Multi-modality Multi-region Radiomics Prognosis And Feature Robustness Analysis

Head and neck cancer(HNC)is the seventh most prevalent malignancy worldwide,and nasopharyngeal carcinoma(NPC)is a specific type of HNC.The combination of radiotherapy and chemotherapy has improved the control rate of HNC,however 40%of patients still suffer from treatment failures.Early prognosis prediction can provide guidance for developing individualised treatment strategies for patients.Radiomics becomes an important role in the clinical management of oncology patients as a powerful tool for measuring tumour heterogeneity.However,current radiomics workflow still encounter a number of challenges including direct extraction of imaging features based on the whole tumour that ignores the regional heterogeneity within tumour;it is difficult to comprehensively characterise tumour heterogeneity based on relatively single modality information;differences in acquisition devices and scanning protocols among multi-centre data may hinder the generalization of predictive model;and the robustness of radiomics features has not been rigorously tested.To deal with these challenges,we focused on intra-tumoural sub-region identification,multimodal information fusion,multicentre dataset harmonization applied for HNC prognosis prediction,features robustness evaluation.Hence,we performed the work based the following five aspects:(1)Identification of intra-tumoral high-risk subregions of NPC based on PET/CT images and its application in prognostic analysis.The 18F-FDG PET/CT images of 128 patients with NPC were collected,and a two-step K-means clustering algorithm based on individual-and population-level separately was used to split the whole tumor into several subregions.Radiomics features were then extracted from each intratumoral subregions and the whole tumor,and their imaging biomarkers were used to construct prognostic prediction models.(2)Multimodality radiomics analysis based on 18F-FDG PET/CT and multisequence MRI images applicated to NPC prognosis.PET/CT and multisequence MRI images of 132 NPC patients were collected.Single-sequence,singlemodality,multisequence and multimodality radiomics prognostic models were sequentially constructed.Four fusion strategies based on feature-and decision-level were applied for constructing multimodality models.We compared and analyzed the performance of these models.(3)Application of Combat strategies,sub-volumes characterization and automatic segmentation in the radiomics prognostic analysis of HNC.The HECKTOR 2021 Challenge dataset was adopted,containing 325 patients from 6 centers with PET/CT images.To explore the impact of primary Combat and its 8 modified strategies on multicenter dataset harmonization;and the potential of automatic segmentation algorithms(with 2 thresholds and 1 deep-learning based methods)in radiomics prognosis pipeline compared to manual segmentation.Moreover,we further validated the complementary value of intra-tumoral subregion signature in the prognosis prediction of HNC.(4)Joint nnU-Net and radiomics to construct a one-stack segmentation and prognosis prediction system for HNC.The HECKTOR 2022 challenge dataset containing 883 HNC patients from nine centers was adopted.3D nnU-Net was used to automatically segment the primary tumor and lymph nodes of HNC.Then based on the segmentation,radiomics framework was used to develop prognostic model.This study demonstrated the potential of radiomics approaches based on automatic segmentation for HNC prognosis prediction within a large and multi-center dataset.(5)Evaluation and optimization of radiomics features stability to respiratory motion in 18F-FDG 3D PET imaging.A model simulated respiratory motion was designed to acquire PET images under different respiratory patterns.We evaluated the robustness of PET radiomics features to respiratory motion,and further optimized feature robustness in terms of feature extraction parameters and matrix aggregation strategies.