Evaluation of proton and photon dose distributions recalculated on 2D and 3D Unet-generated pseudoCTs from T1-weighted MR head scans

Abstract

Introduction: The recent developments of magnetic resonance (MR) based adaptive strategies for photon and, potentially for proton therapy, require a fast and reliable conversion of MR images to X-ray computed tomography (CT) values. CT values are needed for photon and proton dose calculation. The improvement of conversion results employing a 3D deep learning approach is evaluated. Material and methods: A database of 89 T1-weighted MR head scans with about 100 slices each, including rigidly registered CTs, was created. Twenty-eight validation patients were randomly sampled, and four patients were selected for application. The remaining patients were used to train a 2D and a 3D U-shaped convolutional neural network (Unet). A stack size of 32 slices was used for 3D training. For all application cases, volumetric modulated arc therapy photon and single-field uniform dose pencil-beam scanning proton plans at four different gantry angles were optimized for a generic target on the CT and recalculated on 2D and 3D Unet-based pseudoCTs. Mean (absolute) error (MAE/ME) and a gradient sharpness estimate were used to quantify the image quality. Three-dimensional gamma and dose difference analyses were performed for photon (gamma criteria: 1%, 1 mm) and proton dose distributions (gamma criteria: 2%, 2 mm). Range (80% fall off) differences for beam's eye view profiles were evaluated for protons. Results: Training 36 h for 1000 epochs in 3D (6 h for 200 epochs in 2D) yielded a maximum MAE of 147 HU (135 HU) for the application patients. Except for one patient gamma pass rates for photon and proton dose distributions were above 96% for both Unets. Slice discontinuities were reduced for 3D training at the cost of sharpness. Conclusions: Image analysis revealed a slight advantage of 2D Unets compared to 3D Unets. Similar dose calculation performance was reached for the 2D and 3D network.

A review on medical imaging synthesis using deep learning and its clinical applications

“…The unpooling layers in the decoder, which up‐sample and therefore reverse pooling layers in the encoder and produce sparse feature maps, were also replaced with deconvolutional layers that produce dense feature maps and the skip connections were replaced with residual shortcuts, inspired by ResNet, to further save computational memory 14 . Neppl et al also replaced the ReLU layer with a generalized parametric ReLU (PReLU) to adaptively adjust the activation function 15 . Torrado‐Carvajal et al added a dropout layer before the first transposed convolution in the decoder to avoid overfitting 16 …”

“…14 Neppl et al also replaced the ReLU layer with a generalized parametric ReLU (PReLU) to adaptively adjust the activation function. 15 Torrado-Carvajal et al added a dropout layer before the first transposed convolution in the decoder to avoid overfitting. 16 Various loss functions have been investigated in the reviewed studies.…”

MR-guided proton therapy: a review and a preview

“…The generic target volume is marked in red, the 95% iso-dose line in green. Adapted from [49] for single-field-uniform-dose (SFUD) proton plans was 0.1 mm. Using the same method (without identification of internal air cavities), Depauw et al [50] also concluded that, based on DVH parameter analysis, clinically acceptable proton dose calculation accuracy can be achieved.…”

“…The relative proton range error was 0.14 ± 1.11%. In parallel, Neppl et al [49] investigated the feasibility of utilizing deep 2D and 3D Unets for sCT generation of the head (Fig. 5b).…”

Deep learning for brain disorders: from data processing to disease treatment

“…Image translation has also been used in a cross-modality context, such as to generate computed tomography (CT) from MR images [69][70][71][72][73][74][75][76][77][78][79][80][81][82][83][84] or to generate MR images of a certain sequence from MR images of another sequence [74,[85][86][87][88][89][90] or a set of other sequences [83,[91][92][93][94][95][96][97]. A large number of methods have been applied to improve attenuation correction on PET/MR scanners [73,76,78,81,98,99] or to enable radiotherapy treatment planning from MRI only [71,77,80,84]. Other studies aim to improve subsequent image processing steps such as segmentation or registration [90], improve classification in case of missing data [86,91,100,...…”