Abstract

souha aouadi¹, Rabih Hammoud², Tarraf Torfeh¹, Satheesh Paloor¹, Noora Al-hammadi¹

¹Hamad Medical Corporation, National Centre for Cancer Care and Research, Doha, Qatar; ²Hamad Medical Corporation, National Centre for Cancer Care and Research, Doha, Qatar

To create a synthetic CT (sCT) from T2-weighted brain MRI using 3D convolutional neural networks  (CNN) algorithm and to assess the resulting image quality in comparison to reference CT.

Conventional T2-weighted MRI (1.5T GE MRI, PROPELLER, TR = 6144.9 ms, TE = 89.82 ms, FA = 160º) and CT datasets from 13 patients who underwent brain radiotherapy were included in this retrospective study. CT and MRI were coregistered and resampled to resolution of 1x1x1mm³. The mask of the background was extracted from MRI using the levelsets algorithm.

A high-resolution, compact 3D convolutional network was used for the generation of sCT. It used a stack of residual dilated convolutions with increasingly large dilation factors which incorporated large volumetric context. The root mean square error was used as the loss function between sCT and CT. The algorithm is available in the open source NiftyNet library as «highresnet».

Geometric assessment of the sCT was performed for all patients using leave one out cross-validation. Voxel-wise Mean Absolute Error (MAE) and Mean Errors (ME) were computed to assess sCT intensities. Bone, soft-tissues and air cavities geometry were quantified by dice (DI), sensitivity(SE) and specificity (SP) indices. MAE Water Equivalent Path Length (MAE_WEPL) was computed for a multitude of spokes starting from the center point of the brain and crossing the whole skull to evaluate the radiologic path length. A comparison with the multiscale and dual contrast patch based method (MDPBM) that was previously published was performed.

Figure 1 gives the visual assessment of the generated sCT using CNN and shows the average MAE in bins of 20 HU. Mean MAE, ME and MAE_WEPL values for sCT evaluation, using CNN, were 121.92 (σ=21.13), -24.28 (σ=33.73), and 2.18(σ=0.46), respectively. Mean MAE, ME and MAE_WEPL values for sCT evaluation, using MDPBM, were 99.69 (σ=11.07), 2.01 (σ=11.18), and 1.72(σ=0.54), respectively. Table 1 shows DI, SE and SP indices for bone, soft-tissues and air cavities using CNN and MDPBM. MDPM demonstrated better performance than CNN in our dataset.

A promising study on the generation and validation of CT-substitute from standard clinical T2 MRI is presented. The algorithm is based on convolutional neural networks. A dosimetric evaluation of the sCT for radiotherapy treatment planning is to be done. Also, further work will be done to assess and improve the method on more patients and different clinical sites.