Nasopharyngeal Carcinoma Subtyping Based On Radiomics Of MRI

Nasopharyngeal carcinoma(NPC)is the most common cancer originating in the nasopharynx.It is more common in certain regions of East Asia and Africa,especially in south China.The American Joint Committee on Cancer(AJCC)/Union for International Cancer Control(UICC)TNM staging system is the most widely used prognostic tool for predicting survival outcomes for patients with NPC.Intriguingly,the outcomes of patients with the same stage of NPC often varies.Recently,several clinical and genetic factors have been identified to be potentially applicable as independent prognostic factors for NPC.However,their use is limited in clinical practice.With the development of radiomics,Magnetic Resonance Imaging(MRI)has been considered as one of the powerful non-invasive prognostic tools that could help clinicians stratify NPC patients for tailored treatment.This thesis aims to subtype NPC patients based on radiomics features of MRI.To achieve this goal,the thesis proposes a representative feature model to extract representative features from the radiomics features of MRI.Guided by the survival information of patients,the representative feature model selects a representative slice from the MRI slices with tumor,and takes the radiomics feature of the selected slice as the representative features of the patient.After that,feature selection and risk prediction are performed to construct patients subtyping model using univariate COX model and random survival forests.The thesis also proposes a low-rank sparse metric learning method for multi-view subspace clustering.The method can discover the underlying common pattern among the multi-view subspace representation while maintaining the informative structure of the original subspaces.Experiments on synthetic datasets and real-world datasets demonstrate the efficiency of the proposed method.It is then applied to the multi-modality MRI dataset to obtain efficient patients subtyping.Multiple experimental results show that features in the frequency domain of MRI demonstrate strong prognostic ability in survival prediction and patients subtyping.Risk evaluation model based on these features can subtype patients well into low-risk group and high-risk group.The main contributions of this thesis are listed as follows: a)This thesis proposes a representative feature model guided by survival information,which provides a new method for radiomics feature extraction;b)This thesis proposes a low-rank sparse metric learning method for multi-view subspace clustering,which provides a new solution for common structure discovery in multi-modality data;c)This thesis finds that radiomics features in frequency domain of MRI have a strong prognostic ability in survival prediction and patients subtyping,which is meaningful and inspiring for researches aiming to discover the NPC prognostic factors from radiomcis of MRI.