Quantitative characterizations of ultrashort echo (UTE) images for supporting air-bone separation in the head

Abstract

Accurate separation of air and bone is critical for creating synthetic CT from MRI to support Radiation Oncology workflow. This study compares two different ultrashort echo-time sequences in the separation of air from bone, and evaluates post-processing methods that correct intensity nonuniformity of images and account for intensity gradients at tissue boundaries to improve this discriminatory power. CT and MRI scans were acquired on 12 patients under an institution review board-approved prospective protocol. The two MRI sequences tested were ultra-short TE imaging using 3D radial acquisition (UTE), and using pointwise encoding time reduction with radial acquisition (PETRA). Gradient nonlinearity correction was applied to both MR image volumes after acquisition. MRI intensity nonuniformity was corrected by vendor-provided normalization methods, and then further corrected using the N4itk algorithm. To overcome the intensity-gradient at air-tissue boundaries, spatial dilations, from 0 to 4 mm, were applied to threshold-defined air regions from MR images. Receiver operating characteristic (ROC) analyses, by comparing predicted (defined by MR images) versus 'true' regions of air and bone (defined by CT images), were performed with and without residual bias field correction and local spatial expansion. The post-processing corrections increased the areas under the ROC curves (AUC) from 0.944 ± 0.012 to 0.976 ± 0.003 for UTE images, and from 0.850 ± 0.022 to 0.887 ± 0.012 for PETRA images, compared to without corrections. When expanding the threshold-defined air volumes, as expected, sensitivity of air identification decreased with an increase in specificity of bone discrimination, but in a non-linear fashion. A 1 mm air mask expansion yielded AUC increases of 1 and 4% for UTE and PETRA images, respectively. UTE images had significantly greater discriminatory power in separating air from bone than PETRA images. Post-processing strategies improved the discriminatory power of air from bone for both UTE and PETRA images, and reduced the difference between the two imaging sequences. Both post-processed UTE and PETRA images demonstrated sufficient power to discriminate air from bone to support synthetic CT generation from MRI data.

Figure 1

“True” air (magenta) and bone (blue) regions generated from CT images using thresholds of < −400 HU and > 200 HU, respectively.

Figure 2

Uncorrected MR images (a,d,g), estimated bias fields using N4itk (b,e,h), and corrected MR images (c,f,i) of two patients. (a)–(c): UTE images acquired using pre-scan normalization on patient 1; (d)–(f): UTE images acquired using B1 normalization on patient 2; (g)–(i): PETRA images acquired using B1 normalization on patient 2. The images before and after using N4itk are windowed identically.

Figure 2

Uncorrected MR images (a,d,g), estimated bias fields using N4itk (b,e,h), and corrected MR images (c,f,i) of two patients. (a)–(c): UTE images acquired using pre-scan normalization on patient 1; (d)–(f): UTE images acquired using B1 normalization on patient 2; (g)–(i): PETRA images acquired using B1 normalization on patient 2. The images before and after using N4itk are windowed identically.

Figure 3

ROC curves of discrimination of air from bone using UTE (12 patients) and PETRA (5 patients) images without and with bias field correction (BC).

Figure 4

(a) Sensitivity vs. threshold and (b) specificity vs. threshold for UTE and PETRA images without and with bias field corrections. The error bars represent standard errors of the means.

Figure 5

ROC curves of air and bone discrimination from UTE images with two different normalization methods, pre-scan (7 patients) and B1 (5 patients), applied during image acquisition. Solid lines represent the curves with additional bias-field correction.

Figure 6

“True” air (magenta) and bone (blue) regions in the nasal cavity as well as air masks (yellow) dilated by (a) 0 (no dilations), (b) 1 and (c) 3 mm after applying a threshold of 350 on the bias-field corrected and normalized UTE images.

Figure 7

ROC curves of air and bone discrimination using UTE and PETRA images with bias field corrections and with various spatial expansions (from 0 to 4 mm). The curves are shown at high specificities (0.7–1.0) to emphasize reductions in sensitivity with increasing expansion.

Figure 8

(a) Sensitivity vs. threshold and (b) specificity vs. threshold for air-bone separation from UTE (12 patients) and PETRA (5 patients) images without and with 1-mm mask dilation. The error bars represent standard errors of mean.

Figure 9

(a) CT and (b)–(h) synthetic CT from seven different air masks generated by: (b) UTE images, (c) bias field-corrected UTE images, (d) bias-field corrected UTE images with a 1-mm expansion, (e) PETRA images, (f) bias field corrected PETRA images, (g) bias-field-corrected PETRA images with a 1-mm expansion, and (h) bias field-corrected PETRA images with a 2-mm expansion.

Figure 9

(a) CT and (b)–(h) synthetic CT from seven different air masks generated by: (b) UTE images, (c) bias field-corrected UTE images, (d) bias-field corrected UTE images with a 1-mm expansion, (e) PETRA images, (f) bias field corrected PETRA images, (g) bias-field-corrected PETRA images with a 1-mm expansion, and (h) bias field-corrected PETRA images with a 2-mm expansion.