1 Introduction

Soft tissue sarcomas (STSs) are rare malignancies that arise from mesenchymal cells [1]. They account for about 1% of all adult cancers and have a wide range of histological subtypes. Pre-therapeutic tumour grading and assessment are crucial in making treatment decisions, as they provide prognostic information and guide the choice of the proper approach (surgery, chemotherapy, and radiotherapy) [2].

In diagnosing suspected STSs, essential imaging techniques such as ultrasound and Magnetic Resonance Imaging (MRI) are fundamental, with MRI being vital for comprehensive evaluation. Thoracic Computed Tomography (CT), or PET-CT scans are instrumental in identifying metastatic sites, while precise imaging is key for biopsy guidance to accurately localise the lesion [3].

One of the most recenty introduced techniques in radiological science is Radiomics, a high-throughput approach for the analysis of imaging data; this method offers quantitative information that augments expert assessments [4, 5]. It involves the extraction of a large number of features from medical images that reflect the tumour characteristics such as shape, size, intensity, texture, and heterogeneity [6]. Radiomics has been applied to various types of cancers, including STSs, with promising results in terms of diagnosis, prognosis, and prediction [7, 8]. Figure 1 depicts an illustrative radiomics workflow applied to a left thigh myxoid fibrosarcoma, sourced from an open-source anonymized database (https://doi.org/10.7937/K9/TCIA.2015.7GO2GSKS) [9, 10].

Fig. 1
figure 1

An illustrative radiomics workflow applied to a left thigh myxoid fibrosarcoma

Full size image

This review aims to provide an overview of recent publications in the field of STS radiomics. It categorizes these studies into various domains of interest, highlighting the diverse applications and limitations of radiomic analysis in STS.

2 Materials and methods

We conducted a systematic review according to the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) guidelines [11]. A comprehensive search was conducted in PubMed/MEDLINE and Scopus using the search terms: ‘radiomics [All Fields] AND ("soft tissue sarcoma" [All Fields] OR "soft tissue sarcomas" [All Fields])’. After removing duplicates, original published articles were included in the analysis.

All single, comparative studies, and primary studies that met the following PICO criteria were selected: P (patients): Patients with STSs; I (interventions): Radiomics; C (comparison): Conventional diagnostic imaging (including CT, MRI,PET/CT); O (outcome): The impact of radiomics on STSs on diagnosis, prognosis, risk stratification, genetic/histological prediction and technical feasibility aspects.

The following exclusion parameters were applied: (1) not original articles (e.g. letters, reviews, editorials, book chapters, congress communications); (2) papers not concerned radiomics topic; (3) researches not referred to humans (e.g. STSs in mice); (4) only articles in English, French, Spanish, Italian or German were included.

Two radiologists (RC, RDE) initially analysed all articles. An independent validation was performed by one other radiologist (MAB), by one resident in radiology (AM), and by one physicist (NC). The complete procedure, along with the results and any discussion regarding probable inconsistencies, was verified by one other independent radiologist (PS), expert in the field of musculo-skeletal oncological radiology.

The quality assessment of the eligible articles was evaluated using the Radiomics Quality Score (RQS) [12] by 2 evaluators (RC) and (SI). Each of the 16 essential criteria specified by the RQS was individually rated, yielding a composite score ranging from -8 to 36 points. These scores were subsequently transformed into RQS percentages, with a score of -8 to 0 points corresponding to 0% and a score of 36 points corresponding to 100% [12].

3 Results

A preliminary search conducted on PubMed/MEDLINE and Scopus provided 74 and 93 studies respectively. After removing duplicates and applying the aforementioned PICO criteria, a total of 94 papers were retained through evaluation of their titles and abstracts. Finally, after an extensive selection process (Fig. 2), 49 papers were eligible for analysis. In particular, 36 papers were excluded from consideration due to their non-original article status; 6 papers were omitted as they did not pertain to the subject of radiomics; 4 papers were excluded as they did not involve human subjects; 2 papers were disregarded based on their language of publication.

Fig. 2
figure 2

Selection process of literature

Full size image

Among the retained papers, 46 studies were retrospective, and 3 studies were prospective, with a publication range from July 2015 to June 2023. The participant centres (Fig. 3) included China (n = 18; 36.7%), Italy (n = 6; 12.2%), France (n = 6; 12.2%), Germany (n = 5; 10.2%), USA (n = 4; 8.2%), Canada (n = 2; 4.1%), Netherlands (n = 2; 4.1%), Republic of Korea (n = 2; 4.1%), Spain (n = 1; 2%), United Kingdom (n = 1; 2%), Belgium (n = 1; 2%), Qatar (n = 1; 2%). MRI (n = 41; 77.4%), PET-CT (n = 5; 9.4%), CT (n = 5; 9.4%) or PET (n = 2; 3.8%) images were used for radiomics analysis (Fig. 4)—some of above-mentioned articles used more than one imaging techniques. The median number of patients involved in the analysis was 63 (range 11–540). The median number of radiomics features was 160 (range 30–2758).

Fig. 3
figure 3

Participant centres of included articles

Full size image
Fig. 4
figure 4

Imaging techniques of included articles

Full size image

Regarding the domain of interest (Fig. 5): 24 (36.4%) articles were focused on risk stratification, 19 (28.8%) articles on radiogenomics, 9 (13.6%) articles on technical feasibility/reproductive aspects, 9 (13.6%) articles on treatment response and 5 (7.6%) articles on diagnosis—some of aforementioned articles treated more than one domain of interest.

Fig. 5
figure 5

Domains of interest covered by the selected articles

Full size image

In total, 24 studies analysed the use of radiomics for risk stratification. 7 articles assessed the role of radiomics models for predicting lung metastasis [9, 13,14,15,16,17,18]; 3 articles analysed radiomics models for prediction of distant metastasis or metastatic relapse-free survival [19,20,21]. The ability to predict overall survival or free survival was evaluated in 6 studies [22,23,24,25,26,27]; in particular, according to Spraker et al. [26], texture features related to histogram_skewness, histogram_kurtosis, GLZSM_Small zone/low grey emphasis and GLZSM_Zone, obtained from T1-weighted contrast-enhanced images, were selected in the models for predicting overall survival. Fadli et al. [28] found that increase in heterogeneity (visually evaluated) and logarithmic change in radiomics features clusters, in contrast enhancement MRI T1-weighted images, were independent predictors for metastatic relapse-free and local relapse-free survival. One study developed a radiomics model for predicting disease-free survival in patients with STSs of the extremities and trunk who have undergone neoadjuvant radiotherapy [29]. Lee et al. [30] investigated the effectiveness of a radiomics model using T2-weighted Dixon sequence in differentiating the degree of STSs margin infiltration. Zhao et al. [31] evaluated the ability of various PET/MRI fusion methods to extract features for the prediction of recurrence/metastasis in patients with STSs. Tagliafico et al. [32] analysed MRI radiomics features in surveillance of local recurrence in patients with limb STSs. Liu et al. [33] evaluated the accuracy of two deep learning-radiomic nomogram models, in conjunction with clinical parameters, for predicting local recurrence in patients with STSs who underwent surgical resection. Lastly, one recent study presented a methodology employing MRI radiomic features for the prediction of metastasis and recurrence risk in patients with extremity STSs using formal logic models [34].

19 radiogenomics articles aimed to establish a connection among image phenotype, gene expression, mutations, molecular, or pathological findings. In particular, 11 studies have been conducted to develop and evaluate MRI-based radiomics for the differentiation of grade in STS tumours [20, 25, 32, 35,36,37,38,39,40,41,42,43]; furthermore, a study among these aforementioned ones aimed to predict the grade and the Ki-67 expression level by utilising intravoxel incoherent motion MRI and diffusion kurtosis imaging parameter maps [43]. Corino et al. [42] discovered that the GLCM features related to dissimilarity and entropy showed higher values in the high-grade. Peeken et al. [44] further explored the potential of quantitative imaging features in CT radiotherapy planning for predicting the grading. In children, using T2-weighted MRI images, Giraudo et al. [45] discovered that lmc1 feature was associated with high-grade tumours and variance feature was associated with rhabdomyosarcomas histotype. One study [27] assessed the predictive value of FDG PET/CT conventional metrics and textural features in determining histopathological data; in particular, the FNCLCC score (representing a histologic surrogate for tumour aggressiveness) demonstrated a correlation with GCLM_dissimilarity, GLCM_contrast and an inverse correlation with GLCM_homogeneity. Crombé et al. [46] investigated the association between distinct patterns of natural evolution of STSs, based on MRI radiomics features, and differential gene expression. Timbergen et al. [47] assessed whether radiomics can differentiate between desmoid-type fibromatosis and STSs and can predict CTNNB1 mutation types in desmoid-type fibromatosis patients. One study [28] examined the associations between temporal alterations observed in MRI, based on qualitative/semi-quantitative features and radiomics features, and the survival outcomes and histopathological characteristics. Foreman et al. [48] developed radiogenomic models with the purpose of predicting the MDM2 gene amplification status and distinguishing between atypical lipomatous tumours and lipomas based on preoperative MRI scans. Nine studies assessed the role of radiomics features for evaluating the toxicity, management and treatment responses of patients with STSs treated with radiotherapy and/or chemotherapy [23, 44, 49,50,51,52,53,54,55].

There were 9 studies conducted to assess the technical feasibility and reproducibility of radiomics analysis [9, 16, 17, 31, 35, 55,56,57,58]. The influence of ComBatHarmonization on MRI-based radiomics models to differentiate between low-grade and high-grade STS tumours was analysed by Peeken et al. [35]. Thrussell et al. [58] evaluated the repeatability of radiomic features from Diffusion Weighted Imaging (DWI) and Apparent Diffusion Coefficient (ADC) maps in retroperitoneal STSs and compared their repeatability before and after radiotherapy. Vallières et al. [9, 16] assessed the influence of different extraction parameters and acquisition protocols on FDG-PET/MRI models to predict the risk of lung metastasis. Zhao et al. [31] evaluated the ability of various PET/MRI fusion methods to extract features for the prediction of recurrence/metastasis. Sheen et al. [17] analysed the efficacy of four segmentation methods in defining radiomics signatures and prediction models for lung metastases using PET-CT in STSs.

In 5 studies, there were observed associations between imaging features and the ability to diagnose and differentiate STSs from normal tissue or non-malignant lesions. Yue et al. [59] developed a clinical-MRI radiomics nomogram aimed at distinguishing between benign and malignant soft-tissue tumours. Tagliafico et al. [32] investigated the potential use of radiomics in MRI surveillance in patients with limb STSs to differentiate between normal tissue and local recurrence. In another study, Timbergen et al. [47] evaluated MRI radiomics models for distinguishing between desmoid-type fibromatosis and STSs. As previously mentioned, Foreman et al. [48] attempted to develop radiogenomic models with the goal of distinguishing between atypical lipomatous tumours and lipomas by analysing the MDM2 gene amplification status using preoperative MRI scans. Aouadi et al. [60] examined several datasets, in particular the LIPO dataset [61] for distinguishing between well-differentiated liposarcoma and lipoma, and the Desmoid dataset [61] for differentiating desmoid-type fibromatosis from extremity STSs.

Table 1 presents a comprehensive overview of all the aforementioned analysed studies, highlighting the primary objectives, outcomes, and various domains of interest.

Table 1 Detailed summary of the analyzed studies, outlining their objectives, conclusions, and diverse domains of interest
Full size table

The 49 articles showed a median RQS of 12 and the values varied from 2 to 23 (Table 2). Additionally, the median RQS expressed as a percentage was 33%, with the minimum recorded value being 6%, while the highest was 64%.

Table 2 RQS for the 49 selected articles
Full size table

4 Discussion

Imaging has a prominent role in diagnosis and in treatment decision making in patients with STS. When an STS is clinically suspected, the role of diagnostic imaging and multidisciplinary discussions are essential. Ultrasound is often the first-approach modality, but MRI is mandatory to characterize the lesion, assess its anatomical limits and therefore guide treatment decision [3]. Also, some STS features seen on MRI may predict the grade of malignancy of the lesions [62], recurrences [63, 64], post-treatment oedema and seroma [65]. Thoracic CT or PET-CT are used to assess secondary lesions. Imaging is also important to guide a biopsy, which may be necessary to confirm the exact histological nature of the lesion. Recently, the development of new imaging techniques, such as whole-body MRI, hybrid PET/MRI, diffusion weighted MRI, dynamic contrast MRI and advances in artificial intelligence have greatly enhanced the radiologist role in tumour grading and staging assessment.

However, the only use of imaging has still a very limited accuracy and therefore a histological confirmation is often needed. In order to improve imaging accuracy, several new imaging techniques were proposed in latest years. One of these techniques, radiomics, has shown promising results. It may be associated with the notion of radiogenomics, positing that imaging characteristics correlate with genetic signatures [6]. This method surpasses conventional imaging by being non-invasive, objective, and cost-effective. Importantly, radiomics could provide complementary information to histopathology and molecular biomarkers, enhancing tumour evaluation and aiding in personalised treatment strategies.

In recent advancements within oncologic imaging, radiomics has extended its utility to the field of STSs. Our review elucidates the versatile applications of radiomic analysis in STSs, showcasing interesting results across various critical domains. Notably, radiomics has demonstrated significant potential in aiding with diagnosis, facilitating risk stratification, predicting grading or genomic status, and evaluating treatment responses in STSs.

Our review underscores the potential of radiomics to complement and augment conventional imaging approaches. We have observed that radiomics, with its high-throughput data analysis capabilities, can provide deeper insights into the imaging characteristics of STSs. This integration of radiomic techniques with traditional imaging approaches marks a significant advancement in the field.

Albeit radiomics discoveries have shown promises, more work is needed to ensure these results can be replicated with a satisfactory degree of proof by enrolling larger patient datasets. The heterogeneity of features, the variability of acquisition protocols and multiple image modalities further complicate the generalizability of radiomics observations. Moreover, independent validation of proposed predictive models often is lacking, making extensive radiomics studies in STSs challenging. Additionally, the complexities inherent in radiomics may challenge its integration into routine clinical practice, suggesting a need for tailored training and resources to facilitate its use among clinicians and radiologists. An improved implementation of the latest guidelines [12, 56, 66,67,68] and increased knowledge of radiomics will be necessary to enhance the quality of sarcoma radiomics studies and to enable their implementation in the clinical setting.

5 Conclusion

Our review highlights radiomics as a potential adjunct to conventional STS methodologies, enhancing diagnostic and treatment strategies. Yet, the limitations of the current approach necessitate further studies. Additionally, the complexities inherent in radiomics may challenge its integration into routine clinical practice. Future investigations should focus on validating radiomics' clinical application and establishing its role in STS management practices.