Clinical use-based evaluation of SmartFuse: an AI-powered registration software by TheraPanacea
Purpose: Image registration plays a central role in treatment planning and delivery. The goal of this study is to evaluate the
accuracy of a rigid and a spline based free-form deformation registration tool offered by TheraPanacea.
Methods and materials: Validation of registration software systems
remains a challenge due to the lack of standard mathematical
formalism to perform real-world evaluations where noise, distortion,
and complex anatomical variations can occur. In [1], different
evaluation methods and criteria were suggested depending on the
clinical use of the registration. Following these guidelines, in this study
we evaluate the usability of SmartFuse for each of the major fusion
clinical uses cases in the radiotherapy workflow related to replanning,
contouring assistance, electron density transfer and 4D-CT: a)
replanning in case of MRgRT(MR-MR), b) CT-based replanning (CT-CT),
c) planning on MR with CT as a secondary image (CT-MR), d) planning
Fig. 1- Example of contour propagation after deformable
on CT with MR as a secondary image (MR-CT), and e) planning on 4D- registration from planning (1b/2b) to the daily MR images
CT. For each clinical case, a mix of abdominal, pelvic, H&N, brain, and (1a/2a) for prostate (top) and abdominal cases for MRgRT.
Organs like the PTV (blue), kidney (yellow) and aorta (orange)
thoracic patient images (n=36 per case) were evaluated. were used to assess the quality of the propagated contours.
DSCs were computed between the registered contours of the source
and the target images for both types of fusion. In addition, external
medical experts visually evaluated the quality of fused images and
propagated contours through a scoring system with a) contour is
acceptable for clinical use without any modification/fusion results
exceed expectation, b) contour would be acceptable for clinical use
after minor modifications/fusion results meet expectation, c) contour
requires major modifications/fusion results do not meet expectation.
Results: The average mean DSC for the cases where contour
Fig. 2- Qualitative assessment of rigid/deformable registration propagation is needed for replanning purposes was 0.74 and 0.89 for
per clinical case, with the deformable outperforming the rigid
registration for all cases. For the case of 4D-CT, 91.7% of the rigid and deformable registration, respectively. Example of cases of
fusion results exceeded/met the evaluator’s expectations. MR-Linac for both the pelvic (T2) and abdominal areas (TrueFisp) are
shown in Fig.1. For the cases where no propagation of the contours is
needed, following the guidelines outlined in [1], Fig.2 presents the
qualitative evaluation results. In Fig.3, an example of a fusion between
MR and CT brain images after rigid registration is shown. Qualitative
evaluation of the fusion results led to a global acceptance of 90% for
the rigid and 98% for the deformable registration, respectively.
Conclusion: The accuracy and performance of the registration
algorithm for both deformable and rigid registration suggests its utility
for the different clinical purposes outlined in [1] including the
integration into an adaptive radiotherapy protocol of MRgRT.
Fig. 3- Qualitative assessment of rigid registration between an References:
MR (4) and a CT (5) image for a brain case. The overlap using the
checker-board function are shown in the axial (1), coronal (2) [1] Brock KK, Mutic S, McNutt TR, Li H, Kessler ML. Use of image registration and fusion
and sagittal (3) views. Landmarks like the nose, optical nerves algorithms and techniques in radiotherapy: report of the AAPM Radiation Therapy
(green and orange) and brainstem (blue) shown good overlap of Committee Task Group No. 132. Med Phys. 2017; 44:e43–e76
the images, especially when superimposed.
