(LAM, Sai Kit, Department of Health Technology and Informatics, The Hong Kong Polytechnic University)
Presently, Intensity-modulated radiotherapy (IMRT) is a standard-of-care remedy for advanced nasopharyngeal carcinoma (NPC) patients. Notably, the success of treatment relies on an assumption that the patient anatomy remains throughout the entire IMRT course. In response to treatment perturbations, however, tumors and surrounding healthy organs may exhibit significant morphometric volume and/or geometric alterations, which may jointly alter patient anatomy and jeopardize the efficacy of the original treatment plan. Confronted with this, Adaptive Radiotherapy (ART) has been introduced to compensate for these patient-specific variations. Numerous criteria as ART triggers has been introduced. Nevertheless, most of these factors require close monitoring throughout the IMRT course for each patient, and are deficient in capturing inter-patient disparity in intrinsic biologic response of tissue upon receiving treatment perturbation. Therefore, effective pre-treatment prediction of ART eligibility is greatly demanding.
In this study, various machine learning techniques was applied to investigate capability of a variety of prediction models, which were developed by using different types of “-omics” features extracted from a variety of organ structures, for pre-treatment prediction of ART demand in NPC patients, with an ultimate objective to facilitate ART clinical implementation in the future.
First, 124 and 58 NPC patients from Queen Elizabeth Hospital (QEH) and Queen Mary Hospital (QMH), respectively, were retrospectively analyzed. Radiomic features extracted from neck nodal lesions of Computed Tomography (CT) images, patient’s clinical data, and combined types of features were used for developing R, C, and RC models, respectively, for predicting ART event triggered by ill-fitted thermoplastic mask (IfTM) due to significant nodal volume shrinkage. Results showed that the R model performed significantly better than the C model in the external QMH testing cohort (p < 0.0001), while demonstrating no significant difference compared to the RC model (p = 0.5773). Second, pre-treatment contrast-enhanced T1-weighted (CET1-w), T2-w magnetic resonance (MR) images of seventy NPC patients from QEH were processed for extraction of radiomic features from Gross-Tumor-Volume of primary NPC tumor (GTVnp), for developing CET1-w, T2-w, and joint T1-T2 models for ART eligibility prediction. Models were developed using the least absolute shrinkage and selection operator (LASSO) logistic regression. Results indicated promising predictability of MR-based tumoral radiomics, with AUCs ranging from 0.895 – 0.984 in the training set and 0.750 – 0.930 in the testing set. In general, the joint T1-T2 model outperformed either CET1-w or T2-w model alone. Third, pre-treatment CECT and MR images, radiotherapy dose and contour data of one-hundred and thirty-five NPC patients treated at QEH were retrospectively analyzed for extraction of multi-omics features, namely Radiomics (R), Morphology (M), Dosiomics (D), and Contouromics (C), from a total of eight organ structures. Four single-omics models (R, M, D, C) and four multi-omics models (RD, RC, RM, RMDC) were developed under 10-fold cross validation and evaluated on hold-out test set. Results demonstrated that the R model significantly outperformed all other three single-omics models (all p-value<0.0001), achieving an average AUC of 0.942 (95%CI: 0.938-0.946) and 0.918 (95%CI: 0.903-0.933) in training and hold-out test set, respectively. Besides, the R model demonstrated no significant difference as compared to all studied multi-omics models in the hold-out test sets. Intriguingly, Radiomic features accounted for the majority of the final selected features, ranging from 64% to 94%, in all the studied multi-omics models.
In conclusion, a series of studies in this thesis demonstrated that CT-based neck nodal radiomics was capable of predicting If TM-triggered ART events in NPC patients undergoing RT, showing higher predictability over traditional clinical predictors. MRI-based tumoral radiomics was shown promising in pre-treatment identification of ART eligibility in NPC patients. In particular, the joint T1-T2 model outperformed both T1-w and T2-w models. Multi-organ multi-omics analyses revealed that the Radiomic model played a dominant role for ART eligibility in NPC patients. The overall findings may provide valuable insights for future study into developing an effective screening strategy for ART eligibility in NPC patients in the long run, ultimately alleviating the workload of clinical practitioners, streamlining ART procedural efficiency in clinics, and achieving personalized RT for NPC patients in the future.