The evaluation of a hybrid biomechanical deformable registration method on a multistage physical phantom with reproducible deformation

To correct for potential erroneous IM-DIR DVF inside regions of interest, a hybrid method of incorporating a physical model into pure image DIR has been developed. The specific implementation consists of steps like volume mesh generation, boundary condition assignment and DVF interpolation (depicted in Fig. 1).

Generally, BM-DIR methods are more complicate to implement than pure IM-DIR. To investigate the proper situations where BM-DIR is more desirable, a soft-tissue equivalent phantom with reproducible and controllable deformation was developed [43]. Briefly, our phantom is based on a Ultrasound Prostate Training Phantom (CIRS, Norfolk, VA) -- a multi-modality phantom providing anatomical imaging contrast under CT, MRI, ultrasound and elastography similar to that of patient. The prostate along with structures simulating the rectal wall, seminal vesicles, perineal membrane and urethra is contained within clear acrylic container. In our deformable phantom, the prostate gland is programmatically deformed and displaced using an inflatable balloon controlled by a programmable syringe pusher as shown in Fig. 2(a). Ten different stages of deformation were created by controlling the syringe position and imaged under CT (Philips Brilliance Big Bore) and MR (Philips Ingenia 3.0 T Spine Echo, TR 2000 ms, TE 40 ms). The original resolution is 0.03, 0.03, 0.2 cm for MR and 0.05, 0.05, 0.1 cm for CT. Both the CT and MR images are resampled to the same resolution 0.05, 0.05, 0.2 cm before DIR to facilitate direct DVF comparison without interpolation uncertainty. Figure 2 shows the phantom setup and the un-deformed MR images (phase 0). The un-deformed linear dimensions are 4.42, 3.85 and 5.07 cm in anterior–posterior, lateral and craniocaudal directions, respectively. The tube structure simulating urethra shows sharp contrast on MR images and less contrast on CT images (Fig. 2 and Fig. 3). The MRI and CT of the most extreme-deformed stage (phase 10) is shown in the top row in Fig. 3. The fusion of CT and MR images on phase 3, 7, and 10 are shown in bottom row from left to right, where highly reproducible deformation states can be observed between the two imaging modalities. The contours of the prostate phantom and urethra structure were delineated on MR phase 0, and propagated to other phases by using ADMIRE, then reviewed and edited if necessary. The third set of images (Masked MR), was generated by overriding the inner voxels with the average intensity inside prostate contour (the green contour in Fig. 2b and magenta contour Fig. 3a). The purpose of generating the Masked MR is to emulate intensity-homogeneous soft-tissue organs as manifested on CT.

The IM-DIR and BM-DIR were performed between Phase 0 as reference and Phase i (i = 1 to 10) for CT, MR and masked MR with prostate contour as constrain, generating six types of DVFs for each phase. Since MR images had much better contrast inside prostate, the output DVFs were used to warp the MR Phase i images to fuse with MR phase 0 for a qualitative visual inspection. The tubular structure inside this prostate phantom --simulating the urethra-- has excellent contrast on MR image and can serve as surrogate for internal deformation accuracy. The urethra tube was deformed from phase i to phase 0 via different DVFs. The DSC, Hausdorff and mean surface distance as defined in [44] were calculated between the propagated and human-delineated urethra tube to evaluate the accuracy of internal DVF. The Jacobian determinant of DVFs from the MR and masked MR images were calculated and compared between IM-DIR and BM-DIR. In addition, the voxel-wise discrepancy between DVF from IM-DIR and BM-DIR inside the prostate phantom was evaluated on all ten phases. Target Registration Discrepancy (TRD) was calculated for every voxel inside the prostate phantom as the vector distance between the warped points generated using the IM-DIR and BM-DIR. The histogram of TRD was compared between different image types and deformation phases.

Two pairs of MR images (Philips Ingenia 3.0 T, T2W, TR 5187 ms, TE 80 ms) of Head Neck cancer patients have been acquired on two patients before and after radiotherapy under approval from our institutional review board, with the resolution of 0.045, 0.045, 0.306 cm in anterior-posterior, lateral and inferior-posterior direction respectively. The two patients were selected for the parotid shrinkage after radiotherapy and the availability of high-resolution MR with clearly defined internal structure for landmark selection. The parotid contours were delineated on all MR images. In addition, for each parotid, six anatomic landmarks were identified based on the bifurcation of artery and vein as pointed on Fig. 4a, b. The contours and landmarks were reviewed and edited by an experienced radiation oncologist. Most landmarks are located close to the central part of parotids. Figure 4 shows the MR image before (Fig. 4a) and after (Fig. 4b) treatment of patient#1, with physician edited contours. The Gross Target Volume (GTV) in green contour, which was shown on the pre-treatment MR, almost disappears completely on the after-treatment image, which leads to large local deformation. The landmarks of patient#1 are visualized in Fig. 4d with parotid 3D-surface, showing the displacement between two images. The corresponding parotid FEM model with the color rendering of magnitude of DVF is illustrated in Fig. 4e. Similar to the masked phantom MR images, a new set of parotid-masked MR images were generated by overriding the intensity inside parotid to simulate imaging modalities without internal contrast like CT (Fig. 4c). IM-DIR and BM-DIR were performed on both MR and masked MR pairs using pre-treatment image as reference. The target registration error (TRE) was evaluated for the output DVFs based on the expert landmarks.