Deep Learning Model For SFT/HPC And Meningioma Classification

Background: Intracranial hemangiopericytoma/solitary fibrous tumor(SFT/HPC)is a rare type of neoplasm oriented from preicytes that located in the neurovascular unit between endothelial cells,astrocytes and neurons.It has more malignancy due to its infiltration,peritumoral edema,bleeding,or bone destruction.SFT/HPC has similar clincial and radiological characteristics as meningioma but with different clinical managements and outcomes.Recently,deep learning algorithms have been widely applied in medical imaging fields.By utilizing different deep learning methods,the machine could implement recognition,classification and segmentation of CT or MRI images.Thus,this study aims to discriminate SFT/HPC and meningioma via deep learning approaches based on routine preoperative MRI.Methods: We retrospectively enrolled 236 patients with histopathological diagnosis of SFT/HPC(n = 114)and meningioma(n = 122)from2010 to 2020 in Xiangya Hospital.We collected their clinical,radiological and pathological information.Radiological data incudes T1,T2,T1contrast(T1C)and FLAIR images.Firstly,we preprocessed and standardized the images after checking the information of patients.Radiological features including tumor boundary,bone erosion,dural sign,T1 C enhancement patterns,venous sinus invasion,cystic components,and peritumoral edema were extracted manually,and a radiological diagnostic model was established for classification.Then a deep learning pretrained model was adapted to train T1 C single sequence images for predicting tumor types.Finally,the performance of both models were evaluated by area under the receiver operating characteristic(ROC)curve(AUC)with other metrics.Feature maps was collected from the deep learning model and further analysis was implemented by dimensionality reduction,and class activation mapping(CAM)was used to explore the model attention mechanism.Results: By analysing the radiological features,our study reported that SFT/HPC and meningioma had significant different in location distribution(P=0.024),and SFT/HPC was found to have more invasion to venous sinus(P<0.001),more cystic components(P<0.001),and more heterogeneous enhancement patterns(P=0.01)with irregular tumor boundary(P=0.05).However,dural sign appeared more in meningioma.We established a classification model via these radiological features and it achieved the AUC of 0.78 in the validation set.As for Deep learning model,it achieved a high classification accuracy of 0.889 with AUC of 0.91 in the validation set.Feature maps showed distinct clustering of SFT/HPC and meningioma in the training and test cohorts,respectively.And by implementing CAM,we found the attention of deep learning model mainly focused on the tumor bulks that represented the solid texture features of both tumors for discrimination.Conclusions: In this study,We proposed a model to classify preoperative single sequence MRI of SFT/HPC and meningioma based on deep learning,which reached a high and accurate performance.For further analysis,the tumor bulks that represent the solid texture features of both tumors are essential for model discrimination.Hence,our study paves the way toward an improved clinical diagnosis and management of SFT/HPC and meningioma.There are 11 figures,4 tables,and 55 citations in this thesis.