Multimodal Imaging Based Prognosis Research For Lung Metastasis Risks Of Soft Tissue Sarcomas

Soft tissue sarcomas are a group of uncommon malignant tumors made up by soft tissue coming from the fat,muscles,nerves,tendons,blood vessels and lymphatic,which may occur in various parts of the body.They are quite uncommon.Incidence of this disease is 1/50000,and most of the newly discovered cases got deteriorated to a middle and high level.In general,dfferent forms of therapies could cotrol the recurrence of soft tissue sarcomas to some extent,but most patients are easy to have distant metastasis after treatment.Lung is the main distant metastatic place,and lung metastasis is an important factor of survival threats.The prognosis for lung metastasis of soft tissue sarcomas is still poor up to now,most patients will die of this in two or three years.Therefore,an early evaluation for lung metastasis risks in the course of soft tissue sarcomas management is of great significance,and may help doctors to adjust treatment plans,improving the overall survival rate eventually.Radiomic is a new research field of medical imaging,using the advanced image analysis methods,extracted a lot of features with high fidelity from the medical images to be analyzed and to make a comprehensive evaluation for various phenotypes of tumor.Radiomics involves in machine learning,data mining,and statistical learning methods,and the main research contents include medical image acquisition and reconstruction,image segmentation,feature extraction and feature selection,and sample analysis etc.Researches showed that radiomics could reveal the predictive information of tumor's development tendency,making help in the prognosis of tumor.This paper aims to make an early assessment for the risks of lung metastasis of soft tissue sarcomas based on radiomics.51 patient cases with soft tissue sarcomas form TCIA website soft tissue sarcomas database were retrospectively analyzed.Firstly,4 kinds of 3D morphological features were extracted from each patient's CT(computed tomograpy)and MRI(magnetic resonance imaging)T1 and T2 FS sequence images.After the image pretreatment process including wavelet denoising,isotropic resampling and grayscale quantification,42 kinds of 3D texture features were extracted.Then,we classified all the image features to 7 sets according to the combination of single modality,two modalities and three modalities to make preliminary feature selection using filtering feature selection based on Spearman rank correlation coefficient and maximum information coefficient,the features of each set were reduced to just 25 features.Then the seven image feature sets were underwent further feature selection using multiple feature selection methods,and we compared a variety of resampling methods for their classification performance,determined an appropriate method.Finally,selecting the appropriate sample resampling method,the classification performance of the seven feature sets after preliminary screening combined with 11 kinds of further feature selection methods and 13 kinds of classifiers were accessed,and we analysed how the classification accuracy varied with the influnce of feature numbers.The results showed that the KNN classifier together with the Trace Ratio Criterion feature selection method obtained the best performance,with the accuracy is 87.6% and the AUC value is 90.8%.We also found that multimodal imaging dataset had better prognosis value than single imaging dataset.However,the results above were just limited to a small sample set,lack of independent validation sample sets.And there were no innovations in the feature extraction,feature seletion and classification methods,the comparation of methods are not thorough,either.All of these aspects need more efforts in the future research.