Radiomics, machine learning, deep learning in oncologic imaging : applications in nasopharyngeal carcinoma and esophageal squamous cell carcinoma

Abstract

Nasopharyngeal carcinoma (NPC) and esophageal squamous cell carcinoma (ESCC) have for years been the leading causes of death in Hong Kong. In current clinical practice, non-invasive imaging techniques are commonly used for these cancer types. Recently, the application of radiomics, machine learning, and deep learning has provided a novel scope for advanced imaging analysis for cancer patients. However, there is a research gap regarding their applications in NPC and ESCC. In this study, radiomics, machine learning, and deep learning -based investigations of the related tasks including diagnosis, therapy response evaluation, prognosis prediction, and biological characterization were conducted with the aim to improve clinical decision-making on treatment strategy and risk stratification on prognosis. The serial studies in this thesis aimed to apply recent advances of the machine learning techniques for prognosis prediction of NPC and treatment response evaluation, prognosis prediction and lymph node status determination of ESCC. First, 166 NPC patients were recruited from the HKU PET-CT center. A cross-combination of ten re-sampling methods with four machine learning classification models were applied for survival prediction. The model performance could be improved by the application of resampling techniques in clinical imbalanced data. Second, two hundred and thirty-one ESCC patients receiving neoadjuvant chemoradiation (nCRT) followed by surgery at two institutions in Southeast China, were consecutively recruited. The peri-tumoral radiomic features showed additional value in estimating the pathologic complete response following nCRT. Furthermore, a subsequent study showed the model using features extracted from pre-trained deep learning models had a better performance than the handcrafted radiomics model. Third, exploratory radiogenomics analysis demonstrated the potential importance of the tumoral microenvironment in the nCRT assessment. Our results also showed the genomic information was useful for radiomic feature selection, which could improve the radiomics pipeline for prognosis prediction in ESCC patients. Lastly, we analyzed the prevalence of metastatic lymph nodes for each region in ESCC patients receiving upfront surgery, which highlighted that machine learning models provide superior information about lymph node metastasis status than conventional size-based measurements. This study showed the radiological features extracted using radiomics or deep learning models were informative for disease evaluation and prognosis prediction. Re-sampling techniques were effective in improving prognostic performance on imbalanced datasets. Radiogenomics analysis could help to interpret the predictive power of radiomics features in terms of associated pathophysiology, which introduced a novel way of radiological feature selection. Recent advances in machine learning could provide abundant information about tumor heterogeneity, which could significantly improve cancer patient risk stratification in different aspects and benefit precision medicine.