1026 - Fast Tracking MR Only Adaptive Radiotherapy; Automatic Segmentation on Planning, Daily MRI and Synthetic CT

A. Amjad¹, R. Geoffrey¹, D. Genz¹, S. Horache², L. Colombo², X. Chen¹, and E. S. Paulson¹; ¹Department of Radiation Oncology, Medical College of Wisconsin, Milwaukee, WI, ²TheraPanacea, Paris, France

Purpose/Objective(s): CT based deep learning automatic segmentation (DLAS) has proven to improve efficiency of clinical workflows by reducing contouring time. However, MR based clinical workflows, including MR-only adaptive radiotherapy (MROART), remain a challenge. We investigate here the feasibility of using non-sequence specific deep learning algorithms to generate sCT from planning or daily MR images combined with auto-segmentation on both MRI and sCT for prostate MR-guided online adaptive radiotherapy (MRgOART). Materials/

Methods: A unique 3-level architecture comprised of recursive, convolutional, and multi-cephalic neural networks was used to train the MR and sCT segmentation models. A cycle generative adversarial network was used to generate sCTs from MRI. For testing, a planning 3D T2 from a 3T MRI, a daily 3D T2 from a 1.5T MR-linac, and a sCT generated from the daily MRI, were acquired for 2 clinical patients, and 2 male volunteers. The segmentation on MRI and sCT was reviewed on slice-by-slice basis to evaluate clinical accuracy. Additionally, Dice similarity coefficient (DSC), and mean distance to agreement (MDA) were calculated to evaluate the 3D accuracy of the MRI segmentation.
Results: The percent of clinically acceptable segmentations, from slice-based contour analysis, is shown in the table below. Fast segmentation (40 seconds) was successful for the sCT, planning and daily MRIs. Average DSC, MDA/mm were (0.8,0.8,0.8,0.8,0.8), (7,2,6,1.3,1.4) and (0.98,0.87,0.8,0.5,0.7), (1,1.4,1.5,7,2) for bladder, prostate, rectum, seminal vesicles, and penile bulb for the daily and plan MRI, respectively. The following observations were made during evaluations; 1) sCT generation struggled with edge-of-scan artifacts in MRI, 2) accuracy on MRI was dependent on urine in bladder (in the absence of urine the tool was unable to segment the bladder), 3) segmentation was successful in capturing anatomical changes, such as large air pockets in the rectum, 4) this approach eliminates intrafraction differences due to motion, as often in clinical workflows the sCT needs to be generated from a sequence different than the 3D T2.
Conclusion: The combined use of 3D T2 MRI and sCT in a segmentation module can simplify, automate, and expedite plan and re-plan adaptation, especially for pelvic anatomy, where organ filling changes can severely influence efficiency in MRgOART. For our ongoing work, we are evaluating the dosimetric viability of the proposed workflow.

	Bladder			Prostate			Rectum			SeminalVes			Penile Bulb
	Plan	Daily	sCT	Plan	Daily	sCT	Plan	Daily	sCT	Plan	Daily	sCT	Plan	Daily	sCT
1		18%	89%	91%	77%	95%	94%	95%	97%		63%	89%	62%	100%	100%
2	95%	93%%	94%	94%	87%	93%	97%	99%	99%	86%	86%	86%	93%	100%	94%
3	100%	40%	78%	83%	84%	91%	97%	98%	86%	32%	93%	72%	100%	100%	100%
4	72%	0%	55%	84%	69%	93%	96%	73%	89%	86%	94%	80%	88%	100%	100%