2273 - AI Contouring in HDR Prostate Brachytherapy Using a General Foundation Model

J. Beaudry¹, I. Youssef², A. L. Damato¹, J. Jiang¹, H. Veeraraghavan¹, M. A. Kollmeier², and S. Elguindi¹; ¹Department of Medical Physics, Memorial Sloan Kettering Cancer Center, New York, NY, ²Department of Radiation Oncology, Memorial Sloan Kettering Cancer Center, New York, NY

Purpose/Objective(s): To: (i) assess the cross-modality capabilities of a transformer-based deep learning network; and (ii) evaluate the model for auto-segmentation of the prostate, urethra, rectum, bladder, and seminal vesicles in preimplant high-dose rate (HDR) brachytherapy prostate transrectal ultrasound (TRUS) images. Materials/

Methods: 3D TRUS images were acquired prior to needle insertion through axial slice acquisition. A Foley catheter was inserted to assist with visualization of the urethra and bladder. An in-house developed transformer-based deep learning model was initially pre-trained using an unsupervised learning approach on 10,413 CT images. A dataset of 65 prostate TRUS images was divided into 54 training, 6 validation, and 5 test images was used to fine tune and assess the usability of this pre-training across modalities for the purpose of auto-segmentation of common HDR prostate contours in brachytherapy. AI generated contours were qualitatively assessed and were compared against expert-delineated contours using the Dice similarity coefficient.

Results: The transformer model achieved Dice coefficients (mean ± SD) of 0.88 ± 0.03 for the prostate, 0.77 ± 0.03 for the rectum, 0.72 ± 0.11 for the urethra, 0.53 ± 0.24 for the bladder, and 0.61 ± 0.11 for the seminal vesicles, respectively, on the validation and test sets combined. Visual inspection revealed better similarity mid-gland, with more deviations notes at the apex and base of the prostate.

Conclusion: A transformer-based model pre-trained using one modality can be successfully trained cross-modality with TRUS images. Initial results suggest that preimplant automated HDR contours may be clinically useful, reducing contouring time and assisting with needle placement in an intraoperative setting. Abstract 2273 – Table 1

	Dice Similarity Coefficient
Contours	Median	Mean	St Dev.	Min	Max
Bladder	0.60	0.52	0.27	0.00	0.86
Prostate	0.88	0.88	0.03	0.82	0.93
Rectum	0.77	0.77	0.04	0.71	0.85
Seminal Vesicles	0.59	0.61	0.11	0.42	0.80
Urethra	0.75	0.72	0.11	0.46	0.84