Quantitative MRI is a multi-step process that requires image acquisition, reconstruction, segmentation and often mathematical modelling of parameters prior to feature extraction ( Figure 1 ) (2). A semi-quantitative MRI parameter already in use in clinical practice is contrast-enhanced (CE-)MRI, where the difference in signal after intravenous injection of a gadolinium-based contrast agent is used as a surrogate for tumour vascular perfusion (22). Examples of fully quantitative parameters are Dixon MRI which enables fat content in tumours to be visualised and quantified and apparent diffusion coefficient (ADC) measurements from diffusion-weighted (DW)-MRI which maps tissue cellularity (22, 25–27). ADC is rapidly growing in use in oncology because of the widespread availability of DW-MRI on clinical scanners, ease of acquisition (not requiring additional equipment or injection of contrast agents), excellent repeatability and large number of biological and clinical validation studies (22, 27, 28). These factors make it an attractive IB, and as such, ADC has emerged as a useful quantitative IB in STS which has led to its incorporation in the European Organisation of Research and Treatment of Cancer (EORTC) guidance on standard of care MRI in STS (29–31).

The emerging field of “radiomics” may also serve to identify IBs for use in STS. Radiomics describes the extraction of features from radiological images, converting them to mineable high-dimensional data and searching for correlations with defined clinical variables. It relies on the extraction of both semantic (radiology lexicon such as size and shape) and agnostic (quantitative descriptors such as texture and histogram) features, interrogating tumour morphology and behaviour on a deeper scale (32, 33). Correlative assessment of these features with clinical variables may drive generation of diagnostic, prognostic or predictive models for variables of interest, to act as complimentary non-invasive decision support tools ( Figure 1 ) (24). Radiomics may identify an individual feature or set of features (a signature) that could be utilised as an IB. Radiomics is particularly attractive as it can explore and discover clinically relevant IBs without the need for new or complex imaging systems.

A typical radiomics workflow consists of multiple steps: high quality image acquisition, preprocessing of the images to ensure uniformity of scans, tumour segmentation and feature extraction. Extracted features are combined with other mineable data and correlations for clinical outcomes of interest are explored (24). Cancer types where radiomics has shown promise include lung cancer, glioblastoma and prostate cancer where radiomics has been used to distinguish benign from malignant tissue, measure the aggressiveness of tumours, aid diagnosis and selection of biopsy sites (24, 34–39). In addition, radiomics has identified imaging phenotypes suggestive of patient prognosis and predictions of treatment response (24, 34, 36–42). The use of radiomics in a highly heterogeneous malignancy like STS could deepen our understanding of the biology across the entire tumour volume, and its correlation with clinical outcomes. To date, preliminary radiomic studies in STS have focused on correlation of radiomic features with pathological grading and prognosis (43–47). Multiple studies have demonstrated an association of final histopathological grade in the surgical specimen on tissue examination with extractable radiomic features, including the ability to discriminate between tumours of low and high grade (43, 44, 47). In addition, radiomic features were predictive of distant metastases, early response to neoadjuvant chemotherapy and clinical outcome (46–50). The studies performed in STS remain limited to single centre cohorts with small patient numbers and imaging performed on single scanners. Radiomics requires large datasets to allow confident training of models and for robust conclusions to be drawn, and therefore, the rarity of STS and resultant small datasets has significantly hindered the progress of applying radiomics to “real life” clinical practice. Solutions for dealing with smaller datasets are emerging, such as using internal cross-validation to test radiomic models removing a dependence on large independent datasets (51). Overcoming this challenge would pave a bright future for radiomics in STS and its potential to supply unparalleled information about tumour heterogeneity and its impact on tumour behaviour and clinical outcome.

Quantitative MRI and radiomics require time-consuming processing and analysis and produce a vast amount of complex data which make interpretation challenging and prohibit translation as clinical decision making tools. Recently, AI is gaining traction within oncology as offering a complementary suite of technologies to drive uptake of IBs in clinical practice. Defined as the “development of computer systems able to perform tasks normally requiring human intelligence”, AI offers a solution to the time-consuming and complex image processing analyses required for some MRI parameters and for radiomics. It encompasses machine learning and its subset deep learning, each with multiple algorithms, and may prove an indispensable tool for complex data analysis (2, 52, 53). The integral radiomics workflow step of image delineation requires expert radiological input and is time consuming, particularly in the context of multiple tumour lesions or large, complex tumours. Deep learning can train and use algorithms for automated segmentation of tumours without the requirement for human intervention (54). In addition, it can analyse numeric data from predefined radiomic features as well as designing its own radiomics features from the direct analysis of images (55). This holds potential as a attractive tool, maximising the amount of “hidden data” that can be automatically extracted from radiological images, whilst addressing the real life constraints of time consuming segmentation and post processing.

The potential of biopsies to misrepresent grade in STS is well established (62). With regards to grading tumours, the primary goal of imaging has remained to achieve the equivalent of in vivo microscopy, or provide a complimentary diagnostic aide. In order to overcome the limitations of invasive core biopsies in STS, many MRI studies have sought to test the correlative power of imaging with tumour grade, comparing findings to histopathological examination of specimens based on FNCLCC. In 2008, Liu et al. correlated peripheral tumour growth pattern on MRI with histopathological grade in 59 STS patients, and found poorly defined margins in 60% of high grade tumours, versus well-defined margins in 60% of low grade tumours (66). This has been reinforced in a retrospective study by Zhao et al. where increased intensity of peritumoural contrast enhancement on imaging in higher grade tumours was the strongest independent indicator of histopathologically confirmed high grade STS (67). In addition, a study by Crombe et al. validated findings by Zhao et al. and demonstrated that the three following features of peritumoural enhancement, necrosis and heterogeneity in signal intensity on MRI could allow up-grading of underestimated tumour grade (61). This study also undertook survival analyses and determined that two out of three of these criteria were independent prognosticators of worse metastasis free survival and overall survival through uni- and multi-variable analysis (61).

Computed tomography (CT) sensitivity to detect tissue necrosis has also been utilised to demonstrate improved accuracy of grading in leiomyosarcoma (LMS) when combined with histopathological assessment of tissue, versus histopathology alone (68). Further to tumour grade, studies have explored the CT features of STS that could aid diagnosis, exploiting features such as fat content, margin status and presence of septations. An example of successful utilisation of these imaging parameters is in liposarcoma, a group of STS which can be frequently distinguished radiologically, without the requirement of a diagnostic biopsy (69). Contrast-enhanced CT, which is generally the standard investigation for retroperitoneal sarcomas, can be used for site confirmation and often tissue composition, and can confidently delineate well-differentiated liposarcoma (WDLPS) and angiomyolipoma (70, 71).

The value of PET/CT in assessing histological grade in STS has been explored. Using ¹⁸flurodeoxyglucose-positron emission tomography (¹⁸FDG-PET)/CT, the maximun standardised uptake value (SUVmax) of a lesion was demonstrated to correlate with the FNCLCC. In two meta-analyses, SUVmax was able to distinguish high-grade versus lower-grade STS such that this imaging technique could improve the diagnostic accuracy of core biopsies (72, 73). However, both meta-analyses reported poor methodological quality and lack of comparable parameters across studies, preventing confident conclusions to be drawn.

Advancements in MRI capabilities has allowed the derivement of quantitative parameters that allow insight into biological features of tumours for example fat, vascularity and cellularity, as well as the spatial heterogeneity across an entire tumour. These parameters may represent translatable IBs that could improve accurate diagnosis. Although the multitude of potential parameters is expanding, some remain in early stages of exploration ( Table 1 ). Exploiting MRI features of fat fraction and contrast sequences, liposarcomas can be distinguished from other STS, and provide information about the margin, shape, as well as internal architecture and properties (70, 75, 76). In a group of retroperitoneal sarcoma patients, MRI fat fraction strongly correlated with histopathological fat content on the surgical excision sample (22).

Schnapauff et al. in 2009 prospectively undertook DW-MRI in 30 patients prior to histological evaluation of regions of interest (ROIs), illustrating an inverse correlation of tumour cellularity and ADC, with low ADCs generated for areas of higher cellularity (77). This correlative relationship was not influenced by prior anti-cancer treatment. This finding was reproduced in two further retrospective studies, with lower mean ADC in higher grade sarcomas representative of the increased cellularity on histological analysis (78, 79). In a prospective study of 45 patients, the use of ADC to discriminate between malignant and benign soft tissue lesions was demonstrated. This study showed that malignant lesions had a significantly lower mean ADC value, further supporting the complimentary value of DW-MRI and ADC to soft tissue mass imaging (80). In addition, whilst most studies exploring ADC focus on the use of median or mean ADC values to objectively measure cellularity, this can fail to reflect baseline or post-treatment heterogeneity for any given tumour. Winfield et al. demonstrated ADC may be measured for various regions of interest on MRI, providing a deeper insight into ITH both within a region and across the entire tumour (22). An extension of DW-MRI is intravoxel incoherent motion (IVIM) which allows separation of perfusion from the true tissue diffusion. This allows calculation of quantitative parameters that reflect true tissue diffusivity (D), perfusion-related pseudodiffusion (D*) and perfusion fraction (f) which represents the contribution of water in capillaries (81). This could improve the accuracy of a quantitative parameter of diffusion by informing on the contribution of the microvascular circulation. Wu et al. demonstrated the combination of ADC and D were most useful in differentiating malignant and benign STS than ADC alone, and that the combination of ADC and f improved the differentiation between benign, intermediate and malignant STS (82). However, in Winfield et al’s study, the IVIM parameters exhibited poor repeatability when compared with ADC (22). A modification of DW-MRI is diffusion tensor imaging (DTI) and its advancement diffusion kurtosis imaging (DKI) (83, 84). These techniques add a quantitative assessment of the direction of diffusion of water molecules and may provide further diagnostic value by providing more structural information about lesions. DTI studies in muscoskeletal imaging are still in early stages but have been used to provide information about nerve fascicle visualisation for peripheral nerve sheath tumours and the integrity of axons (85–87). Similarly DKI studies remain evolutionary, however, have shown value in helping differentiate sarcoma from benign lesions and distinguishing recurrence from post surgical changes (88, 89).

Dynamic contrast-enhanced MRI (DCE-MRI) is being increasingly applied to obtain temporal information about tumour microvascularity and there have been a wealth of studies in the past few decades (90–94). It is well recognised that the tumour vascular system has an altered pathophysiology and role in tumourigenesis, supporting the diagnostic clinical application of DCE-MRI (95, 96). DCE-MRI utilises serial images capturing the arrival, duration and exit of a contrast agent in a tissue of interest, generating a signal intensity curve that may provide detailed information on both the physical and physiological properties of tissue in response to the agent (91, 97–99). The signal intensity curve and individual quantitative parameters such as the transport constant of the contrast agent from the blood plasma into the extracellular extravascular space (K^trans), and the extracellular extravascular space volume (V_e) may inform on underlying tissue microvasculature (99). For example tumour areas of higher perfusion and permeability having a higher and quicker uptake of contrast compared with tumours areas of necrosis (100). An early study in STS has demonstrated that DCE-MRI can aid in discriminating between low and high grade tumours, with K^trans differentiating grade I from grade III and grade II from grade III tumours (101). Combining K^trans with two semi-quantitative parameters also derived from DCE-MRI, initial area under the gadolinium concentration-time curve (iAUC) and time to peak (TTP), had the highest diagnostic performance with an area under the curve (AUC) of 0.841.

Magnetic Resonance Spectroscopy (MRS) is a technique which can provide information about the biochemical processes without the requirement for invasive intravenous contrast agents. By detecting protons (¹H-) containing metabolites in addition to water, this can provide information about the micro-metabolic environment within tissue. In STS, a prospective study used this technique to leverage on the measurement of choline which is an established marker for increased cell membrance turnover and can indicate malignancy. The authors showed that this can be useful in differentiating grade of muscoskeletal tumours (102). In addition, in a heterogeneous lesion like STS tumours, it might highlight more aggressive areas. However, this technique has poor sensitivity and specificity in STS and can be technically challenging is some anatomical sites such as extremities (103).

Recent efforts have centred on radiomic model development prediction of tumour grade given the crucial role of grade in preoperative diagnosis and treatment planning. Multiple studies have focused on identification of MRI-based radiomic features distinguishing low from high grade tumours (43, 45, 47, 78, 104). Using T2-weighted MRI images, Zhang et al. tested 5 radiomic features in a predictive model of STS grade in 35 patients with good accuracy (44). This was further explored by studies training machine learning algorithms based on radiomic features extracted from T2-weighted and/or T1-weighted MRI sequences to identify the optimal model of tumour grade prediction (43, 45). These studies used larger sample sizes (105 and 113 patients respectively) to form their complete cohort for training, testing and validation of their radiomic models, and achieved area under the curves (AUC) of over 0.9, demonstrating both feasibility and utility in their models for distinguishing grade (43, 45). In a study with the largest sample size to date, Navarro et al. retrospectively collected a training cohort of 148 patients to train a deep learning model comprising of features extracted from contrast-enhanced fat-saturated T1 and fat saturated T2 weighted images which was capable of distinguishing low from high grade tumours (105). They went on to externally validate this result in a cohort of 158 patients with AUCs of 0.75 and 0.76 for their T1 and T2 weighted radiomics models respectively. Corino et al. used DWI-MRI and it’s corresponding ADC maps in a group of 19 histologically confirmed STS patients for retrospective radiomic analysis to distinguish intermediate and high grade tumours. This study concluded that using a maximum of three features from first order statistics could achieve a model distinguishing grade with relatively good accuracy (78). Texture analysis tools can be applied to routine MR images to interrogate the inter-pixel relationships and grey level pattern of a ROI to provide a measure of heterogeneity (106, 107). This was utilised in a retrospective study of 29 patients by Meyer et al. to match Ki-67 index to discriminate between low and highly proliferating sarcomas (108). Ki-67 proliferation index is a widely used immunohistochemical test that reflects the activity of tumour cells, aiding the discrimination between low and high grade tumours, treatment response to neoadjuvant radiotherapy and can act as a predictor of clinical outcome in several cancers. Meyer et al. demonstrated several textural analyses correlated with Ki-67 index, namely run length and grey-level non uniformity (run length matrix features) had a positive correlation with Ki67 whilst sum average (grey-level histogram feature) and grey-level kurtosis (absolute gradient feature) had a negative correlation. Radiomics may be a valuable non-invasive tool for supplementing histological methods of distinguishing low and high grade STS, and may improve rates of undergrading tumours. However, MRI-based radiomic studies are usually retrospective, often limited to few imaging sequences and rarely externally validated, and therefore remain some way from constituting reliable and reproducible IBs. Despite its inferiority to MRI in soft tissue contrast, CT-based radiomics has also been explored in STS. Peeken et al. used three retrospective, independent cohorts for training, testing and validating (83, 87 and 51 patients respectively) a model to demonstrate feasibility of using CT-radiomics for grading STS, able to differentiate grade 3 from non-grade 3 tumours (47). There remains occasions, particularly for STS located within the retroperitoneum, when MRI is not conducted and therefore CT radiomics remains a useful avenue to explore. Whilst this study successfully validated its radiomics model, its AUC was 0.64, and therefore does not perform well enough to be a confident tool alongside the usual histopathological work-up for these patients.

Image guided biopsy could circumvent many issues related to tissue sampling, although this has undergone relatively little change for over 30 years (109). Biopsy usually takes place under CT or ultrasound (US) guidance alone, and whilst these modalities benefit from low cost, speed, and familiarity, they do not convey the same level of biological information as next generation imaging techniques (e.g. quantitative MRI).

In other tumour types including prostate cancer bone metastases, lymphoma and neuroendocrine tumours, functional imaging has been used to select the biopsy sites thought to be most deterministic in terms of biological behaviour, response to treatment and disease outcomes (110–112). Since tumour grade is known to be an important prognostic indicator of recurrence in sarcomas, this ‘precision biopsy’ approach could be adopted to provide more accurate grading and better inform treatment decisions e.g. alternative or adjuvant therapies (113). Practically speaking, a precision MRI biopsy requires either a direct ‘in-gantry’ approach, or some form of image fusion to leverage the benefits of the two imaging modalities - either ‘cognitive’ (visual estimation) or actual (software), the latter leveraging the benefits of two imaging modalities. Sampling tumours in multiple regions would afford the ability to capture intratumoural heterogeneity and can provide an accurate ground truth for imaging biomarker validation, with a view to “virtual biopsy” in the future. For example, biopsy in a single region may not capture the necrotic or myxoid elements, which are a key component of the FNCLCC grading scheme for adult sarcomas (114). A recent example of multiregional fusion biopsy has been reported by a group who sampled different regions of ovarian tumours using CT/US fusion, guided by radiomic habitats (115). Precision biopsy would also facilitate discovery of IBs with radiogenomic study of co-localised tissue samples, integrating data-rich radiological images with biological profiling inclusive of deeper histological and molecular metrics. There is research applying “-omics” approaches such as transcriptomics and proteomics to comprehensively profile STS and their functional status, particularly in the context of heterogeneity (116–118). Further, we are increasingly understanding the importance of the role the immune environment plays within these heterogeneous tumours, and its potential to inform on behaviour and clinical outcome of STS (119–122). Combining these powerful –omics platforms with next generation imaging and biopsy techniques could accelerate the discovery and validation of IBs in STS.

Studies involving ¹⁸F-FDG PET/CT have shown value in using the reduction in SUVmax following treatment to indicate possible therapeutic response even when overall tumour volume has not reduced (74, 156–158). Benz et al. demonstrated that a decrease in SUVmax of ≥ 35% from baseline to post-treatment scans was a sensitive predictor of histopathological response to neoadjuvant therapy in a group of 50 patients with resectable high grade STS (156). Schuetze et al. demonstrated a reduction in SUVmax of ≥ 40% was an independent predictor of improved prognosis (disease specific and overall survival) in 46 patients with high grade sarcoma receiving neoadjuvant chemotherapy (157). However, these PET/CT studies are limited to small patient numbers, single centres and are usually retrospective and therefore, confidence in SUVmax as a clinical IB will be dependent on ongoing prospective studies.

PET/MRI has also shown promise for monitoring of patients, adding the excellent soft tissue contrast possible with MRI to information obtained through PET. A study by Erfanian et al. reported improved accuracy in the detection of local recurrence of STS of 90% when compared to conventional MRI alone at 83% (159). Like PET/CT, progress for this imaging modality will rely on robust prospective studies. A further exciting avenue for PET research in STS is immuno-PET imaging as a potential non invasive and serial tool to monitor pre-selected immune markers or aid the selection of patients for immunotherapy. This remains in early stages, however, a study demonstrated feasibility of using a radiotracer and immuno-PET to detect and monitor programmed cell death protein 1 (PD1, an immune checkpoint protein) in humanised mouse models (160).

Frequently ADC values change following radiotherapy signifying possible treatment response, even when overall tumour volume remains unchanged (22, 161–163). Winfield et al. also illustrated that in the same group of retroperitoneal STS, patient MRIs following radiotherapy demonstrated an increase in ADC, perhaps reflecting biological changes to a less cellular tumour suggestive of response. In this study, ADC was shown to have excellent repeatability, with a coefficient of variation (CoV) of 2.5% for median ADC (22). The repeatability of ADC adds to the confidence of its usage as a robust, standardised IB. However, it is important to note that at present, there remains no standardised method to measure ADC in STS, particularly with relation to therapeutic response (164). This can range from variations in the acquisition of diffusion-weighted images, the ROI selection process (for example 2D ROIs versus 3D volume ROIs) and the ADC parameter selected (for example minimum ADC versus mean ADC) (22, 77, 162). As such, this hinders our ability to standardise study findings and like other IBs, it also suffers from lack of reproducibility and validation efforts in prospectively designed studies.

DCE-MRI has also been used to demonstrate changes following therapy, suggestive of possible response (98, 165–168). Recently, in a subset of 10 patients with myxoid liposarcoma from the Dose Reduction in Preoperative Radiotherapy in Myxoid Liposarcomas (DOREMY) multicentre clinical trial, a higher baseline K^trans was linked to response following neoadjuvant radiotherapy and may allow prediction of early response to radiotherapy (169). Although time consuming analysis and some limitations on anatomical coverage preclude routine clinical use, DCE-MRI represents an interesting avenue for further research into the STS vascular microenvironment and its role in therapeutic response.

Oxygen enhanced (OE-)MRI is an emerging technique that may offer insight into tumour hypoxia; a well-recognised negative prognostic factor in cancer (170–172). To date, there remains an unmet need to measure hypoxia through a biomarker, and MRI offers an attractive non-invasive and serial option (173). OE-MRI is capable of mapping and quantifying the spatial distribution of tumour oxygen delivery. By measuring the longitudinal relaxation rate of hydrogen nuclei (R₁), which changes in relation to the dissolved oxygen concentration in tissue or plasma, OE-MRI can characterise well- and poorly-oxygenated tissue. Inhalation of hyperoxic gas results in delivery of excess oxygen to well-oxygenated tissue, however, given well- oxygenated tissue will have saturation of haemoglobin molecules with oxygen, the excess oxygen remains within the plasma and interstitial tissue fluid, increasing the tissue R₁. Preclinical and clinical studies in renal cell carcinoma, glioma, prostate cancer and lung cancer have demonstrated the feasibility and accuracy of OE-MRI in mapping hypoxic subregions, including changes following radiotherapy (173–179). Whilst human studies in sarcoma are lacking, the early work by Cao-Pham et al. in rhabdomyosarcoma xenograft models showed that R₁ as measured by OE-MRI is sensitive to changes in oxygenation (180). Further exploration in STS is required, however in a group of heterogeneous cancers, OE-MRI could optimise our treatment planning for these patients. Visualising areas of hypoxia across a tumour could aid radiotherapy planning, stratifying patients accordingly.

Similarly, magnetic resonance elastography (MRE) is gaining attention as a quantitative imaging method capable of assessing the mechanical properties of tissue. MRE informs on the viscoelastic properties of tissue by measuring the propagation of mechanical waves through tissue (181). Already a well-established technique in patients with chronic liver disease, there is a strong motivation to explore the applicability of MRE to other pathologies (182, 183). Preclinical studies have demonstrated that altered tissue “stiffness” (quantified using the complex shear modulus (G*) and related parameters) can be used to indicate malignant masses within brain parenchyma, and that changes in mechanical properties of a tumour may precede volume change following treatment indicative of response in colon cancer and non-Hodgkin’s lymphoma (181, 184–186). There are currently limited studies within STS, however, Pepin et al. demonstrated technical feasibility of MRE in a pilot study of 13 sarcoma patients, and possible utility in assessment of early response to radiation therapy (in 4 of these 13 patients) (187). Further studies are underway, and this may offer a further non-invasive method to interrogate deep and complex masses such as those within the retroperitoneum.

Much of the radiomic research in STS is focused on predicting therapeutic response and clinical outcome. In an externally validated study, Spraker et al. found that features extracted from T1-weighted MRI were independently associated with overall survival (OS), and a radiomics model combined with clinical features (age and grade) performed best (46). In a further study, the addition of T2-weighted MR radiomics to the clinical American Joint Committee on Cancer (AJCC) staging system into a nomogram improved the clinical net benefit for stratification of patients for OS, and was superior to using AJCC system alone (188). The same group went on to apply deep learning to the two cohorts (that had been slightly expanded), concluding that the addition of deep learning to their previous radiomic model yielded comparable prediction performance for OS. This finding was supported by a further nomogram constructed by Yan et al, demonstrating successful patient risk stratification based on progression free survival (PFS) when MR-based radiomics was used in addition to clinical factors, with an increased benefit to using AJCC staging system alone (189). In CT, the same study by Peeken et al. using radiomics to distinguish tumour grade in STS explored the predictive performance of radiomic features for OS, distant progression free survival (DPFS) and local progression free survival (LPFS). This study concluded that again, despite CT’s inferior soft tissue contrast, there was value in using a radiomics model to stratify patient risk based on their survival outcomes. This finding was successfully validated in an external cohort, and results were more consistent than using a clinical model (47).

In the context of response to therapy, delta-radiomics, the change in a given radiomics feature in a set of longitudinal images is of particular interest. Gao et al. used longitudinal DW-MRI images in 30 STS patients receiving neoadjuvant RT at three time points, to develop a predictive model of pathological therapeutic response (190). The model demonstrated that radiomic features alone were not sufficient to predict response, however, the inclusion of delta-radiomics features at mid- and post-therapy time points optimised prediction of response to RT. Delta-radiomics was also used for prediction of pathological response to neoadjuvant chemotherapy in a retrospective cohort of 65 STS patients. A radiomics model was developed based on longitudinal MRI scans taken at baseline and following two cycles of anthracycline-based therapy, prior to surgical excision from 50 of the 65 patients. The findings were then validated in the remaining 15 patients. Using 3 STS features of shape, heterogeneity and surrounding tissue, the study concluded an improved performance of the radiomics model when compared to RECIST 1.1, and a promising tool for the evaluation of early response following only two cycles of chemotherapy (50).

Previous studies have utilised quantitative parameters derived from diffusion weighted imaging, namely ADC, the transverse relaxation time (T₂) and proton density (M₀) to objectively segment tumours into distinct subregions and in relation to therapy (191–195). In a xenograft mouse model of radiation-induced fibrosarcoma (RIF-1), Henning et al. was able to segment tumours into regions based on two viable tumour groups and two non viable (necrotic) tissue groups which were confirmed on histopathological staining of a hypoxia marker (192). The group also analysed the viable tissue groups further and characterised them into well oxygenated radiosensitive and hypoxia and radioresistant groups (193). This is important ground work for methods of objectively characterising tissue based on quantitative markers for underlying biological properties. Building on this are habitat maps which incorporate multiple quantitative parameters obtained from numerous MRI sequences. These maps characterise tissue into distinct biological habitats based on their combination of multiparametric MRI parameters and provide insight into multiple aspects of biological heterogeneity and their interaction with one another (21). In a cohort of 30 retroperitoneal STS, automated segmentation of tumours into tissue subtypes based on fat fraction (FF), enhancement fraction (EF) and ADC, demonstrated distinct habitat maps and changes following radiotherapy suggestive of response, including when overall tumour volume remained the same or increased ( Figure 3 ). These techniques may improve our understanding of the differing characteristics within this heterogeneous cancer and how distinct regions respond variably to the same treatment. In addition, it can illustrate results in a timely and clinician friendly manner that is not reliant on interpretation of numerical values, easing its transition into practice.

The integration of powerful imaging technology to inform on the biology of these heterogeneous tumours is needed in the clinical setting for this complex disease. Indeed, there is potential to enable a more personalised approach to patient management and uncover candidate therapeutic options. However, there are a number of challenges that need to be addressed prior to the successful implementation of “virtual biopsy” methods into practice. If the ultimate goal of quantitative and radiomic imaging techniques is the development of an IB, consideration of the steps of biomarker development is required (5). At present, despite a heavy investment into the search for IB discovery, most sarcoma work has faltered at the initial stage of technical (assay) validation (22).

Quantitative MRI sequences have a varying complexity of delivery, particularly those which require intricate post-processing of images, and therefore, make reliable and confident translation across radiology departments challenging. This is mirrored in the variation seen across different MRI scanners in image acquisition and measurement output, usually requiring adaptations at every centre and vendor. This weakens the reproducibility potential of quantitative MRI techniques and may oblige different centres to perform reproducibility/repeatability studies to conclude some clinical capability. As a result, most studies remain limited to a single centre using the same MRI machine. This hinders the progression of novel and promising quantitative MRI techniques through imaging biomarker validation. Similarly, a crucial step in the radiomics workflow is standardisation of acquired images to allow reliable image delineation and radiomic feature extraction, and this can pose significant challenges at multi-centre level. Each step in the radiomics workflow can in fact affect results, and similar to quantitative MRI, requires careful or expert processing of data. This makes reproducibility efforts challenging. To date, few studies have spanned multiple centres, and the majority of studies have not been tested in an independent external cohort. Attempts to provide standardised guidance for the translation of acquired imaging to high-throughput IBs are underway. The Image Biomarkers Standardization Initiative (IBSI), an independent international collaboration aims to provide detailed guidance on the mandatory steps for radiomic analyses, and Lambin et al. have defined the Radiomics Quality Score (RQS) to provide an objective estimate of the quality of the study (196, 197). Future studies may benefit from these to aid transition from proof of concept to real life clinically applicable tools.

STS is a disease of rarity, and it’s diverse histological subtype profile confounds effort to accrue patient cohorts comparable to other cancer types. This can make assessment of the quality of findings within imaging studies challenging. In particular, radiomics is dependent on large datasets for reliable feature extraction and robust conclusions to be drawn from results. Furthermore, a crucial component of translating an IB is it’s independent validation in an external cohort, which is a recognised challenge in the sarcoma community. Whilst there are methods to deal with smaller datasets such as cross-validation, these still remain limited to the individuals studied and findings cannot be generalised to the rest of the population, not overcoming the issues pertaining to external validation (198).

A solution to this may be data-sharing. A powerful drive through the NIH’s Quantitative Imaging Network (QIN) and its leveraging on The Cancer Imaging Archive (TCIA) has demonstrated feasibility in data sharing of clinical images across various sites (199). This partnership holds promise to overcome challenges related to data-sharing, facilitating multi-site collaboration and support the development of IBs through validation (200–202). Builiding on this, there are efforts to harmonise quantitative MRI data across centres and develop a robust infrastructure to allow this, such as the National Cancer Imaging Translational Accelerator (NCITA) consortium which includes 9 UK centres aiming to develop a robust multicentre IB pipeline (65). This would provide opportunities for both multi-centre data to increase the size of initial datasets and external cohorts for the validation of discoverable IBs.

The challenges related to the low incidence and heterogeneity of the disease can in turn make prospective collection of data difficult. It can take multiple years to recruit a desired cohort size, and studies may be closed early upon failing to reach target accrual. As such, alongside the readily available retrospective data, many imaging studies within STS remain retrospective in nature. This carries risk of confounding and bias (such as selection or recall bias), and results in an inferior level of evidence. Furthermore, retrospective designs negate the ability to carry out repeatability tests for quantitative MRI and radiomic studies due to a lack of repeat scans performed around the time of the initial scan. Going forward, confident validation and translation of IBs is heavily dependent on prospective studies.

Within STS, a Cancer Research United Kingdom (CRUK) Sarcoma Accelerator consortium partnering centres across the UK, Spain and Italy aims to develop a multicentre, expansive data pipeline to develop clinical and biological models to further our understanding of sarcoma biology in patients with high risk STS undergoing perioperative management. This will involve prospective collection of imaging data across 5 years and spanning these multiple centres. This will greatly benefit IB discovery and validation in sarcoma, where limited patient numbers for studies will increasingly rely on close multi-centre collaborations and data-sharing to propel their progress to clinical translation.