Open AccessDeep‐learning‐assisted algorithm for catheter reconstruction during MR‐only gynecological interstitial brachytherapy
Author(s) -
Shaaer Amani,
Paudel Moti,
Smith Mackenzie,
Tonolete Frances,
Ravi Ananth
Publication year - 2022
Publication title -
journal of applied clinical medical physics
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 0.83
H-Index - 48
ISSN - 1526-9914
DOI - 10.1002/acm2.13494
Subject(s) - brachytherapy , contouring , algorithm , computer science , magnetic resonance imaging , workflow , catheter , deep learning , nuclear medicine , thresholding , artificial intelligence , medicine , radiology , radiation therapy , image (mathematics) , computer graphics (images) , database
Magnetic resonance imaging (MRI) offers excellent soft‐tissue contrast enabling the contouring of targets and organs at risk during gynecological interstitial brachytherapy procedure. Despite its advantage, one of the main obstacles preventing a transition to an MRI‐only workflow is that implanted plastic catheters are not reliably visualized on MR images. This study aims to evaluate the feasibility of a deep‐learning‐based algorithm for semiautomatic reconstruction of interstitial catheters during an MR‐only workflow. MR images of 20 gynecological patients were used in this study. Note that 360 catheters were reconstructed using T1‐ and T2‐weighted images by five experienced brachytherapy planners. The mean of the five reconstructed paths were used for training (257 catheters), validation (15 catheters), and testing/evaluation (88 catheters). To automatically identify and localize the catheters, a two‐dimensional (2D) U‐net algorithm was used to find their approximate location in each image slice. Once localized, thresholding was applied to those regions to find the extrema, as catheters appear as bright and dark regions in T1‐ and T2‐weighted images, respectively. The localized dwell positions of the proposed algorithm were compared to the ground truth reconstruction. Reconstruction time was also evaluated. A total of 34 009 catheter dwell positions were evaluated between the algorithm and all planners to estimate the reconstruction variability. The average variation was 0.97 ± 0.66 mm. The average reconstruction time for this approach was 11 ± 1 min, compared with 46 ± 10 min for the expert planners. This study suggests that the proposed deep learning, MR‐based framework has potential to replace the conventional manual catheter reconstruction. The adoption of this approach in the brachytherapy workflow is expected to improve treatment efficiency while reducing planning time, resources, and human errors.