Abstract
Gastro-intestinal (Gl) toxicity plays an important role in decision-making for pelvic radiotherapy (RT), often limiting the amount of therapeutic dose that can be delivered. This thesis demonstrates how medical imaging, deep learning (DL), and normal tissue complication probability (NTCP) modelling can facilitate the introduction of RT workflows that increase therapeutic ratio through the reduction of radiation-induced Gl side effects. The combination of imaging and DL is explored in 3 key areas of the pelvic RT planning pathway: pre-treatment imaging, anatomical segmentation and treatment planning.The first study relates to the enhanced soft-tissue visualisation provided by magnetic resonance (MR) imaging. Exclusive use of MR (MR-only RT) requires generation of ‘synthetic’ CTs (sCT) for accurate dose. A phase 1 trial was established to scan 40 patients using MR in the RT treatment position. A DL network was trained to create sCT from treatment planning MR, delivering Hounsfield Unit (HU) mean error of -2.8HU, mean absolute error of 36.8HU, and rectal dosimetric accuracy within 0.1% of gold standard CT calculation. Calculation of grade 2 (G2) late rectal bleeding toxicity risk using DL sCT was accurate for both volume- and surface-based risk models, offering improvements over alternative bulk density (BD) and tissue stratification (TS) methods. Whilst both TS and DL approaches reduced errors to a clinically appropriate level, DL- based sCT offered the highest accuracy in addition to efficiency savings over TS. Clinical commissioning of MR-only prostate RT with DL-sCT is underway at Guy’s and St. Thomas’ NHS Foundation Trust, UK.
ln the second study, DL autosegmentation for RT planning CT and positron emission tomography (PET) was evaluated. A DL model was trained for auto- segmentation of bladder, rectum, sigmoid and bowel loops. Quantitative testing of segmentation and dosimetric accuracy, combined with qualitative scoring and review of editing times confirmed sufficient accuracy for implementation in RT planning. Evaluation of 18-fluorodeoxyglucose (18-FDG) tracer uptake in non- malignant organs is an active area of research, where bowel uptake may act as prognostic marker for treatment outcome. A pipeline for quantitative analysis of 18-FDG bowel uptake was developed, incorporating DL auto-segmentation and PET-guided post-processing. Accuracy was tested against manual contouring for internal and external test cohorts. High accuracy and linearity was found for near-maximum and mean tracer activity, and for tracer heterogeneity metrics. The pipeline allowed additional metrics to be evaluated compared to conventional visual scoring of bowel uptake, and extraction was performed in an automated manner with minimal user-input. Automated analysis of bowel tracer uptake is now directly applicable to large cohorts, facilitating evaluation as a prognostic tool for predicting treatment related Gl complications.
ln the third study DL-based dose prediction is evaluated as a decision-support tool. The placement of rectal spacers (RS) between the prostate and rectum reduces rectal dose from prostate RT, but costs can be prohibitive. A DL network was developed and trained to rapidly predict dose distributions for dose- escalated prostate RT. The DL model was able to predict dose-derived G2 late rectal bleeding risk and G2 late faecal incontinence risk with a high degree of accuracy. Clinical utility is demonstrated, allowing stratification of high-risk patients to RS insertion, and identification of patients close to stratification threshold who would require manual planning for decision-making.
This thesis has established applications of image-based DL and NTCP modelling that are being translated into the clinic to reduce Gl radiotoxicity.
| Date of Award | 1 Oct 2023 |
|---|---|
| Original language | English |
| Awarding Institution |
|
| Supervisor | Sally Barrington (Supervisor) & Isabel Dregely (Supervisor) |