| Background:Head and neck(H&N)cancer includes tumors occurred in nasopharynx,oropharynx and hypopharynx etc.In southern China,especially nasopharyngeal carcinoma(NPC)occurs with high frequency,with a peak annual incidence approaching 0.3‰.The increased uptake in inflammatory tissue(e.g.chronic nasopharyngitis,CN)on PET images hampered the diagnosis specificity of NPC.Though local-regional control has been improved due to the implementation of chemo radiotherapy in the past years,the risk of distant metastasis is still high and long term survival remains low.Conventional PET parameters and TNM staging are unable to characterize the tumor heterogeneity,which showed limited prognostic value.Exploring new effective diagnostic and prognostic features are thus crucial to make more informed decisions regarding diagnosis and treatment.Radiomics analysis,usually by quantifying intensity distribution pattern and extracting high throughput features from single modality image to characterize tumor heterogeneity,has been extensively explored as image-based biomarkers for improved diagnosis,staging,prognosis and treatment response assessment in many kinds of tumors.It is of great challenge while meaningful to construct multi-modality radiomics model that with high robustness and generalization in a multi-center setting.Thus,this study aims to investigate the robustness,diagnostic and prognostic performance of radiomics features extracted from PET/CT images in NPC and H&N cancer patients.Materials and methods:1.Investigate the impact of segmentation and discretization on 88 radiomic features in 18F-FDG and 11C-choline positron emission tomography/x-ray computed tomography(PET/CT)imaging of 40 patients with nasopharyngeal carcinoma.2.Investigate the impact of parameter settings(e.g.direction,symmetry,averaging strategy,distance,neighborhood and window size)as used for the generation of radiomics features on their robustness and disease differentiation(69 patients with NPC vs.37 patients with CN in FDG PET/CT imaging).3.Investigate the prognostic performance of radiomics features,as extracted from PET and CT components of baseline 18F-FDG PET/CT images and integrated with clinical parameters,in 128 patients with NPC.4.Propose a multi-modality multi-level fusion strategy to combine PET and CT information at the image-,matrix-and feature-levels towards improved prognosis for multi-center 296 patients with H&N cancer.Results:1.There are 8 FDG features and 11 choline features showed high robustness.2.While some features showed low robustness with respect to different strategies,they can still show high diagnostic performance.For differentiating NPC from CN,some radiomics features(e.g.SumEntropy,SZLGE,LGZE)showed higher AUC than that of conventional metrics SU-Vmax and MATV(0.91-0.93 vs.0.72-0.88,p=0.01-0.04).3.Multi-modality model showed limited prognostic performance in testing set,while outperformed single-modality model in patients with local-regional advanced stage(C-index:0.67-0.84 vs.0.64-0.77,p=0.001-0.059).4.Among 7 different training and testing partitions,the highest C-index in 5,6 and 5 partitions was achieved by image-level fusion strategies for RFS,MFS and OS predictions,respectively.Conclusion:1.The features with high robustness can be used for robust model construction in future.2.The clinical usefulness of features mainly depend on the definition of features themselves,and was partially affected by different feature extraction strategies.Some radiomics features outperformed conventional metrics.3.Radiomics features extracted from the PET and CT components of baseline PET/CT images provide complementary prognostic information compared to clinical parameters,and multi-modality model showed improved outcome prediction for NPC patients with local-regional advanced stage compared to single-modality model.4.Integrating information at image level(i.e.merging metabolic information in PET and anatomic information in CT voxel by voxel)holds potential to capture more useful characteristics. |
PDF Full Text Request