Synopsis

Keywords: Muscle, Cancer

Dermatofibrosarcoma protuberans （DFSP） is a rare low to intermediate grade soft tissue sarcoma of skin, but the fibrosarcomatous DFSP （FS-DFSP） is a clearly malignant pathological subtype. The identification of malignant pathological subtypes by radiomics plays an important role preoperatively. The non-invasive machine learning method based on T1WI and FS-T2WI imaging is potential prognostic tool by distinguishing different levels of DFSP pathological subtypes before operation in this study to provide a new idea for the diagnosis and treatment of DFSP.

Abstract

Introduction
Dermatofibrosarcoma protuberans（DFSP）is a rare low to intermediate grade soft tissue sarcoma of skin. DFSP accounts for 0.1% of all malignant tumors and 1.8%-6% of all soft tissue sarcomas [1-3]. It is most common in adults between 20 and 50 years. There is a slight male preponderance. DFSP can occur in any part of the body, the most common site of which is the trunk, followed by the limbs and head and neck [4-6]. DFSP contains multiple histopathological subtypes, including classic, pigmented, myxoid, atrophic, sclerotic, fibrosarcomatous and granular cell DFSPs, and giant cell fibroblastoma [7,8]. The incidence of fibrosarcomatous DFSP（FS-DFSP）is relatively low, but the local recurrence and distant metastasis rates are significantly higher than other DFSP variants. There is a certain risk of death after metastasis which results in generally poor prognosis of patients, so WHO clearly defines the biological behavior of FS-DFSP as malignant [9-11]. The treatment approach to DFSP is surgical resection, while FS-DFSP usually comprehensive treatment of surgery, radiotherapy and chemotherapy and molecular targeted therapy. Accurate identification of subtypes has a certain value in determining treatment options and judging prognosis [12]. Recently, the concept of radiomics has attracted wide attention, which refers to the extraction of a large number of quantitative image feature data from images to comprehensively evaluate the various phenotypes of tumors, so as to judge tumor heterogeneity, clinical stage, histopathological grading, curative effect and so on [13,14]. At present, there is no single research and report on the imaging characteristics of DFSP at home and abroad, and most of the studies focus on soft tissue sarcoma (STS) [15]. The purpose of this study is to establish a diagnostic model for classic DFSP and FS-DFSP image information to explore the feasibility of predicting DFSP histological subtypes.
Methods
53 DFSPs were retrospectively included in this study: 9 subjects with FS-DFSP and 44 subjects with classic DFSP (mean age±standard deviation: 35.58±13.38 years). Their T1-weighted imaging (T1WI) images and fat-suppressed T2-weighted imaging (FS-T2WI) images constituted the primary dataset used to train multiple machine learning algorithms for constructing DFSP histological subtype prediction model. The subjects were randomly divided into training group and testing group. The training group was used for machine learning and the testing group was used to evaluate the performance and generalization of the training model. All patients were scanned using a 3.0T MR scanner (Ingenia CX, Philips Healthcare, the Netherlands) including T1WI (TR/TE, 600/12) and FS-T2WI (TR/TE, 4500/83). Abdominal coil was used for extremities and trunk lesions. Head and neck coil was used for a few head and neck lesions. ROIs were manually delineated by open-source software (itk-SNAP version 3.8.0) on each slice of the T1WI images and FS-T2WI images by one musculoskeletal radiologist with 5 years of professional experience, and each ROI segmentation was tested by another radiologist with 10 years of professional experience. The ROIs contained the whole tumor as much as possible. Least absolute shrinkage and selection operator (LASSO) was used to select features from preoperative imaging data. The prediction models of random forest (RF) and k nearest neighbor (KNN) classifiers were constructed on T1WI and FS-T2WI respectively by using the extracted imaging features. The prognostic performance was assessed in training cohort and testing cohort by means of AUC, sensitivity, specificity and accuracy.
Result
15 imaging features were selected by using LASSO algorithm. The AUC of RF classifier based on T1WI was 0.902; sensitivity, 91.2%; specificity, 89.2%; and accuracy, 90.4%. The AUC of RF classifier based on FS-T2WI was 0.913; sensitivity, 90.7%; specificity, 91.1%; and accuracy, 88.9%. The AUC of KNN classifier based on T1WI was 0.798; sensitivity, 88.7%; specificity, 89.7%; and accuracy, 80.1%. The AUC of KNN classifier based on FS-T2WI was 0.815; sensitivity, 83.5%; specificity, 86.3%; and accuracy, 82.1%.
Discussion
Machine learning based on radiomics has certain significance for the identification of histological subtypes of DFSP. In this study, four groups of data show good prediction performance, and compared with KNN classifier, RF appears better performance. RF has the advantages of fast learning speed, random selection of samples and features, efficient processing of big data, and high classification accuracy, so many studies on STS image groups tend to use RF as a feature selection algorithm or classifier. One deficiency of this study is that the sample size of FS-DFSP is too small, which may lead to biased results, and more patient data will be introduced later. Another one is that only one feature selection algorithm is selected, which may result in differences in the selection. Other appropriate algorithms will be added according to the previous literature to make the results more perfect.
Conclusion
This study puts forward for the first time the application of radiomics to distinguish the histopathological subtypes of DFSP, which provides a new and effective assistant diagnostic method for recognizing DFSP variants. The non-invasive machine learning method based on T1WI and FS-T2WI imaging is potential prognostic tool by distinguishing different levels of DFSP pathological subtypes before operation to improve the treatment strategy.

Acknowledgements

No acknowledgement found.