Pediatric soft tissue sarcomas are heterogeneous rare tumors with rhabdomyosarcoma being the most frequent histotype. The recent creation of international working groups allowed the development and implementation of risk-adapted protocols and better supportive care which, overall, increased the survival rates.1,2 Improvements in this oncological field have been accompanied by technical developments in diagnostic imaging. In fact, in the last decade, the use of quantitative magnetic resonance (MR) techniques like diffusion wighted imaging (DWI), dynamic contrast enhanced MR imaging (DCE-MRI), and MR spectroscopy as well as the application of complex volumetric histograms analyses, significantly contributed to primary tumours characterization.3 In adults affected by soft tissue sarcomas, also radiomics demonstrated to provide advanced information about lesions histology and grade.4,5 As part of the current technical innovations that we can apply in our practice, [18F] fluorodeoxyglucose (FDG) positron emission tomography (PET)/MR (PET/MR) has started to be introduced in the diagnostic work-up of adults and children with soft tissue sarcomas.6 In fact, this technique not only provides metabolic information with a lower radiation dose than PET/computed tomography (PET/CT) but also allows the simultaneous collection of quantitative MR data.1 Despite the growing evidence in this diagnostic field, up to now, the role of radiomics features extracted from PET/MR images for pediatric soft tissue sarcomas has not been investigated yet.
The aim of this study was two-fold: i) assess the prognostic value of radiomic features extracted from T2 weighted MR images of the primary tumour of paediatric patients affected by soft tissue sarcomas examined with PET/MR; ii) evaluate if such features may act as biomarkers of tumour grade and histotype.
Children with histology proven soft tissue sarcomas who underwent PET/MR for staging in our tertiary center from January 2016 to March 2021, including in the protocol a whole-body axial Half-Fourier Acquisition Single-shot Turbo spin Echo imaging sequence (HASTE: Echo Time 95 ms, Repetition Time 1600 ms, and 5 mm slice thickness) were included in this retrospective Institution Review Board approved study. One radiologist with four years of experience in oncological imaging used an open-source software (3D Slicer, www.slicer.org) to segment each primary lesion along tumours margins including the entire volume (Figure 1). The same software was applied to extract 33 radiomic features of first and second order. For each included patient, indices of metabolic activity simultaneously acquired by PET (
Graphic representation of the development of the proposed radiomic analysis demonstrating the segmentation and extraction of radiomic features from the primary lesion of children affected by soft tissue sarcomas examined for staging by PET/MR and the factor analysis that led to the selection of five highly correlating variables.
Factor analysis was used to select highly correlating features and logistic regression analysis was applied to determine the prognostic value of the selected radiomic features on the overall outcome. The Student’s t-test and the Chi-square test (for continuous and categorical variables, respectively) were used to evaluate if differences in radiomic, demographics, metabolic, and laboratory variables occurred according to tumour grade (subdivided as low and high grade, the former comprising low and intermediate grades without metastatic spread) and histotype (grouped as rhabdomyosarcomas and non-rhabdomyosarcomas). The applied level of significance for all statistical analyses was p<0.05. If a statistically significant difference occurred, then the diagnostic value of the variable/s was computed using the receiver operating curves and the results expressed as accuracy, sensitivity, and specificity. To test the reliability of the method, a second radiologist with ten years of experience in musculoskeletal imaging repeated all segmentations and data extraction. Then, for the features selected by the factor analysis, the intraclass correlation coefficient (ICC) was computed using the two-way random effects, consistency, two raters method. Values above 0.750 were considered as excellent.7 All statistical analyses were performed using SPSS Statistics 25.0 (IBM Corp, Armonk, NY, USA).
Eighteen children (11 female; mean age 7.8 ± 4.6 years old) matched the inclusion criteria and were examined. Twelve (66.6%) children were affected by rhabdomyosarcoma; nine patients had high grade lesions and five showed metastatic spread. Overall, four patients deceased (22.2%).
The factor analysis allowed the selection of five highly correlating features: minimum, lmc1, cluster shade, long run emphasis, and variance (Figure 1). None of the features showed a significant prognostic role on the overall outcome (p > 0.05, each). The feature lmc1 was significantly higher in low grade soft tissue sarcomas (-0.17 ± 0.05
None of the other variables, including metabolic parameters, showed any statistically significant difference in any of the performed comparisons (p > 0.05, each) (Table 1).
Comparison of demographics, metabolic, and first-line laboratory variables according to tumor histotype and grade
Histotype* | Grade# | |||||
---|---|---|---|---|---|---|
Rhabdomyosarcoma | Non-rhabdomyosarcoma | Low-grade | High grade | |||
Age (years) | 7.2 ± 4.8 | 9 ± 4.2 | 0.464 | 6.1 ± 3.4 | 9.5 ± 5.2 | 0.115 |
Gender (female/male) | 7/5 | 4/2 | 1.000 | 7/2 | 4/5 | 0.335 |
SUVmax | 5.1 ± 3 | 5.6 ± 3 | 0.765 | 5.3 ± 4 | 5.3 ± 2 | 0.998 |
SUVmean | 1.8 ± 1 | 1.5 ± 1 | 0.505 | 1.1 ± 0.3 | 1 ± 0.2 | 0.960 |
Hemoglobin (g/L) | 11.4 ± 1.3 | 11.7 ± 1.4 | 0.660 | 11.7 ± 1.6 | 11.4 ± 1 | 0.657 |
Red (×10blood 12· |
4.2 ± 1 | 4.7 ± 0.5 | 0.215 | 4.3 ± 1 | 4.5 ± 1 | 0.594 |
White (×109· |
5.8 ± 3 | 5.8 ± 5 | 0.996 | 5.7 ± 4 | 6 ± 3.9 | 0.889 |
Platelets (×109· |
349 ± 171 | 386 ± 77 | 0.636 | 375 ± 187 | 351 ± 100 | 0.748 |
Lactate (U/L) dehydrogenase | 335 ± 86 | 369 ± 172 | 0.608 | 358 ± 131 | 331 ± 94 | 0.652 |
D-(ng/dimer mL) | 214 ± 152 | 2125 ± 3530 | 0.07 | 1370 ± 2992 | 331 ± 297 | 0.315 |
*rhabdomyosarcoma
We presented the first PET/MR-based application of radiomics for paediatric soft tissue sarcoma which demonstrated that the radiomic features lmc1 and variance extracted from a T2w echo-planar fast spin echo sequence can distinguish between paediatric soft tissue sarcomas of different grade and histotype, respectively. Similar results were obtained by Corino
On the other hand, our results are partially in disagreement with previous evidence in terms of prognosis.5,9 Indeed, for instance, Crombé
This initial study is affected by several limitations. Firstly, we did not perform any radiomic analysis of the PET dataset because of the well-known challenges associated with PET-based radiomic analysis (
In conclusion, specific radiomic features seem to act as biomarkers of pediatric soft tissue sarcomas grade and histotype. Further multicenter studies are expected to confirm this preliminary evidence and assess its role on the diagnostic work-up and therapeutic management of this group of patients.