Open Access Dissertation
Date of Award
Presented in this work is an investigation of the application of artificially intelligent algorithms, namely deep learning, to generate segmentations for the application in functional avoidance radiotherapy treatment planning. Specific applications of deep learning for functional avoidance include generating hippocampus segmentations from computed tomography (CT) images and generating synthetic pulmonary perfusion images from four-dimensional CT (4DCT).A single institution dataset of 390 patients treated with Gamma Knife stereotactic radiosurgery was created. From these patients, the hippocampus was manually segmented on the high-resolution MR image and used for the development of the data processing methodology and model testing. It was determined that an attention-gated 3D residual network performed the best, with 80.2% of contours meeting the clinical trial acceptability criteria. After having determined the highest performing model architecture, the model was tested on data from the RTOG-0933 Phase II multi-institutional clinical trial for hippocampal avoidance whole brain radiotherapy. From the RTOG-0933 data, an institutional observer (IO) generated contours to compare the deep learning style and the style of the physicians participating in the phase II trial. The deep learning model performance was compared with contour comparison and radiotherapy treatment planning. Results showed that the deep learning contours generated plans comparable to the IO style, but differed significantly from the phase II contours, indicating further investigation is required before this technology can be apply clinically. Additionally, motivated by the observed deviation in contouring styles of the trial’s participating treating physicians, the utility of applying deep learning as a first-pass quality assurance measure was investigated. To simulate a central review, the IO contours were compared to the treating physician contours in attempt to identify unacceptable deviations. The deep learning model was found to have an AUC of 0.80 for left, 0.79 for right hippocampus, thus indicating the potential applications of deep learning as a first-pass quality assurance tool. The methods developed during the hippocampal segmentation task were then translated to the generation of synthetic pulmonary perfusion imaging for use in functional lung avoidance radiotherapy. A clinical data set of 58 pre- and post-radiotherapy SPECT perfusion studies (32 patients) with contemporaneous 4DCT studies were collected. From the data set, 50 studies were used to train a 3D-residual network, with a five-fold validation used to select the highest performing model instances (N=5). The highest performing instances were tested on a 5 patient (8 study) hold-out test set. From these predictions, 50th percentile contours of well-perfused lung were generated and compared to contours from the clinical SPECT perfusion images. On the test set the Spearman correlation coefficient was strong (0.70, IQR: 0.61-0.76) and the functional avoidance contours agreed well Dice of 0.803 (IQR: 0.750-0.810), average surface distance of 5.92 mm (IQR: 5.68-7.55) mm. This study indicates the potential applications of deep learning for the generation of synthetic pulmonary perfusion images but requires an expanded dataset for additional model testing.
Porter, Evan, "Segmentation Of Intracranial Structures From Noncontrast Ct Images With Deep Learning" (2022). Wayne State University Dissertations. 3604.