumor habitat and peritumoral environment should be evaluated to depict a tumor’s behavior and potential for invasion (14, 22).

This study aimed to establish and validate a radiomics nomogram that incorporated tumor habitat and peritumoral features to predict progression-free-survival (PFS) in patients with STS. We hypothesized that such a nomogram would show enhanced prognostic value.

Median PFS overall 148 patients was 12.5 months (range, 1–88). Mean PFS in patients who experienced STS progression and those who did not was 11 months and 13 months, respectively. Patients from the training and validation cohorts had similar baseline characteristics except for age, FNCLCC grade, NCI grade, American Joint Committee on Cancer (AJCC) stage, histopathological grade, depth, heterogeneous signal intensity on T2WI, radiotherapy, chemotherapy, and tumor location (Table 1).

In this study, we verified that a radiomics model combining intratumoral habitat features and peritumor features can predict tumor progression in patients with STS. PFS in our cohort ranged from less than 1 month (5 patients) to over 5 years (7 patients). Compared with analyzing radiomics features derived from intratumoral habitats or regions, the peritumor region, or tumoral expansion, the combined radiomics features signature yielded better predictive performance. Moreover, in the validation cohort, the nomogram showed a convincing level of performance (C-index, 0.777), less prediction error (IBS ≤0.135), good calibration, and convincing clinical usefulness.

Conventionally, radiomics has focused on analyzing the primary tumor as a whole. However, in consideration of the inherent internal heterogeneity and peritumoral aggressiveness of the tumor, it is conceivable that subregions within the tumor and regions surrounding it contain complementary useful information (28). In a previous study, the survival prediction performance of integrated features was better with integrated features than with intra- and peritumoral features alone (22). Another study suggested that a radiomics model based on tumor region habitats enabled accurate patient risk stratification (18). In our study, intra- and peritumoral features were integrally analyzed to construct a survival prediction model, which achieved a convincing performance and revealed that comprehensive analysis of multi-regional and multi-scale radiomics information can quantify tumor heterogeneity. The integrated model appears to have considerable potential in prognostication of STS patients.

Empirical evidence has shown that the tumor microenvironment might have an indispensable role in STS tumor relapse (29). Morphologic changes in the microenvironment that influence survival can be detected by peritumoral radiomics and peritumoral radiomics has potential for predicting progression (28). Dou et al. (30) analyzed radiomics features derived from a 3 to 9 mm region outside the tumor margin to predict distant metastasis of lung adenocarcinoma. Other studies have suggested that radiomics based on a region 15 mm outside of the tumor can stratify patients according to prognosis and predict the response to neoadjuvant therapy (21, 22). In a study conducted by Braman et al. (21), the peritumoral radiomics features included in the final prediction signatures were all derived from the region within 12 mm of the tumor margin; no feature from beyond 12 mm was included. In our study, the region 15 mm outside the STS lesion contained a large amount of bone, large vessels and air; therefore, we defined the peritumoral region boundary as 10 mm from the tumor margin. The peritumoral signature yielded better performance than other single-region signatures, demonstrating that the peritumoral region contains important information regarding STS progression.

Considering the significant variability observed across intratumoral regions, image-based partitioning has been used to identify relevant subregions important for prediction of tumor biological behavior (15, 16). High-throughput radiomics features can be screened for constructing quantitative models for oncology diagnostics. Verma et al. manually partitioned subregions within glioblastomas on the basis of multi-sequence MRI and analyzed the radiomics features derived from each subregion to predict tumor progression (18). However, manual partitioning is reliant on radiologist experience and can only be applied in partitioning of contiguous subregions, which may result in poor reproducibility and objectivity. The clustering of voxels in multi-sequence MRI is a data-based analysis method that enables segmentation of subregions of similar tissue at a voxel-wise level (31). Previous studies have demonstrated that voxel number or fraction of cluster-segmented habitats in functional or structural MRI is an efficient biomarker for tumor biological behavior prediction (32, 33). Nevertheless, these studies focused on analyzing a small number of habitat baseline characteristics and neglected high-dimensional radiomics features that depict tumor habitat heterogeneity. In our study, we considered the potential of integrating high-throughput radiomics feature analysis and voxel-based habitat segmentation to predict STS progression. We showed that the combination of radiomics features derived from intratumoral cluster-segmented habitats and peritumoral features yielded the best predictive performance, validating that intratumoral habitat radiomics features at the voxel level adds predictive value.

Neoadjuvant radiotherapy treatment plays a dominant role in improving prognosis in STS patients (34). Hence, it is vital to identify patients with high risk of progression and treat them accordingly. Our study demonstrated that the radiomics nomogram, which integrated voxel-based and multiregional radiomics features with clinical information, yielded favorable performance for PFS prediction and provided convincing risk stratification ability. Our nomogram generated two risk stratifications (low or high risk of progression) and should help fellow clinicians with management of individual STS patients. For patients with a low risk of progression, surgery without adjuvant therapy might be considered initially to avoid the side effects of chemoradiotherapy. For those with a high risk, postoperative systemic adjuvant chemoradiotherapy and targeted therapy should be considered. In current clinical practice, the most important prognostic indicator for STS is histopathological grade (35). According to our study, the stratification ability of the nomogram was further proved by the sub-cohort analysis in the low- and high-grade patients defined by histopathological grade system. Thus, use of our nomogram can provide incremental information to clinicians and STS patients and help guide treatment decisions.

Several study limitations should be mentioned. First, owing to its retrospective design, selection bias was probably present. Second, the radiomics generalizability and robustness across inconsistent MRI parameters and multiple institutions should be validated. Although we used standardization processes at the imaging and feature levels, more prospective data is needed to validate our findings. Finally, tumor boundaries were defined manually (first outlined by a junior radiologist and corrected by a senior one). Semi-automatic or automatic delineation should be used in future studies to minimize delineator bias.

In conclusion, we constructed a nomogram based on intratumoral habitat and peritumor radiomics that predicts tumor progression in STS patients and stratifies them according to risk of progression. Performance of the nomogram was superior to that of other habitat- and radiomics-based models.