Radiomic features from MRI distinguish myxomas from myxofibrosarcomas

There are several benign and malignant myxoid soft tissue tumors. Benign myxoid tumors include myxomas and angiomyxomas; and malignant myxoid tumors include fibromyxoid sarcomas, extraskeletal myxoid chondrosarcomas, ossifying fibromyxoid tumors, myxoid liposarcomas, myxoinflammatory fibroblastic tumors and myxofibrosarcomas [1]. Each myxoid neoplasms has key chromosomal translocations or immunohistochemical markers that are pathognomic for its diagnosis except for myxomas and myxofibrosarcomas [1‐8]. Myxofibrosarcomas and myxomas are not associated with any particular translocation or gene expression product, and their diagnoses are based on their histological appearances [1, 8‐10]. Therefore, one of the greatest diagnostic dilemmas for a pathologist lies in differentiating a myxoma from a myxofibrosarcoma, and particularly differentiating a cellular myxoma from some low-grade myxofibrosarcomas [1, 8‐10]. Differentiation between these two entities is based on morphologic and histologic criteria [1, 8‐10]. The challenge for pathologists to differentiate these two entities increases with core needle biopsies because of the heterogeneity of myxofibrosarcomas and because the core biopsy is subject to sampling error [11‐13].

Differentiating myxomas from myxofibrosarcomas is also challenging for radiologists because these lesions have overlapping imaging features - namely they are both hyperintense on T2-weighted magnetic resonance imaging (MRI) sequences, have variable signal intensity on T1-weighted sequences, and both have heterogeneous enhancement patterns [14‐18]. Myxomas have been reported to be isointense or hypointense to skeletal muscle using T1-weighted sequences [15]. Myxofibrosarcomas have been reported to be isointense to skeletal muscle on T1-weighted sequences [18, 19]. The difference is clinically significant because the surgical approach and treatment is different for these two entities. MRIs provide global assessment of the tumor whereas biopsies are limited to certain areas of the tumor; therefore, imaging is likely to be less subject to sampling error.

We hypothesized that quantitative analysis using radiomic feature extraction and classifier model analysis from preoperative MRI studies could predict whether a myxoid tumor is likely to be a myxoma or myxofibrosarcoma over volume-based and MRI signal intensity (SI) value analysis. The purpose of this study was to assess the performance of image intensity information, tumor volume, and radiomic features extracted from MRI for distinguishing myxomas from myxofibrosarcomas.

There were a total of 56 patients identified: 29 with myxomas, 27 with myxofibrosarcomas. Subject demographic and clinical variables are shown in Table 1. There was a higher proportion of female patients with myxomas than myxofibrosarcomas, but the difference was not significantly different (p = 0.186 by Pearson’s Chi-squared test). None of the patients had a diagnosis of fibrous dysplasia.

As shown in Fig. 2, myxomas had lower normalized T1-weighted signal intensity values than myxofibrosaromas (p = 0.006, Wilcoxon rank sum test), and AUC was 0.713 as shown in Fig. 3. Figure 4 shows the T1 SI of the myxofibrosarcomas by tumor grade. There was no substantial difference in the T1 SI between the myxofibrosarcomas by tumor grade (Kruskal-Wallis p = 0.88). The radiomics features for all the subjects were demonstrated in Fig. 5. The classification model built upon radiomic features obtained an AUC of 0.885 (accuracy = 0.839, sensitivity = 0.852, specificity = 0.828), which outperformed the classification model built upon the T1SI values (p = 0.039, DeLong test), and the classification model built upon volume features (AUC = 0.838, p = 0.285 by DeLong test) as shown in Fig. 3.

To investigate how different features contributed the classification, the top 15 features with high importance regarding classification are demonstrated in Fig. 6. Seven of them were shape-based measures, indicating high association between tumor type and their morphologic properties. This also supported the result that the volume features based classifier had better performance than that based on intensity. The remaining 8 were texture features, and the feature GLSZM_SizeZoneNonUniformity had the highest importance, indicating that texture features were more discriminative, and could provide complementary information to the shape-based features. This suggests that myxofibrosarcomas were more heterogeneous on T1-weighted sequences than myxomas.

The results show that the T1SI of myxofibrosarcomas are on average higher than that of myxomas, however, there was significant T1SI overlap for both lesions. We hypothesize that a more cellular tumor has higher protein content, and would therefore result in increased T1 shortening (higher T1 signal) as our results have shown. An alternative explanation is that these malignant lesions may contain small foci of hemorrhage. Myxofibrosarcomas also tended to have volumetric features that were slightly different than myxomas. Myxofibrosarcomas have been noted to have a “tail sign” and have a known propensity for spreading along the myofascial planes [18, 21]. This feature may have been detected as part of the volumetric features.

Radiomic (texture) features were the best for differentiating myxofibrosarcomas from myxomas. Quantitative analysis using a classification model based on radiomic features outperformed the classification models using volume-based and T1SI value analysis. Myxomas tend to be more paucicellular and bland (unless a cellular myxoma), and therefore have less T1 signal heterogeneity. T1-weighted signal heterogeneity of myxofibrosarcomas was greater than that for myxomas, and we speculate that the T1-weighted signal intensity heterogeneity mirrors the intrinsic histologic tumor heterogeneity seen in myxofibrosarcomas and possibly intratumoral hemorrhage.

Prior reports support our findings. Myxomas have been shown to be hypointense to skeletal muscle on T1-weighted sequences [15, 16], whereas myxofibrosarcomas have been shown to be more isointense to skeletal muscle on T1-weighted sequences [19]. However, no reports have shown that volumetric and radiomic texture features can be utilized to differentiate myxomas from myxofibrosarcomas from preoperative MRIs.

The results have potential significant clinical implications. Core biopsies are limited by the fact that these lesions are heterogeneous, and the sample cannot entirely represent the lesion’s functional and histologic properties [22]. Image-guided core biopsies targeting areas of necrosis in one of the samples may even add to the correct grade specifically in myxofibrosarcoma [23]. Radiomics offers a non-invasive, cost-effective method for assessment of a lesion’s entire tumor spatial and temporal heterogeneity [22].

We have shown that MRI image-derived radiomic features can quite accurately differentiate two extremely rare tumor types (myxomas and myxofibrosarcomas) which are challenging for pathologists and radiologists. This is particularly exciting because these tumors are so rare, most radiologists rarely encounter these tumors in daily practice, so most radiologists have limited experience in differentiating these two entities.

Radiomic feature extraction and analysis has been applied broadly to other subspecialties in radiology with successful applications in discerning molecular alterations in tumors, predicting and stratifying tumor response to therapy, and to determine patient prognosis [24‐29]. However, literature on musculoskeletal applications of radiomics is scarce. To our knowledge, this is the first study to use radiomic feature extraction and classifier prediction models for this purpose.

The study had some limitations. First, it was retrospective in nature, and subject to ascertainment bias. Myxomas and myxofibrosarcomas are rare tumors (1–4 cases per million people) and this analysis represents one of the largest analyses in the published literature of myxomas and myxofibrosarcomas. Another limitation is that pre-treatment MRIs were obtained using variable parameters and field strengths; however, this was adjusted for by using image intensity normalization. We found that all tumor MRI sequences typically included a T1-weighted sequence, which was why the analysis was restricted to T1-weighted sequences. This makes our findings more broadly applicable to clinical practices which use T1-weighted sequences in their tumor protocols. It is conceivable that there would be additional discriminative information contained in T2 or STIR and post-contrast sequences. We did not analyze the T2/STIR sequences because these were not always available (some patients had T2-weighted sequences, others had proton density-weighted sequences and others had STIR sequences), which would have left us with a much smaller sample size with limited power due to missing data. Myxofibrosarcomas often have perilesional edema and often have a tail-sign on T2-weighted sequences, which would likely add to the discriminatory ability of the model. Additional research is required to assess whether additional information can be obtained from the T2/STIR sequences to differentiate myxomas from myxofibrosarcomas. We did not analyze the contrast-enhanced T1-weighted sequences because not all patients had contrast enhanced studies and because the time from injection of contrast to imaging was not uniform across patients and we thought this would introduce more noise into the analysis and not be definitive.

In summary, we have demonstrated that radiomic features from T1-weighted sequences can provide better discriminative information in distinguishing myxoma from myxofibrosarcomas compared to T1-weighted signal intensity values and tumor volume.