Towards Early Treatment Response Prediction using Longitudinal Distortion-free Diffusion-weighted Magnetic Resonance Imaging
- Author(s): Gao, Yu
- Advisor(s): Hu, Peng;
- Low, Daniel A.
- et al.
Radiotherapy is an effective tool to treat tumors by delivering high-energy photons or charged particles to destroy malignant cancer cells. In the current fractionated radiotherapy treatment workflow, the same radiation plan is delivered to the patient in every fraction during the entire course of treatment, whereas tumor changes and patient response are not taken into consideration. Early patient response prediction is appealing as it offers a window of personalized treatment adaptation for potentially improved treatment outcome.
Diffusion-weighted imaging (DWI) has been shown to be a promising non-invasive biomarker for treatment response assessment. However, the conventional diffusion-weighted single-shot echo-planar-imaging (DW-ssEPI) has strong spatial distortion, which is unacceptable for radiotherapy applications as high geometric accuracy is required for accurate tumor delineation. To achieve the ultimate goal of response-based adaptive radiotherapy, this dissertation sought to develop distortion-free diffusion sequences, and evaluate the possibility of early treatment response assessment using longitudinal DWI on sarcoma patient.
In Chapter 3, a diffusion-prepared turbo spin-echo (DP-TSE) sequence was programmed and compared with the DW-ssEPI sequence on a 0.35T MRI-guided radiotherapy system. The DW-ssEPI failed the spatial integrity test due to severe distortion and low signal intensity, whereas the DP-TSE passed the test successfully. The diffusion phantom study showed that noise correction must be performed for the DW-ssEPI sequence to avoid apparent diffusion coefficient (ADC) quantification errors, whereas DP-TSE had desirable ADC accuracy and ADC reproducibility. Good geometric fidelity and ADC quantifications were obtained in the patient study involving two glioblastoma (GBM) patients and six sarcoma patients.
Shot-to-shot k-space magnitude inconsistency is a common problem in multi-shot diffusion-prepared imaging. In Chapter 4, a magnitude stabilizer strategy was proposed to convert the malignant magnitude inconsistency to phase inconsistency, which is easier to resolve. We demonstrated that the proposed diffusion-prepared magnitude-stabilized balanced steady-state free precession sequence (DP-MS-bSSFP, abbreviated as DP-MS) had satisfactory ADC accuracy, and significantly improved geometric accuracy on both phantom and volunteers compared to the conventional DW-ssEPI approach.
To meet the requirement of high spatial integrity and high spatial resolution for treatment planning and adaptation, the 2D DP-MS sequence was extended to 3D in Chapter 5. A locally low-rank constrained reconstruction was applied to correct the k-space inconsistency. Similar as Chapter 4, the 3D DP-MS sequence was verified on the diffusion phantom and five healthy volunteers for geometric fidelity and ADC accuracy. Overall, the 3D DP-MS sequence had submillimeter geometric accuracy and satisfactory ADC accuracy on both phantom and volunteers.
In the second half of this dissertation, we focused on early treatment response prediction using longitudinal DWI on sarcoma patients treated with hypofractionated radiation therapy. Diffusion images were acquired using DW-ssEPI three times through the treatment. In Chapter 6, a radiomics-based approach was used to prediction treatment effect score, which is obtained from the post-surgery pathology. Logistic regression and support vector machine (SVM) models were constructed to predict the treatment effect using radiomics features selected by univariate analysis and sequential forward selection. The SVM model outperformed logistic regression and had an area under the receiver operating characteristic curve (AUC) of 0.91�0.05.
To overcome the small sample size problem in Chapter 6 and further improve the prediction, deep learning-based data augmentation and prediction were implemented in Chapter 7. An ACGAN network was trained to augment the data based on training patients, and a prediction model based on the VGG-19 was trained using the synthesized data and validated on the training patient dataset. This trained model was then tested on the hold-out test patients. Overall, the training, validation, and test accuracies was 94.3%, 90.1%, and 87.7%, indicating good performance of the data augmentation and response prediction models.