PhD Defence: PREDICTING HEAD AND NECK CANCER TREATMENT OUTCOMES WITH QUS, CT AND MRI RADIOMICS ENHANCED WITH DEEP TEXTURE ANALYSIS, A MACHINE LEARNING APPROACH

Student: Aryan Safakish

Supervisors: Dr. Greg Czarnota and Dr. Ana Pejovic-Milic

Abstract

Cancer treatment can be grueling, with outcomes varying widely. This dissertation explores utilization of machine learning (ML) models to forecast treatment outcomes for head and neck (H&N) cancers, which are close in proximity and categorize malignancies arising from more than one anatomical location. For this reason, in this investigation radiomics features were determined from pathological lymph nodes (LNs) with the hypothesis that the degree of LN involvement may reveal insights regarding future treatment outcomes, independent of the primary tumour.

Predictive models were trained using radiomics textural features determined from medical images of de novo H&N cancer patients. For one cohort of patients, features were determined from quantitative ultrasound spectroscopic (QUS) parametric maps, treatment-planning CT scans, and T1 weighted MRI scans in the first, second, and third studies respectively. Moreover, throughout the three studies, a method called Deep Texture Analysis (DTA) was developed to enhance model performances. The best performance was found with performance was demonstrated in models trained with MRI radiomics features from three total layers with a sensitivity of 92.5% and specificity of 74%, followed by a model trained with QUS radiomics features with sensitivity of 85% and specificity of 80%. Models trained with CT radiomics features demonstrated the weakest discriminative capability, with the best model showing sensitivity of 74% and specificity of 68%.

This thesis presents a unique opportunity to compare predictive models trained using radiomics features from three different imaging modalities for one cohort of patients. Contrast in ultrasound, CT, and MRI images stem from distinct physical principles–including sound wave reflection, X-ray absorption, and magnetic resonance interactions–which offer insights into underlying physiology, thus potentially highlighting relevant mechanisms that may impact treatment efficacy. Reliable models that accurately predict cancer treatment outcomes from pre-treatment imaging would provide clinicians and patients with another tool to cater to patient-centric individualized medicine. Moreover, the novelty of this thesis comes in the form of development of the DTA method over the course of three studies, and findings from the three studies which suggest that this methodology may be worth considering in future radiomics studies.