1253

Deep Learning based pseudo-CT computation and its application for PET/MR attenuation correction and MR-guided radiation therapy planning

Sandeep Kaushik¹, Cristina Cozzini², Mikael Bylund ³, Jaewon Yang⁴, Dattesh Shanbhag¹, Joakim Jonsson³, Josef Lundman³, Thomas Hope⁴, Tufve Nyholm³, Andrew Leynes⁴, Peder Larson⁴, and Florian Wiesinger²

¹GE Global Research, Bangalore, India, ²GE Healthcare, Munich, Germany, ³Umeå University, Umeå, Sweden, ⁴UCSF, San Fransisco, CA, United States

Synopsis

In this work, we demonstrate a generic deep learning (DL) model that computes pCT images (i.e. continuous density bone) using a single channel ZTE MRI data and is robust to protocol and coil variations (as dictated by application needs). The method was evaluated for PET/MR attenuation correction protocol (low resolution for speed) and MRgRTP dose planning protocol (higher resolution for spatial accuracy). The advantages include a single model for multiple protocols, pCT which are very much like real CT in appearance, as well as excellent quantitative accuracy of estimated bone values in the computed pCT.

Introduction

Generation of pseudo CTs (pCT [HU]) from MR images is of interest for applications such as PET/MR attenuation-correction (AC), and MR-guided radiation therapy planning (MRgRTP). In most previous methods, pCT was computed from multi-channel MRI data: Dixon based LAVA-Flex MR, T1-MR+Atlas, or ZTE+LAVA-Flex MR [1, 2]. In this work, we demonstrate a generic deep learning (DL) model that computes pCT images (i.e. continuous density bone) using a single channel ZTE MRI data and is robust to protocol variations (as dictated by application needs). The method was evaluated for PET/MR attenuation correction protocol (low resolution for speed) and MRgRTP dose planning protocol (higher resolution for spatial accuracy).

Methods and Materials

Patient data: MR scans were performed using a 3T, time-of-flight (TOF) Signa PET/MR scanner (GE Healthcare, Chicago, IL, USA). For PET/MR, 17 patients were scanned at site-A using a fast, low-resolution ZTE protocol for head & neck region: 2.4mm isotropic resolution, 41s scan time, GEM head and neck array coil and 8 channel brain coil. For MRgRTP, 18 patients were scanned at site-B at a higher-resolution: 1.4 mm isotropic, 171s, GEM head and neck array coil. Other ZTE parameters were chosen identical: FA=1°, BW=±62.5kHz, FOV=26.4cm. In addition, we have ZTE-MRI data from multiple other sites with protocol similar to site-A (30 cases) and used only for model training purposes. For all patients, accompanying CT scans were available from earlier examinations. All patient studies were approved by respective Institutional Review Boards, including signed informed consent.

ZTE Pre-processing: Intensity correction was performed on ZTE images using ITK N4 algorithm

CT to ZTE-MR registration: CT images were registered to ZTE using a combination of rigid and diffeomorphic dense registration algorithms developed in ITK [3, 4].

Deep learning based pCT computation: A 3D convolutional neural network of the U-net regression architecture to predict pCT from ZTE-MRI across protocols: 8-layers, Adam optimizer, RMSE cost function. This was implemented using Keras and Tensorflow libraries [5, 6]. The model operates on 3D ZTE patches of size 64x64x32, overlap=75%, and predicts a pCT image patch of the same size. Training was performed on 81856 patches from a total of 50 patients (11 site-A, 10 site-B, 29 other sites) with a 90% train -10% validation split. For testing, 6 site-A and 8 site-B cases (30500 patches) were used. Predicted patches were reconstituted back to form the whole pCT volume. Constant tissue value: Since neither of the applications are sensitive to the soft-tissue density variations, the entire soft-tissue (-50 < HU < 200) region in pCT was assigned a value of 42 HU.

pCT Evaluation: We computed Dice overlap coefficient between CT bone region and DL predicted bone region as well as mean absolute error (MAE) over all tissue in the body as a measure of quantitative accuracy [2].

PET-AC Evaluation: PET image reconstruction was performed offline using the petrecon toolbox v1.26 (GE Healthcare, Chicago, IL, USA) and standard parameter settings (2 iterations, 28 subsets, point spread function kernel, 3.0mm full-width at half-maximum (FWHM) in-plane Gaussian filter followed by axial filtering with a three slice 1:4:1 kernel).

Dose evaluation: The pCT datasets were imported into a Radiation Therapy Planning software (RayStation, RaySearch, Sweden). Dose distributions for both evaluated plans were analyzed, and exported from the treatment planning system MICE [8] to perform a gamma analysis over the full volume as defined by the CT scan.

Result & Discussion

Qualitative comparison with ZTE and real CT suggests that pCT-DL maintained the structural match with ZTE in all regions across protocols, and matched real CT in overlapping regions (Fig.1). Fig.2 demonstrates that using DL based method improves both the Dice overlap and MAE metrics. The comparison of histograms of real CT, pCT-DL, and pCT-a in a typical case (Fig.3) indicates consistent estimation of CT values across tissue densities with pCT-DL as compared to the previous method (pCT-a) which shows significant deviation from real CT. In Table 1 (Fig.4a), the average dose difference between pCT-DL and CT on the Target Volume (TV) is <1% which again indicates the quantitative accuracy of DL based pseudo-CT. In Fig.4, the PET-AC and radiation treatment planning for one patient each from site-A (Fig 4b) and site-B (Fig 4a) respectively are shown as comparison between real CT and pCT-DL utility in target applications.

Conclusion

We have presented a generic ZTE only based pseudo-CT generation using deep learning. The advantages include a single model for multiple protocols and coils, pCT which are very much like real CT in appearance, as well as excellent quantitative accuracy of estimated bone values in the computed pCT.

Acknowledgements

No acknowledgement found.

References

[1]. Mehranian A, et al, Med. Phys. 2016;

[2]. Edmund J.M, et al, Radiat.Oncol. Lond. Engl. 2017

[3]. B.B Avants, et al, Penn Image Computing and Science Laboratory, 2009;

[4]. H. Patel, et al, 36th Ger. Conf. Pattern Recognit. 2014;

[5]. Chollet, F., et al, https://github.com/fchollet/keras;

[6]. Abadi, M., et al, arXiv:1603.04467v2;

[7]. C. Cozzini, et al, ISMRM, 2017;

[8]. T. Nyholm, et al, 3rd ESTRO Forum, 2015

Figures

Fig.1: Head pCT images generated using the trained DL model maintains structural match to ZTE FoV, and bears realistic appearance of true CT. The SNR improvement at higher resolution is also preserved by pCT. The provides the best from both modalities.

Fig.2: In site-B dataset, from 8 subject images, (a) Dice coefficient as a function of CT values across subjects. The Dice values are consistent, and higher in pCT-DL. (b) MAE of predicted tissue values in pCT-DL is lower and narrower compared to CT reference tissue values.

Fig.3: Histogram comparison of CT values in the body region in site-B P02. pCT-a deviates from real CT in both low-density, and high-density tissue regions. pCT-DL closely follows the distribution pattern of real CT values over the full value range of tissues.

Fig. 4: (a) Treatment planning comparison example between pCT-DL (top) and real CT (bottom). The dose distribution on the Target Volume and various organs at risk in its surrounding is shown by various contours. pCT is at <1% error from real CT dose calculations. (b) The computed pCT and PET reconstructed using it look visually close to real CT and PET from CT respectively. Last column shows amplified errors (3x) in CT and PET. CT error is a result of combination of registration, interpolation, and estimation errors. On PET, SUVmax error is at 1.38±1.18% in a typical image from site-A.

Proc. Intl. Soc. Mag. Reson. Med. 26 (2018)

1253