Auto-Segmentation of Bone in MRI-only Based Radiotherapy Using Ultra Short Echo Time

Purpose/Objective: Treatment planning based on MRI-only has shown a promising potential if bulk density assignment for tissue, air and bone are taken into consideration. A major issue is the need to autosegment these bulk tissue structures in the MRI image to make the approach feasible. This is, however, complicated by an extremely short T2 relaxation time (~ 1 ms) of the bone resulting in no signal using conventional MRI sequences. Here, we present an approach adapted from PET/MRI attenuation maps to automatically segment the bone using MRI sequences based on ultra short echo times (UTE).

Materials and Methods: A cutaway from the front leg of a calf including the knee-joint was used as a phantom. The MR images were acquired on a 3.0-T MRI scanner (Philips Achieva) using a cardiac coil to cover the entire phantom. The UTE sequence applies two different echo times, TE1 and TE2, which were 0.2 and 1.9 ms, respectively, a flip angle of 10 o , and a TR of 4.0 ms. An isotropic voxel dimension of 1.8 mm was obtained with a FOV of 240 mm. A reference CT scan (Philips Big Bore CT) was also acquired for comparison. Processing of the TE1 and TE2 MR images was done in MatLab using the DICOM toolbox. First, a TE1-2 image is created by subtracting TE2 from TE1 (figure, row 1). This image is then masked with a binary image of TE2 creating a TE1-2* image with a more well defined outer contour and discrimination between air and tissue (figure, row 2). Two different filters based on the most insensitive pixel (MIP) were applied to autosegment TE1-2* into tissue and bone. Method 1 (M1): tissue (outer contour) > MIP/2.4 > bone > MIP/5 > tissue > 0. Method 2 (M2): bone > MIP/5 > tissue > 0 (figure, row 3). The TE1-2* and segmented M1 and M2 scans were registered with the reference CT scan in Eclipse v.10 (Varian Medical Systems). A structure set containing (auto-segmented) bone and tissue (auto-segmented body - bone) were created for the CT-, M1- and M2-scan (figure, row 3). A CT based treatment plan containing two opposing APPA fields giving 2 Gy to the iso-center (middle of phantom) was created and re-calculated on M1 and M2 using tissue = 0 HU and bone = 362 HU (CT average).

Results: 87 % and 72 % of the M1 and M2 bone agreed with the CT reference (intersection of volumes). Non-bone was 13% and 28 % for M1 and M2 bone (M1/M2 volume-intersection). Bone-miss was 28% and 15% for M1 and M2 (CT volume-intersection). The dosimetric differences were less than 1.5 % in the iso-center. The DVHs of the bone and tissue (figure, row 4) show good agreement between M1 and CT, ΔD98% = 2/-4 % (tissue/bone) while M2 shows a distinct deviation from CT, ΔD98% = 12/-18 % (tissue/bone).