3.1 Principle of radiomics approaches

The radiomics technique involves the extensive quantification of the tumor shape and texture based on any imaging modality beyond the conventional radiologist's depiction, using mathematical operators, such as histograms, gray level matrices, and wavelet or fractal analyses (Figure 1). Subsequently, hundreds of resulting numeric variables, termed radiomics features (RFs), are incorporated and mined using machine-learning algorithms to develop predictive models aimed at improved tailoring of patient management [20, 23]. Radiomics approaches rely on complex pipelines, including the selection and quality control of imaging, homogenization of an imaging dataset to ensure comparability, manual or semi-automatic segmentation of the volume of interest (for instance, the primitive tumor itself, its surrounding tissues, or metastases), RF extraction, exploratory data analysis, the use of technics to correct for imbalanced datasets, dimensionality reduction, training multiple machine-learning algorithms using resampling methods (such as nested cross-validation), and testing on novel original datasets to evaluate and select the best model objectively. This empirical process implies defining a metric for measuring the performance of the trained machine-learning models depending on the final objective of the radiomics study and the distribution of the outcomes in the study population. For instance, the area under the receiver operating characteristic curve (AUC) and accuracy, Harrell concordance index (c-index) and root mean squared error are commonly encountered for binary classification, prognostication and regression problems, respectively [24]. Calibration curves and decision curve analysis are additional means to evaluate the radiomics models [20, 25]. Due to this complexity, radiomics approaches are qualified as “handcraft.” Recently, researchers have developed alternative RFs, termed deep-learning RFs, using transfer learning, namely convolutional neural networks (CNNs) previously trained on the Imagenet dataset, such as Xception [26], VGG16 [27] and 19 [28], ResNet50 [29], or InceptionResNetV2 [30]. These deep-learning RFs supposedly provide higher reproducibility than handcraft RFs [31]. Moreover, radiomics composite scores are often integrated into nomograms along with the clinical, biological, and histopathological features, which generally improve their predictive performances [32-37].

3.2 Applications of radiomics approaches for patients with STS

Since its emergence in the early 2010s, clinicians have commonly used radiomics in the field of STS. STS exhibits various radiological presentations with different patterns of heterogeneity that are barely explainable using human vocabulary and usual radiological qualitative or semi-quantitative variables (also termed “semantic” features). Nonetheless, a key assumption behind radiomics is that these heterogeneity patterns reflect different subtypes of STS in terms of molecular features, treatment sensitivity, and prognosis [38].

Tables 1 and 2 summarize the major published sarcoma radiomics studies using CT and MRI. First, MRI-based radiomics from conventional sequences (T1-weighted imaging [WI], T2-WI, and CE T1-WI) and DWI have demonstrated high diagnostic performance in discriminating between benign soft-tissue tumors and STS (accuracy: 0.65–0.93 and AUC: 0.77–0.97 for approximately 609 unique patients) [32, 39–42]. Some authors have obtained good results for the peculiar but routine issues of discriminating lipoma from liposarcomas (AUC: 0.8–0.98 and accuracy: 0.87–0.95) [43-48] and leiomyoma from uterine sarcoma (AUC: 0.83–0.96 and accuracy: 0.74–0.88) [49-53]. Second, radiomics approaches have successfully predicted the histologic grade (according to the French “Federation Nationale des Centres de Lutte Contre le Cancer”) from pre-treatment MRIs (AUC: 0.76–0.92, accuracy: 0.82–0.98 for approximately 1,080 unique patients) [31, 33-35, 54, 55]. However, these authors used different definitions of high grade (i.e., grade II and III versus only grade III), and the grade was occasionally evaluated on biopsy samples, despite the possibility of grade underestimation and biased results [56, 57].

Clinicians have applied radiomics to predict patient prognosis, predominantly metastatic relapse-free survival (MFS), local relapse-free survival (LFS), overall survival (OS), and the risk of presenting lung metastases following initial treatments (either curative surgery alone or neoadjuvant radiotherapy/chemotherapy followed by curative surgery). The resulting radiomics prognostic models demonstrated good to strong performances (c-index: 0.77 for LFS, 0.84–0.93 for MFS, and 0.73–0.80 for OS) [36, 58–65]. Eventually, radiomics has demonstrated the ability to predict the histologic response following neoadjuvant treatments using baseline RFs and delta-radiomics, which correspond to a quantitative change in RFs between two radiological evaluations. Despite variations in the definitions for the histological response across studies (<5%, <10%, and <50% of the viable tumor cells on surgical samples), all studies demonstrated significant associations between the radiomics models and response (AUC: 0.63–0.91 and accuracy: 0.75–0.92) [66-69]. The major secondary findings were as follows: (i) radiomics approaches appeared to provide better predictions than conventional tools (RECIST v1.1 and semantic radiological analysis alone), (ii) delta-radiomics improved performances compared with single timepoint radiomics, and (iii) models directly designed to predict prognosis were more effective than those initially designed to predict the grade [33, 66, 68, 69].

These encouraging results support the need to better incorporate quantitative information from medical imaging because the radiophenotype of STS is related to each clinically relevant outcome. However, no study has investigated the association between any imaging biomarker (including radiomics) and the gene expression profiles in STS (including the CINSARC signature) or their potential synergy with the circulating tumor DNA [70]. Eventually, combining multimodal multi-parametric imaging and AI could enable performing a “virtual biopsy,” i.e., a comprehensive and non-invasive assessment of the tumor's pathological, molecular and prognostic features in cases where an actual biopsy is rarely feasible or insufficiently contributive.

3.3 Current limitations to the clinical applications of radiomics and the means to overcome them

Researchers have to address several obstacles common to imaging biomarkers [71], namely biological correlations, reproducibility and repeatability of the radiomics model, feasibility in other centers, external validation, comparisons against the reference methods with a distinct added value, and prospective validation in decision-making situations. Thus, no radiomics signature is currently in routine clinical use regardless of the cancer type.

However, implementing radiomics in the routine setting has several limitations, which the radiomics community is aware of and has endorsed several proposals to overcome. First, they have created an international collaboration, termed the Imaging Biomarker Standardization Initiative (IB, to standardize the methodology behind radiomics and the definitions of RFs, and provide several recommendations to perform distinct radiomics studies (https://ibsi.readthedocs.io/en/latest/) [72, 73]. Second, Lambin et al. [20] have proposed a score, termed the radiomics quality score (RQS), to quantify the quality of any radiomics study. The RQS consists of 16 items covering each study step, which can be used as support while designing a radiomics signature. Third, sharing databases with well-labeled imaging datasets and open-source codes to test the radiomics signatures would undeniably increase the confidence in radiomics and fill the translational gaps. The cancer imaging archive is the most famous of the initiatives to host freely available imaging datasets (https://www.cancerimagingarchive.net/) [74]. A small sarcoma dataset with PET-CT and conventional MRI is already present and included in radiomics studies; however, greater efforts are needed.

Moreover, the tedious and time-consuming manual segmentations of the tumor volume and the lack of practical and user-friendly applications for clinicians impose restrictions on radiomics. (Semi) automated segmentations using a deep-learning U-net CNN are expected to facilitate the segmentation step in the future [75]. Eventually, fully-automated pipelines that rely on deep-learning algorithms directly applied to the raw 3-D images should be able to capture the information better, owing to numerous and more personalized features for sarcoma.

4.1 Role of ICIs in patients with STS

Despite the recent success of ICIs, there are limited and controversial human data to support the efficacy of immunotherapy in STS across immunotherapy clinical trials conducted in patients with STS [83].

A non-comparative randomized phase II study conducted on 85 patients with refractory STS compared the treatment with nivolumab versus nivolumab + ipilimumab, revealing a 5% response rate (RR) with nivolumab alone, whereas the combination displayed promising antitumor activity with a 16% RR [84]. Another phase II study, SARC028, demonstrated an RR of 40% for undifferentiated pleomorphic sarcoma and 20% for de-differentiated liposarcoma under pembrolizumab treatment [85]. In contrast, a phase II study conducted by the French Sarcoma Group demonstrated only one response among 50 evaluated STS patients treated with pembrolizumab plus oral cyclophosphamide [86].

We recently reported a pooled analysis of clinical trials investigating the efficacy of ICIs on sarcomas [87]. This study included 384 patients from nine clinical trials; of these patients, 153 (39.8%) received a programmed cell death protein 1 (PD-1) or programmed cell death-ligand 1 (PD-L1) antagonist as a single agent. The overall objective response rate (ORR) was 15.1% (95% confidence interval [CI]: 8.6%–25.1%). However, it dropped to 9.8% upon excluding alveolar soft-part sarcoma, an extremely rare sarcoma subtype displaying exquisite sensitivity to PD-1/PD-L1 monoclonal antibodies, from the analysis [88, 89].

Overall, all studies investigating ICIs included patients without applying any biomarker-based selection strategy (Table 3). Researchers have observed anecdotal responses in some histological subtypes (e.g., undifferentiated pleomorphic sarcomas or de-differentiated liposarcomas), which might be more sensitive to treatment [84, 85]. PD-L1 expression, currently used as a biomarker for immunotherapy in several epithelial tumors (non-small cell lung cancer, head and neck, and triple-negative breast cancer), is not relevant for STS. An overall PD-L1 expression of ≥1% in tumor cells was observed in >15% of the patients, without distinct correlation with the clinical benefits [87]. Moreover, the tumor mutation burden was low, with a median burden of fewer than two mutations/Mb across all histological subtypes [90].

For an in-depth investigation of the therapeutic impact of ICIs in patients with sarcoma and its correlation with the sarcoma microenvironment, we conducted an extensive analysis of the immune landscape of sarcomas [91]. By analyzing transcriptomic data from more than 600 STSs, we identified a subgroup of sarcomas classified as “immune high,” characterized by an elevated expression of a B cell-related (BCR) gene signature, which was predictive of survival, independent of the level of CD8+ T cell infiltration [91]. In addition, an immunohistochemistry (IHC) analysis revealed that this class of sarcoma is characterized by the presence of intratumoral tertiary lymphoid structures (TLSs) [91]. TLSs are ectopically formed aggregates of B-cell follicles, follicular dendritic cells, and CD4+ and CD8+ T cells, which play a critical role in anticancer immunity in several tumor types [92]. Moreover, a retrospective analysis of the biopsies from 47 patients included in the SARC028 study indicated that the BCR gene signature was highly predictive of the response to the anti-PD-1 antibody pembrolizumab, thereby suggesting that the presence of TLSs may serve as a robust biomarker for immunotherapy customization in patients with sarcoma [91].

Subsequently, we amended the PEMBROSARC study (NCT02406781; sponsor: Institut Bergonié, Bordeaux, France) to include a recent cohort, which was based on the presence of TLSs to investigate the efficacy of pembrolizumab in patients with advanced sarcomas, characterized by the presence of this potential biomarker [93]. We included 35 patients with TLS-positive advanced STS in this study. The 6-month non-progression rate and ORRs were 40% (95% CI: 22.7%–59.4%) and 30% (95% CI: 14.7%–49.4%), respectively, compared with 4.9% (range: 0.6%–16.5%) and 2.4% (95% CI: 0.1%–12.9%), respectively, in previous cohorts of the PEMBROSARC study, which included all comers [86]. Interestingly, a retrospective analysis of the tumor samples from the patients included in the all-comer cohort displayed that all except one patient harbored a TLS-negative sarcoma, thus suggesting that patient selection based on the presence of TLSs would be an efficient approach toward the identification of an appropriate population for sarcoma-targeted immunotherapy [93].

TLSs are observed in approximately 20% of the patients with STS and can be identified in almost all histological subtypes, despite the high proportion of TLS-positive cases in some subtypes (e.g., de-differentiated liposarcoma versus leiomyosarcoma) [92]. The TLS status can be assessed by real-life samples, including small biopsies. Larger samples are more sensitive to detect TLS, which predominate at the tumor front; however, the biopsy samples can be efficiently used to screen TLSs [94]. We recommend a double CD20/CD23 staining to detect TLSs, which offers the advantage of simultaneously ensuring the presence of TLSs comprising B cells (CD20+) and the co-localization of follicular dendritic cells (CD23+).

The identification of a subgroup of sarcoma characterized by the presence of intratumoral TLSs represents an important step towards the development of immunotherapy for patients with sarcoma; nonetheless, researchers are yet to address at least two major questions before emphasizing immunotherapy as a novel and efficient standard of care in patients with STS as follows: how to improve the response rate of TLS-positive sarcomas and how to make immunotherapy an efficient approach in TLS-negative sarcomas?

4.2 Improving the response rates to ICIs in TLS-positive sarcomas

Simultaneously targeting different immune checkpoints is a potential approach to improve immunotherapy efficacy in patients with TLS+ sarcomas. TLS+ sarcomas are characterized by the increased expression of immune checkpoints besides PD-1, such as cytotoxic T-lymphocyte antigen 4 (CTLA-4), TIGIT or lymphocyte activation gene 3 (LAG3) [91]. LAG3 (CD223) is the third immune checkpoint recently targeted in clinical practice and has garnered considerable interest and scrutiny [95]. LAG3 upregulation is required to control the overt activation and prevent the onset of autoimmunity. However, persistent antigen exposure in the tumor microenvironment (TME) results in sustained LAG3 expression, thus contributing to a state of exhaustion that manifests as impaired proliferation and cytokine production. The exact signaling mechanisms downstream of LAG3 and their interplay with other immune responses remain largely unknown. However, the synergy between LAG3 and PD-1 in multiple settings, coupled with the contrasting intracellular cytoplasmic domain of LAG3, compared with the other immune responses, highlights the potential uniqueness of LAG3. Our research findings [91] and others [96] have demonstrated the strong upregulation of LAG3 in inflamed sarcomas. LAG3+ tumor-infiltrating lymphocytes (TILs) were observed in approximately two-thirds of complex genomic sarcomas, and their expression was strongly correlated with disease outcomes [96]. Based on this rationale, a clinical trial is currently investigating the activity and safety of nivolumab combined with relatlimab, a monoclonal antibody that targets LAG3, versus nivolumab alone, in patients with STS (https://www.clinicaltrials.gov/ct2/show/NCT04095208).

4.3 Successful immunotherapy in 80% of the patients with TLS-negative sarcoma

Investigating combination strategies with ICIs to transform “cold tumors” into “hot tumors” and improve their sensitivity to ICIs is key to addressing this issue. Several clinical trials are currently investigating the combinations of ICIs with different classes of agents, namely chemotherapy, radiotherapy, toll-like receptor agonists, anti-angiogenics, oncolytic viruses, and epigenetics, that enable the conversion of “cold tumors” to “inflamed tumors”. To better understand the impact of such combinations on the TME, investigators and patients should realize the importance of sequential blood and tissue sampling or a randomized design to avoid any bias in data interpretation. A recent study investigated the combination of doxorubicin and pembrolizumab in patients with advanced STS [97]. The RR (13%) and PFS observed in this phase II study were similar to those observed when doxorubicin was used as a single agent [97]. Unfortunately, the lack of sequential tumor biopsies and randomization precludes any conclusion about the impact of the TME combination and correlations with the clinical benefits, thereby limiting the value of the study in terms of an improvement of knowledge for the scientific community.

4.4 Immunotherapy for sarcomas beyond ICIs

Besides ICIs, researchers have investigated other immunotherapy strategies in patients with STS. For this review, we focused on T-cell receptor (TCR) gene-modified cells, which are more advanced in terms of clinical development. This therapeutic approach is based on the ex vivo expansion of autologous T cells following their genetic engineering to express a novel TCR that recognizes specific tumor antigens. Ongoing clinical trials investigating such TCR-based therapies in patients with STS use TCRs directed at validated cancer germline antigens, i.e., MAGE-A4 and NY-ESO-1, expressed in particular STS subtypes, including myxoid/round cell liposarcoma (MRCLS) and synovial sarcomas.

Ramachandran et al. [98] reported that the adoptive transfer of NY-ESO-1c259 T cells in 42 patients with synovial sarcoma (NCT01343043) was associated with an ORR of 35.7% (15 patients; 1 CR and 14 PR cases) by RECIST. D'Angelo et al. [99] demonstrated that the persistence and functionality of these adoptively transferred T cells were associated with sustained responses in a significant proportion of patients. Moreover, treatment with an autologous T-cell therapy targeting NY-ESO-1 displayed antitumor activity up to 40% ORR and long median PFS, with an acceptable safety profile in a study that enrolled 23 patients with advanced and metastatic MRCLS [100]. Similarly, preliminary data suggested that afamitresgene autoleucel, a genetically modified autologous melanoma-associated antigen 4 (MAGE A4) specific T cell therapy, could induce a prolonged response in heavily pre-treated patients with advanced MAGE-A4-expressing SS and MRCLS [101].

Despite promising preliminary results, the development of TCR-based adoptive cell therapies for treating patients with STS is associated with several challenges, including those associated with manufacturing and processing of the final TCR product, patient selection, optimizing lymphodepletion conditioning, and managing adverse events. Ongoing and future studies will determine the role of TCR therapy in patients with STS.