CT and MRI radiomics of bone and soft-tissue sarcomas: a systematic review of reproducibility and validation strategies

No Local Ethics Committee approval was needed for this systematic review. Literature search, study selection, and data extraction were performed independently by two recently boarded radiologists with experience in musculoskeletal tumors and radiomics (S.G. and F.M.). In case of disagreement, agreement was achieved by consensus of these two readers and a third reviewer with radiology specialty and doctorate in artificial intelligence and radiomics (R.C.). The Preferred Reporting Items for Systematic reviews and Meta-Analyses (PRISMA) guidelines [11] were followed.

An electronic literature search was conducted on EMBASE (Elsevier) and PubMed (MEDLINE, U.S. National Library of Medicine and National Institutes of Health) databases for articles published up to December 31, 2020, and dealing with CT and MRI radiomics of bone and soft-tissue sarcomas. A controlled vocabulary was adopted using medical subject headings in PubMed and the thesaurus in EMBASE. Search syntax was built by combining search terms related to two main domains, namely “musculoskeletal sarcomas” and “radiomics.” The exact search query was: (“sarcoma”/exp OR “sarcoma”) AND (“radiomics”/exp OR “radiomics” OR “texture”/exp OR “texture”). Studies were first screened by title and abstract, and then, the full text of eligible studies was retrieved for further review. The references of selected publications were checked for additional publications to include.

Inclusion criteria were: (1) original research papers published in peer-reviewed journals; (2) focus on CT or MRI radiomics-based characterization of sarcomas located in bone and soft tissues for either diagnosis- or prognosis-related tasks; (3) statement that local ethics committee approval was obtained, or ethical standards of the institutional or national research committee were followed.

Exclusion criteria were: (1) papers not dealing with mass characterization, such as those focused on computer-assisted diagnosis and detection systems; (2) papers dealing with head and neck, retroperitoneal or visceral sarcomas; (3) animal, cadaveric or laboratory studies; (4) papers not written in English language.

Data were extracted to a spreadsheet with a drop-down list for each item, as defined by the first author, grouped into three main categories, namely baseline study characteristics, radiomic feature reproducibility strategies, and predictive model validation strategies. Items regarding baseline study characteristics included first author’s last name, year of publication, study aim, tumor type, study design, reference standard, imaging modality, database size, use of public data, segmentation process, and segmentation style. Those concerning radiomic feature reproducibility strategies included reproducibility assessment based on repeated segmentations, reproducibility assessment related to acquisition or post-processing techniques, statistical method used for reproducibility analysis, and cut-off or threshold used for reproducibility analysis. Finally, data regarding predictive model validation strategies included the use of machine learning validation techniques, clinical validation performed on a separate internal dataset, and clinical validation performed on an external or independent dataset.

A flowchart illustrating the literature search process is presented in Fig. 1. After screening 278 papers and applying our eligibility criteria, 49 papers were included in this systematic review. Tables 1 and 2 detail the characteristics of papers dealing with radiomics of bone (n = 12) and soft-tissue (n = 37) tumors, respectively.

All studies were published between 2008 and 2020. Twenty-three out of 49 investigations (47%) were published in 2020, 14 (29%) in 2019, 4 (8%) in 2018, and 8 (16%) between 2008 and 2017. The design was prospective in 6 studies (12%) and retrospective in the remaining 43 (88%). The imaging modality of choice was MRI in 42 (86%), including one or multiple MRI sequences, and CT in 7 (14%) cases. The median size of the database was 60 patients (range 19–226). Public data were used only in 3 (6%) studies.

The research was aimed at predicting either diagnosis or prognosis, as follows: benign versus malignant tumor discrimination (n = 14); grading (n = 10); tumor histotype discrimination (n = 4); proliferation index Ki67 expression (n = 1); survival (n = 12); response to therapy, either chemotherapy or radiotherapy (n = 8); local and/or metastatic relapse (n = 9). It should be noted that the aim was twofold in some studies, as detailed in Tables 1 and 2. In those focused on diagnosis-related tasks, including benign versus malignant discrimination, grading, tumor histotype discrimination, and proliferation index expression, histology was the reference standard in all cases excepting benign lesions diagnosed on the basis of stable imaging findings over time in two papers [12, 13]. In studies focused on prediction of response to chemotherapy or radiotherapy, the reference standard was histology if lesions were surgically treated, based on the percentage of viable tumor and necrosis relative to the surgical tissue specimen, or consistent imaging findings if lesions were not operated. In studies focused on prediction of tumor relapse, the diagnosis was based on histology or consistent imaging findings, as the reference standard. In studies dealing with survival prediction, survival was assessed based on follow-up.

Regarding segmentation, the process was performed manually in 45 (92%) studies and semiautomatically in 4 (8%) studies. In no case, the segmentation process was fully automated. The following segmentation styles were identified: 2D without multiple sampling in 11 (23%) studies; 2D with multiple sampling in 3 (6%); 3D in 35 (71%). Of note, a single slice showing maximum tumor extension was chosen in all studies employing 2D segmentation without multiple sampling, excepting one case [14] where it was chosen based on signal intensity homogeneity.

Eighteen (37%) of the 49 studies included a reproducibility analysis of the radiomic features in their workflow. In 16 (33%) investigations [13, 15‐29], the reproducibility of radiomic features was assessed on the basis of repeated segmentations performed by different readers and/or the same reader at different time points. Two (4%) studies presented an analysis to assess the reproducibility based on different acquisition [30] or post-processing [31] techniques. Of note, segmentations were validated by a second experienced reader in 15 studies [12, 32‐45] without, however, addressing the issue of radiomic feature reproducibility.

The intraclass correlation coefficient (ICC) was the statistical method used in most of the papers reporting a reproducibility analysis [13, 15‐18, 20, 22‐25, 27‐29, 31]. ICC threshold ranged between 0.6 [13] and 0.9 [22] for reproducible features. The following statistical methods were used less commonly: analysis of variance [30, 31]; Cronbach alpha statistic [26]; Pearson correlation coefficient [19], and Spearman correlation coefficient [21].

At least one machine learning validation technique was used in 25 (51%) of the 49 papers. K-fold cross-validation was used in most of the studies [13, 25, 28, 31‐33, 37, 38, 40, 43, 44, 46‐50]. The following machine learning validation techniques were used less commonly: bootstrapping [42, 51]; leave-one-out cross-validation [34, 35, 41]; leave-p-out cross-validation [52]; Monte Carlo cross-validation [23]; nested cross-validation [25, 27]; random-split cross-validation [20]. Figure 2 provides an overview of machine learning validation techniques. Figure 3 illustrates an example of a radiomics-based machine learning pipeline.

A clinical validation of the radiomics-based prediction model was reported in 19 (39%) of the 49 papers. It was performed on a separate set of data from the primary institution, i.e., internal test set, in 14 (29%) studies [15, 16, 22, 24, 28, 31, 32, 35, 37, 38, 41, 46, 47, 52]. It was performed on an independent set of data from the primary institution (related to a different scanner) or from an external institution, i.e., external test set, in 5 (10%) studies [25, 27, 29, 43, 51].

MRI and CT radiomics of bone and soft-tissue sarcomas have progressively gained attention in musculoskeletal and oncologic imaging. The number of papers has rapidly increased over the recent years, and almost half of those included in our review (47%) was published in 2020. Radiomics was used in attempt to answer clinical questions related to both diagnosis and prognosis of musculoskeletal sarcomas. Most studies (88%) were retrospective in nature, as this design allowed including relatively large number of patients with imaging data already available and bone or soft-tissue sarcomas, which are rare diseases. A prospective analysis, while not strictly necessary in radiomic studies [5], may, however, have advantages for controlling data gathering in reproducibility assessment and matching certain patient or imaging characteristics in independent datasets. Public data were used in no study regarding bone sarcomas and in a small proportion of the studies (6%) concerning soft-tissue sarcomas. A public database [55] available on The Cancer Imaging Archive (https://​www.​cancerimagingarc​hive.​net) was used in all these studies. Public databases afford opportunities for researchers who do not have sufficient data at their institution and allow research groups from around the world to test and compare new radiomic methods using common data. Thus, research employing radiomics in this field would certainly be enhanced if further imaging databases are made publicly available in the near future.

Regarding segmentation, the process was performed manually in most of the studies (92%) and semiautomatically in the remaining, both requiring human intervention to some extent. Even though the influence of inter-observer and/or intra-observer variability on the reproducibility of radiomic features can be assessed as part of the radiomic workflow, fully automated segmentation algorithms would ideally achieve higher reliability and deserve future investigation. Annotations included the entire lesion volume (3D segmentation) in most of the studies (71%) and a single slice (2D), without multiple sampling, in the remaining (23%). However, to date no study has compared the outcome of 2D and 3D segmentations in musculoskeletal sarcomas. As 2D annotations are time saving and have recently proven higher performance than 3D segmentation in oropharyngeal cancers [56], this should represent another area of research in the near future. Of note, a limited number of studies (6%) used a 2D segmentation style with multiple sampling as a data augmentation technique to increase the number of labeled slices [26, 48, 57]. This practice can be useful for an uncommon entity as musculoskeletal sarcomas but should be employed with care to avoid the introduction of bias in the final model. The inclusion of samples from the same case in both the training and test sets could lead to overly optimistic results.

A great variability in radiomic features has emerged as a major issue across studies and attributed to different segmentation, image acquisition, and post-processing approaches [4]. Therefore, methodological analyses are advisable prior to conducting radiomic studies in order to assess feature robustness and avoid biases due to non-reproducible, noisy features. This concept is in line with recent literature emphasizing the importance of reproducibility in artificial intelligence and radiology [58]. In our review, we noted that about one third of the included papers described a reproducibility analysis in their workflow. In this subgroup of papers, inter- and/or intra-reader segmentation variability was the main focus of the reproducibility analysis. Segmentation variability-related analyses outnumbered those addressing reproducibility issues due to image acquisition or post-processing differences, which were reported in one paper per each [30, 31]. This finding underlines that further research should deal with dependencies of radiomic features on image acquisition and post-processing. While these analyses may already be performed in retrospective series, when patients underwent more than one study in a short interval, prospective studies could facilitate the identification of reliable radiomic features within this domain. Finally, ICC was the statistical method used in most of the papers evaluating radiomic feature reproducibility. Of note, guidelines for performing and assessing ICC are available and can be followed to achieve consensus on the cutoff and threshold values [59].

Proper validation of radiomic models is highly desirable to bridge the gap between concepts and clinical application [53]. Machine learning validation techniques are employed to avoid any information leak from the test to the training set during model development [60]. Resampling strategies can be extremely useful, especially with relatively limited samples of data, which may not be truly representative for the population of interest, with the aim of reducing overfitting and better estimating the performance of the radiomics-based predictive model on new data (i.e., the test set) [61, 62]. K-fold cross-validation was the most commonly used technique for this task in the studies included in this review.

Ideally, in both prospective and retrospective studies, a clinical validation of the model is performed against completely independent sets of data, i.e., the external or independent test set [4]. We found that clinical validation was performed against an independent dataset from the primary institution (using different scanners) or from a different institution only in a small number of studies (10%) included in this systematic review. More studies (29%) validated the model using a separate set of data from the primary institution, i.e., an internal test set. Therefore, future studies should be carried out in more than one institution and include external testing of the model with large and independent sets of data.