Abstract
Mandibular osteoradionecrosis (ORN) in patients with head and neck cancer undergoing radiotherapy (RT) is a rare radiation-induced toxicity but can highly compromise patients’ quality of life and result in costly clinical interventions.In addition to clinical and demographic risk factors, radiation dose plays an important role in the development of mandibular ORN. Existing ORN prediction models use the dosimetric information extracted from the dose-volume histogram (DVH) of the mandible.
In a DVH, the clinical radiation dose distribution map of the mandible volume is reduced to a 2D representation that omits any spatial dose information. Because the anatomy and the radiosensitivity varies across the mandible, this spatial dose information is clinically relevant. In this thesis I hypothesise that the incidence of mandibular ORN can be predicted based on the clinical radiation dose distribution maps as the dosimetric factor combined with the clinical and demographic factors.
A class-balanced cohort of up to 92 ORN cases and 92 matched controls treated with intensity-modulated radiotherapy (IMRT) between 2011 and 2022 were retrospectively selected from the clinical database. The clinical and demographic data was retrieved from the clinical notes and the DVH and RT DICOM files were exported from the clinical treatment planning system. To facilitate subsequent ORN prediction model development, a pipeline was developed that involves a number of image pre-processing steps.
First, the computed tomography (CT), RT dose and mandible structure volumes were registered to a common space to compensate for inter-patient positioning variations. Then the RT dose map was masked by the mandible structure, thus resulting in the mandible dose map.
The first part of this thesis focuses on exploring machine learning (ML) and deep learning (DL) classification models for the prediction of ORN incidence. A first experiment was performed to initiate the transition from traditional ORN risk factor analysis to ML-based case-by-case ORN incidence prediction. The performance of different ML methods was compared on the task of predicting ORN incidence based on DVH metrics and clinical and demographic data. Although no statistically significant difference was observed between models, the artificial neural network (ANN) model showed the highest prediction accuracy (71%). This study was followed up with a DL-based approach that used the mandible radiation dose map as the input into a 3D deep convolutional neural network (CNN) for binary classification (ORN vs. non-ORN). The predictive performances (AUROC) of three different CNN architectures were compared, including a DenseNet121 (0.64), a DenseNet40 (0.69) and a ShuffleNet (0.65).
The dose map-based deep CNN model prediction performance results were then compared to a DVH-based Random Forest (RF) model (0.61 AUROC). This DL-based ORN prediction approach was expanded to include other non-dosimetric risk factors (clinical variables). This was done following early and late multimodality fusion strategies, which resulted in similar prediction performances (0.68 and 0.70 AUROC, respectively) to the single-modality DL model (0.69), but had a statistically significantly higher performance than a RF model trained on clinical variables only (0.60).
The second part of this thesis focuses on exploring the interpretability of the DL-based ORN prediction model using the 3D Grad-CAM pixel-attribution method and quantitatively analysing its results to draw clinically relevant conclusions. The results obtained were in alignment to existing clinical knowledge derived from the more traditional statistical approaches, which represents an important step towards gaining trust for the clinical implementation of a DL-based ORN prediction model.
Finally, this thesis includes a description of the PREDMORN multi-institutional study, which I designed and developed to obtain the largest and most diverse mandibular ORN dataset worldwide that will allow for the development of robust and generalisable ORN prediction models and further subsequent studies.
Overall, I expect that the work included in this thesis will represent a significant step towards a more individualised treatment of head and neck cancer that will potentially result in an incidence reduction or better prognosis of mandibular ORN.
| Date of Award | 1 Jul 2023 |
|---|---|
| Original language | English |
| Awarding Institution |
|
| Supervisor | Teresa Guerrero-Urbano (Supervisor) & Andrew King (Supervisor) |