(Sunday, 7/14/2019) 2:00 PM - 3:00 PM
Room: 304ABC
Purpose: The incorporation of cone-beam CT (CBCT) has enhanced image-guided radiation therapy. Due to severe artifacts and low soft-tissue contrast in CBCT, accurately defining the prostate volume on daily CBCT images for adaptive dose calculation is challenging. This study’s purpose is to develop a deep-learning-based approach to accurately and automatically segment the prostate from daily CBCT for potential CBCT-guided adaptive prostate cancer radiotherapy.
Methods: We propose to a new prostate segmentation strategy which applies a MRI-based pre-trained deep attentional network to CBCT-based sMRI to accurately segment the prostate on daily CBCT images. This CBCT prostate segmentation consists of three major steps: synthesis, fine-tuning, and segmentation. In the synthesis step, a cycle generative adversarial network (cycle-GAN) was used to estimate synthetic MRIs (sMRI) from CBCT image. Next, a pre-trained deep attentional fully convolution network (DAFCN), which had been previously trained on MRIs and their corresponding prostate contours, was fine-tuned by sMRI and corresponding prostate contours deformed from MRIs. In the segmentation step, a new arrival CBCT was fed into the well-trained cycle-GAN and DAFCN to obtain the segmented prostate. Our segmented prostate contours were compared with the contours manually delineated by physicians on MRI to quantify segmentation accuracy.
Results: This segmentation technique was validated with a clinical study of 20 patients’ CBCT and MR images with leave-one-out cross-validation. The Dice similarity coefficient, sensitivity and specificity indexes between manual and segmented contours were 89.8±5.5%, 86.6±5.7%, and 94.1±4.5%, which demonstrated the segmentation accuracy of the proposed method.
Conclusion: We have investigated a novel prostate segmentation framework using a CBCT-based sMRI approach to improve the prostate segmentation accuracy in CBCT images and demonstrated its feasibility and reliability. This technique warrants further development of a CBCT-guided adaptive prostate radiotherapy workflow.
Keywords
Not Applicable / None Entered.
Taxonomy
IM- Cone Beam CT: Machine learning, computer vision
Contact Email
