
==== Front
Phys Eng Sci Med
Phys Eng Sci Med
Physical and Engineering Sciences in Medicine
2662-4729
2662-4737
Springer International Publishing Cham

38884673
1443
10.1007/s13246-024-01443-8
Scientific Paper
Prediction of prognosis in glioblastoma with radiomics features extracted by synthetic MRI images using cycle-consistent GAN
Yoshimura Hisanori 12
http://orcid.org/0000-0002-8472-9821
Kawahara Daisuke daika99@hiroshima-u.ac.jp

1
Saito Akito 1
Ozawa Shuichi 13
Nagata Yasushi 13
1 https://ror.org/03t78wx29 grid.257022.0 0000 0000 8711 3200 Department of Radiation Oncology, Graduate School of Biomedical and Health Sciences, Hiroshima University, Hiroshima, 734-8551 Japan
2 https://ror.org/05te51965 grid.440118.8 0000 0004 0569 3483 Department of Radiology, National Hospital Organization Kure Medical Center, Hiroshima, Japan
3 grid.257022.0 0000 0000 8711 3200 Hiroshima High-Precision Radiotherapy Cancer Center, Hiroshima, 732-0057 Japan
17 6 2024
17 6 2024
2024
47 3 12271243
24 5 2023
14 5 2024
© The Author(s) 2024
2024
https://creativecommons.org/licenses/by/4.0/ Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article's Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/.
To propose a style transfer model for multi-contrast magnetic resonance imaging (MRI) images with a cycle-consistent generative adversarial network (CycleGAN) and evaluate the image quality and prognosis prediction performance for glioblastoma (GBM) patients from the extracted radiomics features. Style transfer models of T1 weighted MRI image (T1w) to T2 weighted MRI image (T2w) and T2w to T1w with CycleGAN were constructed using the BraTS dataset. The style transfer model was validated with the Cancer Genome Atlas Glioblastoma Multiforme (TCGA-GBM) dataset. Moreover, imaging features were extracted from real and synthesized images. These features were transformed to rad-scores by the least absolute shrinkage and selection operator (LASSO)-Cox regression. The prognosis performance was estimated by the Kaplan–Meier method. For the accuracy of the image quality of the real and synthesized MRI images, the MI, RMSE, PSNR, and SSIM were 0.991 ± 2.10 ×10-4, 2.79 ± 0.16, 40.16 ± 0.38, and 0.995 ± 2.11 ×10-4, for T2w, and .992 ± 2.63 ×10-4, 2.49 ± 6.89 ×10-2, 40.51 ± 0.22, and 0.993 ± 3.40 ×10-4 for T1w, respectively. The survival time had a significant difference between good and poor prognosis groups for both real and synthesized T2w (p < 0.05). However, the survival time had no significant difference between good and poor prognosis groups for both real and synthesized T1w. On the other hand, there was no significant difference between the real and synthesized T2w in both good and poor prognoses. The results of T1w were similar in the point that there was no significant difference between the real and synthesized T1w. It was found that the synthesized image could be used for prognosis prediction. The proposed prognostic model using CycleGAN could reduce the cost and time of image scanning, leading to a promotion to build the patient’s outcome prediction with multi-contrast images.

Keywords

Neural network
GAN
Cycle-GAN
Radiomics
Machine learning
Hiroshima UniversityOpen Access funding provided by Hiroshima University.

issue-copyright-statement© Australasian College of Physical Scientists and Engineers in Medicine 2024
==== Body
pmcIntroduction

Primary brain tumors are subdivided into different categories [1]. Among them, glioblastoma (GBM) is known to be one of the most lethal and refractory primary brain tumors. The GBM has a median survival of just 15 month after diagnosis. Only past exposure to ionizing radiation and certain genetic syndromes are recognized as risk factors for GBM [2]. Standard therapy consists of surgical resection to the extent, which is safely feasible, followed by radiotherapy [3]. In a clinical trial using fractionated radiotherapy followed by adjuvant temozolomide; the median survival time is still less than two years [4]. The endpoint for the evaluation of treatment is primarily prolongation of overall survival (OS) [5].

There are several methods for predicting OS. Conventionally, factors affecting the prognosis include high tumor grade, multiple lesions, age, sex, Karnofsky performance score, the extent of resection, treatment plan, and several biomarkers [6–8]. Regarding the risk of GBM, ionizing radiation is one of the few known risk factors that reliably indicates an increased risk of glioma development [8]. In addition, radiation-induced GBM is typically seen years after therapeutic radiation, which is indicated for another tumor or condition [5]. Including vinyl chloride, pesticides, smoking, petroleum refining, and synthetic rubber manufacturing have been loosely associated with the development of GBM.

In addition to clinical factors, MRI images have been used for prognosis prediction [9–17]. Image-based computational models play an increasingly important role in precise diagnosis and treatment guidance in neuro-oncology [18]. Recently, quantitative imaging features derived from medical imaging were of added value in predicting outcome parameters in oncology in what is called Radiomics. We have been working with the University of Minnesota to develop a prognosis prediction system [19]. Local control of brain metastasis was predicted, and the proposed method was more accurate than visual evaluation. For the GBM patients, there are several reports on GBM Radiomics using multi-modality MRI images [20–23]. Pak et al. proposed a prognosis prediction model obtaining imaging features from the dynamic contrast enhanced MRI image [24]. In addition, Shim et al. proposed a deep learning-based distant recurrence prediction model using T1-weighted MRI image (T1w), contrast-enhanced T1-weighted MRI images (CE-T1w), fluid attenuated inversion recovery MRI image (FLAIR), and dynamic susceptibility contrast-enhanced MRI images [21]. On the other hand, problems with Radiomics include the time consuming to scan for the multi-contrast MRI image scan, the dependence of some features on scan acquisition and image reconstruction parameters, and the fact that the best model varies with sample size [25, 26].

The correction methods for variation of the imaging features have been proposed with normalization and combat techniques [27]. On the other hand, images with artifacts from MRI imaging are excluded because they could not correct and affect the accuracy of the prognosis prediction [28, 29]. These problems are expected to be solved by new deep learning techniques such as style transfer and improved normalization, which will make MRI more robust [30, 31]. In addition, the style transfer could synthesize missing images [29]. Several algorithms for style transfer have been proposed [31, 32]. Recent studies proposed style transfer techniques using a generative adversarial network (GAN) [33]: pix2pix [34] and cycle-consistent generative adversarial network (CycleGAN) [40]. The style transfer models of MRI images were proposed for the medical images [36–39]. These studies developed a highly accurate method for synthesizing medical images, which contributes to improving radiation diagnosis performance, such as detecting lesions using synthetized images.

Recently, there has been a movement to use style transfer techniques such as GAN to synthesize images and use the synthesized images for Radiomics [40–42]. Comparing previous study with ours, de Farias et al. and Chen et al. both implemented CT super-resolution in GAN and performed radiomics analysis on the synthesized CT images [40, 41]. In our study, we perform T1w and T2w transformations, so although the task of image transfer is similar, we consider them to be different studies in terms of synthesizing different contrasts. In addition, Tixier et al. generated images using MRI images, however used it to harmonize and normalize images across multiple centers [42]. The purpose of this study differs from that of Tixier et al. study, as it synthesized MRI images of different contrasts to predict OS. Islam et al. aimed to complement missing MRI images by synthesizing them [29]. In contrast, this study predicts OS using only the imaging features of the synthesized image. In contrast, we predicted the OS using the real and synthesized MRI images. From these comparisons, the novelty of this study is that it predicts OS even with synthesized images alone. A paper on prognosis prediction using Radiomics with synthesized images has been reported [29]. They synthesized the missing modality and proposed the survival rate prediction. However, there has been not reported image quality and contribution as a predictor of the synthesized MRI images by comparing with the real MRI images. Also, Kaplan–Meier survival analysis has a benefit to divide survival risk by considering time in many small intervals [43].

The current study proposes a CycleGAN-based T1w and T2 weighted MRI image (T2w) style transfer model for GBM patients.

Moreover, we predicted the OS in preoperative images of GBM patients between the real and synthesized MRI images with Kaplan–Meier survival analysis. A feasibility of the style transfer model was evaluated by comparing with the real and synthesized MRI images in the point of image quality and the survival prediction performance.

Material and methods

Figure 1 shows the overall workflow from style transfer to prognosis prediction using Radiomics analysis in the current study. The detail of the dataset is described in ‘Dataset’section. The style transfer is described in ‘Style transfer’ section. The evaluation of image quality is described in ‘Assessment of style transfer with radiomics analysis’section. In addition, the assessment of style transfer with Radiomics analysis is described in ‘Statistical analysis’ section.Fig. 1 Overview of the process of Radiomics-based prognosis prediction models using style transfer. The prognosis prediction model constructed consists of the following processes: Style transfer, Segmentation, Feature Extraction, Dimension Reduction, Prognosis prediction. Dimension Reduction is further composed of parts Normality check, Pearson Test, and LASSO Cox regression. The upper part of the figure shows Real image, and the lower part shows the flow of the synthesized image

Dataset

Training dataset for style transfer

We used the Brain Tumor Segmentation Challenge (BraTS) dataset for the style transfer model building. The BraTS provided a large, annotated dataset of low-grade gliomas and high-grade GBMs to encourage the development of modern methods of tumor segmentation. T1w, CE-T1w, T2w, FLAIR, and tumor segmentation were provided for each case. The current study used T1w and T2w from the BraTS 2017 dataset [44]. In this paper, 150 cases from the BraTS 2017 dataset were used. Figure 2 shows an example of the images used in this paper.Fig. 2 Training dataset for style transfer of (a) T1w, (b) T2w, and (c) tumor segmentation

Test dataset for style transfer/radiomics dataset

In this paper, we use the third-party data of The Cancer Genome Atlas (TCGA) GBM collection (TCGA-GBM) [45]. We used the test dataset called “Segmentation Labels and Radiomic Features for Pre-operative Scans of TCGA-GBM collection (BraTS-TCGA-GBM)” [46] as a test dataset for style transfer and Radiomics analysis. The test dataset of 102 cases contains the same kind of multi contrast MRI images and segmentation as the training dataset. Figure 3(a) shows the workflow for excluding cases from the style transfer dataset. The training dataset used is BraTS 2017 above. Figure 3(a) also shows the process of evaluating the synthesized image. The exclusion of cases was done as follows. First, the test dataset includes some BraTS 2017; 9 overlapping cases were excluded. Next, 1 case with missing survival time data was excluded. Finally, the remaining 92 cases were used as the test dataset, which was used for building the prognosis prediction model with Radiomics analysis. As with the training dataset, T1w and T2w were used. Style transfer and synthesized image evaluation will be presented in subsequent sections. Figure 3(b) shows the feature extraction process to the prognosis prediction from real and synthesized images using the test dataset. Details of each process are discussed in the subsequent sections.Fig. 3 Flowchart of patient exclusion for (a) style transfer dataset and (b) Radiomics dataset using BraTS-TCGA-GBM

Style transfer

Figure 4 shows the network that converts T1w and T2w using CycleGAN. To distinguish between the real image and the synthesized image, the real T2w is denoted as “rT2w” and the synthesized T2w as “sT2w”. Simultaneously, the style transfer model of the T2w to T1w image was constructed as the input of the T2w. The real T1w is denoted as “rT1w”, and the synthesized T1w is denoted as “sT1w”.Fig. 4 T1w to T2w and T2w to T1w translation networks using CycleGAN. rT1w and rT2w are real image sT2w and sT1w are synthesized image. GT1→T2 and GT2→T1 are the generator. The generator produces a synthesized image by transforming the contrast of the image. DT1, sT1 and DT2, sT2 are the discriminator. The discriminator distinguishes between real image and synthesized images

The CycleGAN has two different networks, generator (G) and discriminator (D), which are the basic structure of GAN, shown in Figs. 5 and 6, respectively. The structure of G and D is different from conventional GAN. The CycleGAN uses a special G based on U-Net [47], used in several segmentation studies [48–50]. It can synthesize highly accurate images by adding the ResNet [51] structure to the U-Net. In this paper, 9-layer ResNet was used. In addition, CycleGAN uses a special D called PatchGAN [34], which was also used in pix2pix, to stabilize the learning process. PatchGAN discriminates real and synthesized images for each region cut out by the patch size.Fig. 5 An architecture for the CycleGAN generator. The representation size shrinks in the encoder phase, stays constant in the transformer phase, and expands again in the decoder phase. The representation size that each layer outputs is listed below it, in terms of the input image size, K. On each layer is listed the number of channels. Each layer is followed by an instance normalization and Rectified Linear Unit (ReLU) activation function

Fig. 6 An architecture for the CycleGAN discriminator. It calls PatchGAN to discriminate real and synthesized images for each region cut out by the patch size. The representation size from each layer is listed below each layer in terms of the input image size, K. On each layer is listed the number of channels

More detailed definitions of GAN and CycleGAN are given below. First, GAN is a method that trains two different networks, G and D, alternately to obtain highly accurate images. The learning of GAN is described as1 minGmaxDVD,G=Ex∼pdataxlogDx+Ez∼pzzlog1-D(Gz

where GAN is characterized by playing a mini-max game of this function.

Next, learning a GAN is described as2 LGANG,DY,X,Y=Ey∼pdataylogDYy+Ex∼pdataxlog1-DY(Gx

where CycleGAN learns G, which synthesized a highly accurate virtual image G(X), using adversarial loss.

However, for learning GAN, adversarial loss alone is not binding. Therefore, CycleGAN introduces the concept of cycle constancy. The cycle constancy is described as3 LcycG,F=Ex∼pdatax‖FGx-x‖1+Ex∼pdatay‖GFy-y‖1

where F is a function that recovers the distribution X by inverse transformation, as in F(G(x)).

Here, the learning of CycleGAN including cycle constancy is described as4 LG,F,DX,DY=LGANG,DY,X,Y+LGANF,DX,Y,X+λLcycG,F

Finally, learning a CycleGAN is described as5 G∗,F∗=argminG,FmaxDx,DYLG,F,DX,DY

where CycleGAN is characterized by playing a mini-max game of LG,F,DX,DY, just like learning GAN.

The proposed model was implemented using the Ubuntu operating system (18.04) with PyTorch packages (V1.4.0, CUDA 11.2, Python 3.7.12). Moreover, the CycleGAN used the following implementation [52]. Hyper parameter was calculated with the default settings. A model to convert between T1w and T2w was trained with identical hyperparameters and instance normalization. A minibatch of 2D MRI images was randomly selected from the training set at each iteration. We calculated at 200 epochs, although most studies usually stop the calculation at 50 epochs [35, 53, 54]. Moreover, we also used methods “Mini-batch processing”, “Normalization and penalization”, and “Scheduling of learning rate” to prevent mode collapse. These processes were used to run on a 16-GB NVIDIA Tesla P100 GPU.

Evaluation of image quality

The following formula was used to create a difference image, and the synthesized image was visually evaluated:difference=ri,j-t(i,j)2,

where r(i,j) is the value of the pixel (i,j) in the real image, t(i,j) is the value of pixel (i,j) in the synthesized image.

The difference image was normalized by the following formula:Y=xi-minximaxxi-minxi,

where xi is the value of pixel, and minxi and maxxi are the minimum and maximum value of pixel.

The prediction accuracy of the model for synthesized images was evaluated using the following four metrics: mutual information (MI), root mean square error (RMSE), signal-to-noise ratio (PSNR), and structural similarity index (SSIM). These metrics are defined as follows:

The MI(X,Y) is used as a cross-modality similarity measure. It is calculated as6 MI(X,Y)=∑x∈X∑y∈YPX,Yx,ylogPX,Y(x,y)PX(x)PY(y)

where PX,Y is the joint probability mass function of X and Y, and PX and PY are the marginal probability mass functions of X and Y, respectively.

The RMSE is defined as7 RMSE=1nxny∑i,jnxnyr(i,j)-t(i,j)r(i,j)2

where r(i,j) is the value of the pixel (i,j) in the real image, t(i,j) is the value of pixel (i,j) in the synthesized image, and nxny is the total number of pixels.

The PSNR is calculated as8 PSNR=20∙log10MAXMSE

where MSE was mean square error. The MSE is calculated as the square of RMSE.

Here, the SSIM between two images, x→ and y→, can be computed as in Wang et al. [55]:9 SSIMx,y=(2μxμy+c1)(2σxy+c2)μx2+μy2+c1σx2+σy2+c2

10 C1=(k1Q)2,k1=0.01

11 C2=(k2Q)2,k2=0.03

where C1 and C2 are constants used to prevent a zero denominator and to maintain the stability of the formula. Q is the maximum value for the real and synthesized images. The values of k1 and k2 are typically obtained from Qi et al. [56]. σx is an estimate in the discrete form:12 σx=1N-1∑i=1N(xi-μx)2

The covariance between x→ and y→ is defined as σxy. It is defined as13 σxy=1N-1∑i=1Nxi-μxyi-μy

where μx is the mean intensity and can be expressed as14 μx=1N∑i=1Nxi

These metrics calculated each case got the average and standard deviation (SD). SD is defined as15 SD=1n∑i=1n(xi-x¯)2

where x¯ is average and n is the number of data.

Assessment of style transfer with radiomics analysis

Feature extraction

The target image was used as the stacked 2D image of rT2w, sT2w, rT1w, and sT1w to extract 3D imaging features. A segmentation data was used from the BraTS-TCGA-GBM. We extracted imaging features using PyRadiomics [57].

These features were extracted from the tumor regions on rT2w, rT1w and synthesized MRIs sT2w, sT1w, respectively. These features include shapes, first-order, second-order (gray level co-occurrence matrix (GLCM), gray level size zone matrix (GLSZM), gray level run length matrix (GLRLM), and gray level dependence matrix (GLDM) features), and higher-order (Laplacian of Gaussian (LoG) filter and wavelet transform).

Dimension reduction

The normality was confirmed by the Shapiro–Wilk test [58] for these features extracted from real and synthesized images, respectively. Then, referring to the analysis by delta-Radiomics, Pearson’s r was calculated for imaging features of real and synthesized images [59]. The Pearson’s r is calculated as16 ρX,Y=σxyσXσY

where σX and σY are the standard deviation of X and Y, respectively. These features with Pearson’s r greater than 0.9 were removed.

In this study, we used the least absolute shrinkage and selection operator (LASSO) regression with the Cox proportional hazards model (LASSO-Cox regression). The LASSO-Cox regression is a data analysis method that can shrink the coefficients of variables unrelated to survival time to zero [60]. These features with non-zero coefficients were selected based on tenfold cross-validation.

These selected features with real and synthesized images were converted to Radiomics score (rad-score) [61]. The rad-score transformation is defined as17 β^=argminℓβ,subject to∑βj≤s

where β^ is the obtained parameters, ℓβ is the log partial likelihood of the logistic regression model, s > 0 is a constant.

Prognosis prediction

Survival curves were estimated from the rad-score with real and synthesized images for each case. These survival curves were estimated by the Kaplan–Meier method [62]. The Kaplan–Meier method is defined as18 S^t=∏i:ti≤t1-dini

where ti is the time at which at least one event occurred, di is the number of events (e.g., deaths) that occurred at time ti, and ni is the number of individuals that have survived (i.e., events have not yet occurred or been censored) until time ti. In addition, a log-rank (LR) test [63] was performed to determine if the hazard functions of the two samples are equal with the median of the rad-score as the cutoff. Let h1t and h2t be the two hazard functions, and define the null hypothesis as19 H0:h1t=h2t

and the alternative hypothesis as20 H0:h1t≠h2t

To construct a test statistic, let r1j and r2j be the numbers of observations just before time tj in samples 1 and 2, with rj=r1j+r2j. Finally, let ij be a variable that takes the value 1 if the event at time tj occurs in sample 1, and 0 otherwise. Then, the LR test is given as21 ZLR=∑j=1nij-∑j=1nr1jrj∑j=1nr1j∙rj2rj2

where, ZLR can be approximated by a normal distribution. When zobs is the observed value of the test statistic, the null hypothesis can be rejected if 2PZLR≥zobs≤0.05.

Statistical analysis

All the statistical analyses were achieved with Python 3.7.12. The Shapiro–Wilk test was implemented using SciPy 1.7.1. Pearson’s r was implemented using the Numpy 1.19.5. Cox’s proportional hazard’s model with the LASSO penalty was implemented using the scikit-survival 0.17.1. The Kaplan–Meier method and LR test were implemented using lifelines 0.26.4.

Results

Style transfer

Figure 7 shows examples of rT2w, style transfer testing results for the sT2w using BraTS-TCGA-GBM dataset, and the difference image between rT2w and sT2w. The color bar of the difference image was set to the maximum value of the difference image. A more detailed numerical evaluation of the accuracy of sT2w is summarized in Table 1. The MI, RMSE, PSNR, and SSIM between rT2 and sT2 were 0.991 ± 2.10 ×10-4, 2.79 ± 0.16, 40.16 ± 0.38, and 0.995 ± 2.11 ×10-4, respectively.Fig. 7 Example of a style transfer testing: (a) rT2w, (b) sT2w, and (c) difference image between rT2w and sT2w. The color bar of the difference image was set to the maximum value of the difference image

Table 1 Average mutual information (MI), average root mean square error (RMSE), average peak signal-to-noise ratio (PSNR), and average structural similarity index measure (SSIM) computed from the synthesized T2w from the T1w

	MI	RMSE	PSNR	SSIM	
Average SD	0.9912.10 ×10-4	2.790.16	40.160.38	0.9952.11 ×10-4	

Figure 8 shows examples of rT1w, style transfer testing results for the sT1w using BraTS-TCGA-GBM dataset, and the difference image between rT1w and sT1w. The color bar of the difference image was set to the maximum value of the difference image. A more detailed numerical evaluation of the accuracy of sT1w is summarized in Table 2. The MI, RMSE, PSNR, and SSIM between rT1 and sT1 were 0.992 ± 2.63 ×10-4, 2.49 ± 6.89 ×10-2, 40.51 ± 0.22, and 0.993 ± 3.40 ×10-4, respectively.Fig. 8 Example of a style transfer testing: (a) rT1w, (b) sT1w, and (c) difference image between rT1w and sT1w. The color bar of the difference image was set to the maximum value of the difference image

Table 2 Average mutual information (MI), average root mean square error (RMSE), average peak signal-to-noise ratio (PSNR), and average structural similarity index measure (SSIM) computed from the synthesized T1w from the T2w

	MI	RMSE	PSNR	SSIM	
Average SD	0.9922.63 ×10-4	2.4886.89 ×10-2	40.510.22	0.9933.40 ×10-4	

Prediction model with radiomics analysis

For the Radiomics analysis, imaging features were extracted from rT2w, sT2w, rT1w, and sT1w. 1720 features were extracted from the original image and the image with the LoG filter (σ = 2.0 to 5.0 [mm]), and 2752 features were extracted from the image with the wavelet transform. A total of 4528 features were extracted.

These imaging features had no significant difference from the results of the Shapiro–Wilk test (p ≥ 0.05). The Shapiro–Wilk test was performed once for each image feature.

For rT2w and sT2w, 64 features remained after eliminating these features with Pearson’s r greater than 0.9. The LASSO-Cox regression resulted in 20 features for rT2w and 18 features for sT2w. Figure 9 lists the selected features from the LASSO-Cox regression. The same 18 features for rT2w and sT2w were selected. These features included local variability and NonUniformity.Fig. 9 LASSO-Cox regression selected features (a) rT2w and (b) sT2w. The features selected in (a) and (b), respectively, have coefficients obtained by Lasso-cox regression. These coefficients are used to construct the Rad-score

For rT1w and sT1w images, 23 features remained after eliminating these features with Pearson’s r greater than 0.9. The LASSO-Cox regression resulted in 2 features for rT1w and 2 features for sT1w. Figure 10 lists these selected features from the LASSO-Cox regression. These same 2 features for rT1w and sT1w images were selected. The names of these features were log-sigma-5–0-mm-3D_firstorder_Energy and log-sigma-5–0-mm-3D_firstorder_TotalEnergy.Fig. 10 LASSO-Cox regression selected features (a) rT1w and (b) sT1w. The features selected in (a) and (b), respectively, have coefficients obtained by Lasso-cox regression. These coefficients are used to construct the Rad-score

Prognosis prediction

To evaluate the ability of the rad-score to prognosis prediction, patients were divided into two groups according to the median rad-score. Of the two groups, the group with the higher rad-score was designated as the poor prognosis group with a shorter survival time. The group with a lower rad-score was designated as the good prognosis group with a longer survival time.

Figure 11 shows the survival curve estimated from rT2w-rad-score and sT2w-rad-score. In the two groups estimated by the rT2w-rad-score, the survival time of the good prognosis group was longer than that of the poor prognosis group (p < 0.05). Similarly, in the two groups estimated by the sT2w-rad-score, the survival time of the good prognosis group was longer than that of the poor prognosis group (p < 0.05).Fig. 11 The comparison of survival curves between the good prognosis group and poor prognosis group using (a) rT2w-rad-score and (b) sT2w-rad-score, respectively

Figure 12 shows a comparison between the two groups estimated from the rT2w-rad-score and sT2w-rad-score for each survival curve. There was no significant difference in survival curves in the good prognosis group estimated from rT2w-rad-score and sT2w-rad-score, respectively (p = 0.99). Similarly, there was no significant difference in survival curves in the poor prognosis group estimated from rT2w-rad-score and sT2w-rad-score, respectively (p = 0.98).Fig. 12 The comparison of survival curves of the rT2w-rad-score and sT2w-rad-score in (a) good prognosis and (b) poor prognosis

Figure 13 shows the survival curves estimated from the rT1w-rad-score and sT1w-rad-score. In the two groups estimated by the rT1w-rad-score, the survival curve in the good prognosis group was no significant difference from that in the poor prognosis group (P = 0.05). Similarly, in the two groups estimated by sT1w-rad-score, the survival curve in the good prognosis group had no significant difference from that in the poor prognosis group (P = 0.05).Fig. 13 The comparison of survival curves in the good prognosis group and poor prognosis group using rT1w-rad-score (a) and sT1w-rad-score (b), respectively

Figure 14 shows a comparison between the two groups estimated from the rT1w-rad-score and sT1w-rad-score for each survival curve. There was no significant difference in survival curves in the good prognosis group estimated from rT1w-rad-score and sT1w-rad-score, respectively (p = 1.00). Similarly, there was no significant difference in survival curves in the poor prognosis group estimated from rT1w-rad-score and sT1w-rad-score, respectively (p = 1.00).Fig. 14 The comparison of survival curves of the rT1w-rad-score and sT1w-rad-score in (a) good prognosis and (b) poor prognosis

Discussion

In this paper, different contrast MRI image was on contrast MRI images for T1w or T2w using CycleGAN. Yang et al. proposed to synthesize brain MRI using conditional GAN [64]. Yang et al. proposed a cross-modality MRI style transfer framework using a generator with both an L1 loss term and an adversarial discriminator conditioned. The PSNR, MI, and SSIM between rT2w and sT2w were 22.56, 0.862, and 0.866, respectively, which were lower than the current study. The PSNR, MI, and SSIM between sT1w and rT1w were 22.52, 0.777, and 0.854, respectively, which were also lower than the current study. In another previous study, Wang et al. proposed a 3D cross-modality MRI image network [39]. Although our proposed method is trained in 2D, the values of PSNR and SSIM were higher in the current study. These results suggest that CycleGAN can synthesize brain MRI with high accuracy. However, it is unclear whether it can be used in clinical if only the image quality evaluation. Several studies on synthesized images for the dose calculation have been reported [65, 66]. Few papers report whether synthesized images can be used for prognosis prediction [29]. The current study evaluated whether style transfer can be used for prognosis prediction. For the imaging features of rT1w and sT1w, 2 common features were obtained by LASSO-Cox regression. These features were converted to rad-score for prognosis prediction. There was no significant difference in the good prognosis group estimated by rT1w-rad-score and sT1w-rad-score, respectively. There was no significant difference in the poor prognosis group as well. Hence, it is suggested that the synthesized images were equivalent to the real images even when compared by imaging features. For rT2w and sT2w, 18 common features were obtained by LASSO-Cox regression. The survival curves estimated from rT2w-rad-scores and sT2w-rad-scores have a significant difference between the good prognosis group and the poor prognosis group. Zhang et al. showed that the imaging features extracted from T2w correlated with Ki-67 expression were useful for determining prognosis [67]. Among the imaging features common to rT2 and sT2 that we extracted, log-sigma-2–0-mm-3D_glrlm_GrayLevelNonUniformity, wavelet-HLL_gldm_GrayLevelNonUniformity, and log-sigma-5–0-mm-3D_gldm_DependenceNonUniformity were ranked from those with the highest contribution in the construction of rad-score. These imaging features are involved in Gray level variation, homogeneity, etc. In the study by Pak et al. gldm_SmallDependenceEmphasis_NE_Vegldm_SmallDependenceHighGrayLevelEmpha sis_NE_Ve, firstorder_Energy_NE_Ve and other features have been extracted [24]. Their methods are not directly comparable due to differences in the way statistics are taken and the images used. However, the extracted features are consistent with our study in that they represent the complexity of the texture. In Addition, the tumor information in the T2w was reported to be related to tumor cellularity [68, 69]. The imaging features in the T2w have a possibility to show a predictor of the OS related to gene expression and tumor cellularity. There was no significant difference in the good prognosis group estimated by rT2w-rad-score and sT2w-rad-score, respectively. There was no significant difference in the poor prognosis group as well. Hence, it is suggested that the composite images are comparable to the real images when compared in terms of image features.

Multi-contrast image datasets are known to be useful in the diagnosis of GBM [70–72]. In addition, such data sets may show some missing image data [29]. Style transfer can also help to supplement such data. Thus, prognosis prediction using synthetic images is useful, although challenges remain. It was also suggested that synthesizing “prognostic level images” could be utilized as one of the landing points in the category of style transfer.

Our study has several limitations. The imaging features extracted from T1w and T2w were used for GBM prognosis prediction. Yan et al. showed that attributes obtained from CE-T1w and FLAIR are particularly useful for GBM subtyping [73]. GBM MRI images show different features of the tumor, such as heterogeneity patterns in enhanced tumor (ET) regions, non-enhanced tumor (NET) regions, and edema around the tumor (ED), which may influence prognostic prediction [69]. The sample size of datasets containing multi-contrast MRI images is limited. However, multi-contrast MRI has been found to be more accurate in predicting prognosis [74]. We aimed to improve the accuracy of prognostic prediction by complementing the missing modalities. In the future, we will expand the variety of multi-modality images through image synthesis and increase the segmentation for radiomics analysis from synthesized images such as ED in the FLAIR.

Moreover, tumor resection status and specific image features have been suggested to significantly influence predictive power, and Lasocki et al. have demonstrated that certain image features, such as FLAIR mismatch and enhancement [75] and other features such as necrosis and calcification, indicate specific glioma genotypes. In the further study, we will incorporate clinical factors into the prediction model. Furthermore, the architecture in the image synthesis affects the quality of the synthesized image. Recently, an attention mechanism has been used in the image synthesis. Since it has been shown that the Attention mechanism improves the accuracy of GAN synthetization [76], we expect to improve the accuracy by introducing the Attention mechanism to CycleGAN.

Conclusion

The proposed method using CycleGAN was able to synthesize images with the same accuracy as the real images. Moreover, our proposed CycleGAN-based prognosis model is comparable to the conventional prognosis model with real images. It could reduce the cost and time for the image scan, which leads to promoting building the patients’ outcome prediction with multi-contrast images.

Appendix

Definition of radiomics features

Shape features describe the shape of the traced region of interest (ROI) and its geometric properties such as volume, maximum diameter along different orthogonal directions, maximum surface, tumor compactness, and sphericity. [30]

First-order statistics features describe the distribution of individual voxel values without concern for spatial relationships. These are histogram-based properties reporting the mean, maximum, uniformity, etc. [30]

Second-order features include the so-called textural features [77], obtained by calculating the statistical inter-relationships between neighboring voxels [77]. They provide a measure of the spatial arrangement of the voxel intensities and intra-lesion heterogeneity [35]. The second-order features we have extracted are as follows:

GLCM

The GLCM is a matrix defined on an image, which is a particular interest in the field of medical image analysis. The GLCM extracts the features of (Autocorrelation, Joint Average, Cluster Prominence, etc.) [55].

GLSZM

The GLSZM quantifies gray level zones in an image. A gray level zone is defined as the number of connected voxels that share the same gray level intensity. The GLSZM extracts the features of (Small Area Emphasis, Large Area Emphasis, Gray Level Non-Uniformity, etc.) [55].

GLRLM

The GLRLM method is a way of extracting higher-order statistical texture features based on a gray level run. The gray level run can be described as a line of pixels in a certain direction with the same intensity value. The number of times such a pixel appears is called the run length value. The GLRLM extracts the features of (Short Run Emphasis, Long Run Emphasis, Gray Level Non-Uniformity, etc.) [55].

GLDM

The GLDM quantifies gray level dependencies in an image. A gray level dependency is defined as the number of connected voxels within distance δ that are dependent on the center voxel. GLDM extracts these features of small dependence emphasis, large dependence emphasis, gray level non-uniformity, etc. [55].

More details on the second-order features can be found in the official PyRadiomics documentation [55].

Higher-order features are obtained by filtering the image. These features are used to identify repeating and non-repeating patterns, suppress noise, and obtain detail enhancement. These include LoG filter, wavelet transform, etc., which can extract areas with increasingly coarse texture patterns [30]. These filters are defined as follows.

LoG filter

The LoG filter is known as a filter that can highlight tumor characteristics not visible in the original image [77]. The LoG filter is difined asA1 LoGx,y=-1πσ41-x2+y22σ2e-x2+y22σ2

where the LoG filter is the second-order derivative of the Gaussian filter. The Gaussian filter is difined asA2 gx,y=12πσe-x2+y22σ2

where partial differentiation about x,A3 ∂∂xgx,y=-x2πσ4e-x2+y22σ2

where further partial differentiation about x,A4 ∂2∂x2gx,y=x2-σ22πσ6e-x2+y22σ2

where the LoG filter is defined as the addition of the second-order partial derivative for x and the second-order partial derivative for y.∂2∂x2gx,y+∂2∂y2gx,y=x2+y2-2σ22πσ6e-x2+y22σ2

A5 =-1πσ4-x2+y2+2σ22σ2e-x2+y22σ2

=-1πσ41-x2+y22σ2e-x2+y22σ2

In addition, the LoG filter changes the kernel with parameters called σ [mm] and filter size.

For example, when these parameters are σ = 3 mm and filter size = 5×5, then kernel K is asA6 K=001000121012-16210121000100

The parameters of σ [2.0 to 5.0] were used. Figure Fig. 15 An example of image processing by LoG filter (a) original; (b) σ = 1 mm; (c) σ = 2 mm; (d) σ = 3 mm; (e) σ = 4 mm; and (f) σ = 5 mm. σ is standard deviation of pixel values

15 shows an example of image processing by the LoG filter. The LoG filter can produce different textures depending on the value of σ [mm].

Wavelet transform.

The wavelet transform is known as a method to obtain multiscale textures [78]. The wavelet transform uses the discrete wavelet transform (DWT). The wavelet transform is defined asA7 Wψfb,a=1a∙∫-∞∞ψx-ba¯∙fxdx

where a is scale, b is translation, and ψx is called the mother wavelet.

The original function can be obtained asA8 fx=Cψ-1∫-∞∞da∫-∞∞dba-2Wψfb,aψx-ba

DWT is a discretization of the wavelet transform with a and b as a=2-j and b=2-jk, respectively. DWT is defined asA9 dkj=2j2∙∫-∞∞ψ2jx-k¯∙fxdx

In addition, we used the 3D-DWT, which is represented by the following tensor product.V3=Lx⊕Hx⊗Ly⊕Hy⊗Lz⊕HzA10 =LxLyLz⊕LxHyLz⊕HxLyLz⊕HxHyLz

⊕LxLyHz⊕LxHyHz⊕HxLyHz⊕HxHyHz,

Where ⊕ and ⊕ denote a space direct sum and convolution, respectively, H and L denote high-pass and low-pass filters; x, y, and z are the 3D coordinates. Consequently, we extracted these features of multiscale texture types (LLL, HLL, LHL, HHL, LLH, HLH, LHH, and HHH). Mostly, the fine texture is derived from the details (i.e., high-pass filter), while the coarse texture is derived from the approximation (i.e., low-pass filter) [79].

Figure Fig. 16 An example of image processing by 2D-wavelet transform (a) Approximation; (b) horizontal detail; (c) vertical detail; (d) diagonal detail. The color bar in the figure was set by the maximum value of pixel

16 shows an example of image processing by 2D-wavelet transform. In this example, there are four textures that can be obtained for 2D: LL, LH, HL, and HH. These correspond to image approximations, horizontal detail, vertical detail, and diagonal detail. The Fig. 16 shows the rT2w by wavelet transform.

Funding

Open Access funding provided by Hiroshima University. The authors have not disclosed any funding.

Declarations

Conflict of interest

The authors declare no conflicts of interest associated with this manuscript.

Ethical approval

This article does not contain any studies with human participants or animals performed by any of the authors.

Publisher's Note

Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
==== Refs
References

1. American Association of Neurological Surgeons. Classification of brain tumors. https://www.aans.org/en/Media/Classifications-of-Brain-Tumors. Accessed 2/4/2022
2. Baid U Rane SU Talbar S Overall survival prediction in glioblastoma with radiomic features using machine learning Front Comput Neurosci 2020 10.3389/fncom.2020.00061 32848682
Baid U, Rane SU, Talbar S et al (2020) Overall survival prediction in glioblastoma with radiomic features using machine learning. Front Comput Neurosci. 10.3389/fncom.2020.0006132848682
3. Walker MD Alexander E Jr Hunt WE Evaluation of BCNU and/or radiotherapy in the treatment of anaplastic gliomas: a cooperative clinical trial J Neurosurg 1978 49 3 333 343 10.3171/jns.1978.49.3.0333 355604
Walker MD, Alexander E Jr, Hunt WE et al (1978) Evaluation of BCNU and/or radiotherapy in the treatment of anaplastic gliomas: a cooperative clinical trial. J Neurosurg 49(3):333–343. 10.3171/jns.1978.49.3.0333355604
4. Lacroix M Abi-Said D Fourney DR A multivariate analysis of 416 patients with glioblastoma multiforme: prognosis, extent of resection, and survival J Neurosurg 2001 95 190 198 10.3171/jns.2001.95.2.0190 11780887
Lacroix M, Abi-Said D, Fourney DR et al (2001) A multivariate analysis of 416 patients with glioblastoma multiforme: prognosis, extent of resection, and survival. J Neurosurg 95:190–198. 10.3171/jns.2001.95.2.019011780887
5. Nijiati K Zhou Z Luo C A nomogram for individualized prediction of overall survival in patients with newly diagnosed glioblastoma: a real-world retrospective cohort study BMC Surg 2021 21 1 1 11 10.1186/s12893-021-01233-z 33388031
Nijiati K, Zhou Z, Luo C et al (2021) A nomogram for individualized prediction of overall survival in patients with newly diagnosed glioblastoma: a real-world retrospective cohort study. BMC Surg 21(1):1–11. 10.1186/s12893-021-01233-z33388031
6. Yu W Zhang L Wei Q O6-Methylguanine-DNA Methyltransferase (MGMT): challenges and new opportunities in glioma chemotherapy Front Oncol 2020 9 1547 10.3389/fonc.2019.01547 32010632
Yu W, Zhang L, Wei Q et al (2020) O6-Methylguanine-DNA Methyltransferase (MGMT): challenges and new opportunities in glioma chemotherapy. Front Oncol 9:1547. 10.3389/fonc.2019.0154732010632
7. Liang J Lv X Lu C Prognostic factors of patients with gliomas – an analysis on 335 patients with glioblastoma and other forms of gliomas BMC Cancer 2020 20 1 35 10.1186/s12885-019-6511-6 31941467
Liang J, Lv X, Lu C et al (2020) Prognostic factors of patients with gliomas – an analysis on 335 patients with glioblastoma and other forms of gliomas. BMC Cancer 20(1):35. 10.1186/s12885-019-6511-631941467
8. Ellor SV Pagano-Young TA Avgeropoulos NG Glioblastoma: Background, standard treatment paradigms, and supportive care considerations J Law Med Ethics 2014 42 2 171 182 10.1111/jlme.12133 25040381
Ellor SV, Pagano-Young TA, Avgeropoulos NG (2014) Glioblastoma: Background, standard treatment paradigms, and supportive care considerations. J Law Med Ethics 42(2):171–182. 10.1111/jlme.1213325040381
9. Akbari H Macyszyn L Da X Pattern analysis of dynamic susceptibility contrast-enhanced MR imaging demonstrates peritumoral tissue heterogeneity Radiology 2014 273 2 502 510 10.1148/radiol.14132458 24955928
Akbari H, Macyszyn L, Da X et al (2014) Pattern analysis of dynamic susceptibility contrast-enhanced MR imaging demonstrates peritumoral tissue heterogeneity. Radiology 273(2):502–510. 10.1148/radiol.1413245824955928
10. Law M Young RJ Babb JS Gliomas: Predicting time to progression or survival with cerebral blood volume measurements at dynamic susceptibility-weighted contrast-enhanced perfusion MR imaging Radiology 2008 247 2 490 498 10.1148/radiol.2472070898 18349315
Law M, Young RJ, Babb JS et al (2008) Gliomas: Predicting time to progression or survival with cerebral blood volume measurements at dynamic susceptibility-weighted contrast-enhanced perfusion MR imaging. Radiology 247(2):490–498. 10.1148/radiol.247207089818349315
11. Pope WB Qiao XJ Kim HJ Apparent diffusion coefficient histogram analysis stratifies progression-free and overall survival in patients with recurrent GBM treated with bevacizumab: a multi-center study J Neurooncol 2012 108 3 491 498 10.1007/s11060-012-0847-y 22426926
Pope WB, Qiao XJ, Kim HJ et al (2012) Apparent diffusion coefficient histogram analysis stratifies progression-free and overall survival in patients with recurrent GBM treated with bevacizumab: a multi-center study. J Neurooncol 108(3):491–498. 10.1007/s11060-012-0847-y22426926
12. Zhang Z Jiang H Chen X Identifying the survival subtypes of glioblastoma by quantitative volumetric analysis of MRI J Neurooncol 2014 119 1 207 214 10.1007/s11060-014-1478-2 24828264
Zhang Z, Jiang H, Chen X et al (2014) Identifying the survival subtypes of glioblastoma by quantitative volumetric analysis of MRI. J Neurooncol 119(1):207–214. 10.1007/s11060-014-1478-224828264
13. Mazurowski MA Desjardins A Malof JM Imaging descriptors improve the predictive power of survival models for glioblastoma patients Neuro Oncol 2013 15 10 1389 1394 10.1093/neuonc/nos335 23396489
Mazurowski MA, Desjardins A, Malof JM (2013) Imaging descriptors improve the predictive power of survival models for glioblastoma patients. Neuro Oncol 15(10):1389–1394. 10.1093/neuonc/nos33523396489
14. Bonekamp D, Deike K, Wiestler B, et al (2015) Association of overall survival in patients with newly diagnosed glioblastoma with contrast-enhanced perfusion MRI comparison of intraindividually matched T1-and T2*-based bolus techniques. Journal of Magnetic Resonance Imaging. 10.1002/jmri.24756
15. Gutman DA Cooper LA Hwang SN MR imaging predictors of molecular profile and survival: multi-institutional study of the TCGA glioblastoma data set Radiology 2013 267 2 560 569 10.1148/radiol.13120118 23392431
Gutman DA, Cooper LA, Hwang SN et al (2013) MR imaging predictors of molecular profile and survival: multi-institutional study of the TCGA glioblastoma data set. Radiology 267(2):560–569. 10.1148/radiol.1312011823392431
16. Wb Li Kai T Chen Q MRI manifestions correlate with survival of glioblastoma multiforme patients Cancer Biol Med 2012 9 2 120 123 10.3969/j.issn.2095-3941.2012.02.007 23691466
Wb Li, Kai T, Chen Q et al (2012) MRI manifestions correlate with survival of glioblastoma multiforme patients. Cancer Biol Med 9(2):120–123. 10.3969/j.issn.2095-3941.2012.02.00723691466
17. Li J Liu S Qin Y High-order radiomics features based on T2 FLAIR MRI predict multiple glioma immunohistochemical features: a more precise and personalized gliomas management PLoS ONE 2020 15 1 e0227703 10.1371/journal.pone.0227703 31968004
Li J, Liu S, Qin Y et al (2020) High-order radiomics features based on T2 FLAIR MRI predict multiple glioma immunohistochemical features: a more precise and personalized gliomas management. PLoS ONE 15(1):e0227703. 10.1371/journal.pone.022770331968004
18. Zhou M Scott J Chaudhury B Radiomics in brain tumor: Image assessment, quantitative feature descriptors, and machine-learning approaches Am J Neuroradiol 2018 39 2 208 216 10.3174/ajnr.A5391 28982791
Zhou M, Scott J, Chaudhury B et al (2018) Radiomics in brain tumor: Image assessment, quantitative feature descriptors, and machine-learning approaches. Am J Neuroradiol 39(2):208–216. 10.3174/ajnr.A539128982791
19. Kawahara D Tang X Lee C Predicting the local response of metastatic brain tumor to Gamma Knife radiosurgery by radiomics with a machine learning method Front Oncol 2021 10 569461 10.3389/fonc.2020.569461 33505904
Kawahara D, Tang X, Lee C et al (2021) Predicting the local response of metastatic brain tumor to Gamma Knife radiosurgery by radiomics with a machine learning method. Front Oncol 10:569461. 10.3389/fonc.2020.56946133505904
20. Narang S Lehrer M Yang D Radiomics in glioblastoma: current status, challenges and potential opportunities Transl Cancer Res 2016 5 4 383 397 10.21037/tcr.2016.06.31
Narang S, Lehrer M, Yang D et al (2016) Radiomics in glioblastoma: current status, challenges and potential opportunities. Transl Cancer Res 5(4):383–397. 10.21037/tcr.2016.06.31
21. Shim KY Chung SW Jeong JH Radiomics-based neural network predicts recurrence patterns in glioblastoma using dynamic susceptibility contrast-enhanced MRI Sci Rep 2021 11 1 1 14 10.1038/s41598-021-89218-z 33414495
Shim KY, Chung SW, Jeong JH et al (2021) Radiomics-based neural network predicts recurrence patterns in glioblastoma using dynamic susceptibility contrast-enhanced MRI. Sci Rep 11(1):1–14. 10.1038/s41598-021-89218-z33414495
22. Malik N Geraghty B Dasgupta A MRI radiomics to differentiate between low grade glioma and glioblastoma peritumoral region J Neurooncol 2021 155 2 181 191 10.1007/s11060-021-03866-9 34694564
Malik N, Geraghty B, Dasgupta A et al (2021) MRI radiomics to differentiate between low grade glioma and glioblastoma peritumoral region. J Neurooncol 155(2):181–191. 10.1007/s11060-021-03866-934694564
23. Chiu FY Le NQK Chen CY A Multiparametric MRI-based radiomics analysis to efficiently classify tumor subregions of glioblastoma: a pilot study in machine learning J Clin Med 2021 10 9 2030 10.3390/jcm10092030 34068528
Chiu FY, Le NQK, Chen CY (2021) A Multiparametric MRI-based radiomics analysis to efficiently classify tumor subregions of glioblastoma: a pilot study in machine learning. J Clin Med 10(9):2030. 10.3390/jcm1009203034068528
24. Pak E Choi KS Choi SH Prediction of prognosis in glioblastoma using radiomics features of dynamic contrast-enhanced MRI Korean J Radiol 2021 22 9 1514 10.3348/kjr.2020.1433 34269536
Pak E, Choi KS, Choi SH et al (2021) Prediction of prognosis in glioblastoma using radiomics features of dynamic contrast-enhanced MRI. Korean J Radiol 22(9):1514. 10.3348/kjr.2020.143334269536
25. Larue RT Defraene G De Ruysscher D Quantitative radiomics studies for tissue characterization: a review of technology and methodological procedures Br J Radiol 2017 90 1070 20160665 10.1259/bjr.20160665 27936886
Larue RT, Defraene G, De Ruysscher D et al (2017) Quantitative radiomics studies for tissue characterization: a review of technology and methodological procedures. Br J Radiol 90(1070):20160665. 10.1259/bjr.2016066527936886
26. Haga A, Takahashi W, Aoki S et al. (2019) Standardization of imaging features for radiomics analysis. The Journal of Medical Investigation. 66(12): 35–37. 10.2152/jmi.66.35
27. Da-Ano R Masson I Lucia F Performance comparison of modified ComBat for harmonization of radiomic features for multicenter studies Sci Rep 2020 10 1 1 12 10.1038/s41598-020-66110-w 31913322
Da-Ano R, Masson I, Lucia F et al (2020) Performance comparison of modified ComBat for harmonization of radiomic features for multicenter studies. Sci Rep 10(1):1–12. 10.1038/s41598-020-66110-w31913322
28. Granzier RWY Verbakel NMH Ibrahim A MRI-based radiomics in breast cancer: feature robustness with respect to inter-observer segmentation variability Sci Rep 2020 10 1 1 11 10.1038/s41598-020-70940-z 31913322
Granzier RWY, Verbakel NMH, Ibrahim A et al (2020) MRI-based radiomics in breast cancer: feature robustness with respect to inter-observer segmentation variability. Sci Rep 10(1):1–11. 10.1038/s41598-020-70940-z31913322
29. Islam M Wijethilake N Ren H Glioblastoma multiforme prognosis: MRI missing modality generation, segmentation and radiogenomic survival prediction Comput Med Imaging Graph 2021 91 101906 10.1016/j.compmedimag.2021.101906 34175548
Islam M, Wijethilake N, Ren H et al (2021) Glioblastoma multiforme prognosis: MRI missing modality generation, segmentation and radiogenomic survival prediction. Comput Med Imaging Graph 91:101906. 10.1016/j.compmedimag.2021.10190634175548
30. Rizzo S Botta F Raimondi S Radiomics: the facts and the challenges of image analysis European Radiol Exp 2018 2 1 36 10.1186/s41747-018-0068-z
Rizzo S, Botta F, Raimondi S et al (2018) Radiomics: the facts and the challenges of image analysis. European Radiol Exp 2(1):36. 10.1186/s41747-018-0068-z
31. Andrearczyk V Depeursinge A Müller H Neural network training for cross-protocol radiomic feature standardization in computed tomography J Med Imaging 2019 6 2 024008 10.1117/1.JMI.6.2.024008
Andrearczyk V, Depeursinge A, Müller H (2019) Neural network training for cross-protocol radiomic feature standardization in computed tomography. J Med Imaging 6(2):024008. 10.1117/1.JMI.6.2.024008
32. Kingma DP & Welling M. Auto-encoding variational Bayes. arXiv:1312.6114v10.
33. Goodfellow I Pouget-Abadie J Mirza M Generative adversarial nets Adv Neural Inf Process Syst 2014 27 2672 2680 10.1145/3422622
Goodfellow I, Pouget-Abadie J, Mirza M et al (2014) Generative adversarial nets. Adv Neural Inf Process Syst 27:2672–2680. 10.1145/3422622
34. Isola P, Zhu JY, Zhou T, et al. Image-to-image translation with conditional adversarial networks. arXiv:1611.07004v3.
35. Zhu JY, Park T, Isola P et al. Unpaired image-to-image translation using cycle-consistent adversarial networks. arXiv:1703.10593v7.
36. Kawahara D Nagata Y T1-weighted and T2-weighted MRI image synthesis with convolutional generative adversarial networks Rep Practical Oncol Radiother 2021 26 1 35 42 10.5603/RPOR.a2021.0005
Kawahara D, Nagata Y (2021) T1-weighted and T2-weighted MRI image synthesis with convolutional generative adversarial networks. Rep Practical Oncol Radiother 26(1):35–42. 10.5603/RPOR.a2021.0005
37. Han C, Hayashi H, Rund L, et al. 2018 GAN-based synthetic brain MR image generation. In 2018 IEEE 15th International Symposium on Biomedical Imaging (ISBI 2018). 734–738. 10.1109/ISBI.2018.8363678.
38. Kazuhiro K Werner RA Toriumi F Generative adversarial networks for the creation of realistic artificial brain magnetic resonance images Tomography 2018 4 4 159 163 10.18383/j.tom.2018.00042 30588501
Kazuhiro K, Werner RA, Toriumi F et al (2018) Generative adversarial networks for the creation of realistic artificial brain magnetic resonance images. Tomography 4(4):159–163. 10.18383/j.tom.2018.0004230588501
39. Wang J, Chen Y, Wu Y, et al. 2020 Enhanced generative adversarial network for 3D brain MRI super-resolution. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. 3627–3636. 10.1109/WACV45572.2020.9093603.
40. de Farias EC Di Noia C Han C Impact of GAN-based lesion-focused medical image super-resolution on the robustness of radiomic features Sci Rep 2021 11 1 1 12 10.1038/s41598-021-00898-z 33414495
de Farias EC, Di Noia C, Han C et al (2021) Impact of GAN-based lesion-focused medical image super-resolution on the robustness of radiomic features. Sci Rep 11(1):1–12. 10.1038/s41598-021-00898-z33414495
41. Chen J, Wee L, Dekker A, et al. Improving reproducibility and performance of radiomics in low dose CT using cycle GANs. arXiv:2109.07787v1.
42. Tixier F Jaouen V Hognon C Evaluation of conventional and deep learning based image harmonization methods in radiomics studies Phys Med Biol 2021 66 24 245009 10.1088/1361-6560/ac39e5
Tixier F, Jaouen V, Hognon C et al (2021) Evaluation of conventional and deep learning based image harmonization methods in radiomics studies. Phys Med Biol 66(24):245009. 10.1088/1361-6560/ac39e5
43. Altman DG. London (UK): Chapman and Hall; 1992. Analysis of Survival times.In:Practical statistics for Medical research; pp. 365–93.
44. Multimodal Brain Tumor Segmentation Challenge 2017. https://www.med.upenn.edu/sbia/brats2017/data.html. Accessed 2/5/2022
45. TCGA-GBM. https://wiki.cancerimagingarchive.net/display/Public/TCGA-GBM. Accessed 5/2/2022
46. Segmentation Labels and Radiomic Features for the Pre-operative Scans of the TCGA-GBM collection (BraTS-TCGA-GBM). https://wiki.cancerimagingarchive.net/pages/viewpage.action?pageId=24282666. Accessed 5/2/2022)
47. Ronneberger O, Fischer P, & Brox T. 2015 U-Net: Convolutional networks for biomedical image segmentation. In International Conference on Medical Image Computing and Computer-assisted Intervention. 10.1007/978-3-319-24574-4_28.
48. Piantadosi G, Sansone M, & Sansone C. 2018 Breast segmentation in MRI via U-Net deep convolutional neural networks. In 2018 24th International Conference on Pattern Recognition (ICPR). 3917–3922. 10.1109/ICPR.2018.8545327.
49. Lee B Yamanakkanavar N Choi JY Automatic segmentation of brain MRI using a novel patch-wise U-Net deep architecture PLoS ONE 2020 15 8 e0236493 10.1371/journal.pone.0236493 32745102
Lee B, Yamanakkanavar N, Choi JY (2020) Automatic segmentation of brain MRI using a novel patch-wise U-Net deep architecture. PLoS ONE 15(8):e0236493. 10.1371/journal.pone.023649332745102
50. Skourt BA El Hassani A Majda A Lung CT image segmentation using deep neural networks Procedia Computer Sci. 2018 127 109 113 10.1016/j.procs.2018.01.104
Skourt BA, El Hassani A, Majda A (2018) Lung CT image segmentation using deep neural networks. Procedia Computer Sci. 127:109–113. 10.1016/j.procs.2018.01.104
51. He K, Zhang X, Ren S, et al. 2016 Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. 770–778. 10.1109/CVPR.2016.90.
52. Pytorch-CycleGAN-and-pix2pix, https://github.com/junyanz/pytorch-CycleGAN-and-pix2pix. Accessed 4/2/2022
53. Isola P, Zhu JY, Zhou T, et al. Medical Image Generation Using Generative Adversarial Networks. arXiv preprint arXiv:1703.01279.
54. Singh NK, & Raza K. Medical image generation using generative adversarial networks: A review. Health informatics: A computational perspective in healthcare, 77–96.
55. Wang Z Bovik AC Sheikh HR Image quality assessment: From error visibility to structural similarity IEEE Trans Image Process 2004 13 4 600 612 10.1109/TIP.2003.819861 15376593
Wang Z, Bovik AC, Sheikh HR et al (2004) Image quality assessment: From error visibility to structural similarity. IEEE Trans Image Process 13(4):600–612. 10.1109/TIP.2003.81986115376593
56. Qi M Li Y Wu A Multi-sequence MR image-based synthetic CT generation using a generative adversarial network for head and neck MRI-only radiotherapy Med Phys 2020 47 4 1880 1894 10.1002/mp.14075 32027027
Qi M, Li Y, Wu A et al (2020) Multi-sequence MR image-based synthetic CT generation using a generative adversarial network for head and neck MRI-only radiotherapy. Med Phys 47(4):1880–1894. 10.1002/mp.1407532027027
57. Welcome to pyradiomics documentation! https://pyradiomics.readthedocs.io/en/latest/. Accessed 5/2 2022
58. Shapiro SS Wilk MB An analysis of variance test for normality (complete samples) Biometrika 1965 52 3/4 591 611 10.2307/2333709
Shapiro SS, Wilk MB (1965) An analysis of variance test for normality (complete samples). Biometrika 52(3/4):591–611. 10.2307/2333709
59. Lin P Yang PF Chen S A delta-radiomics model for preoperative evaluation of neoadjuvant chemotherapy response in high-grade osteosarcoma Cancer Imaging 2020 20 1 1 12 10.1186/s40644-019-0283-8
Lin P, Yang PF, Chen S et al (2020) A delta-radiomics model for preoperative evaluation of neoadjuvant chemotherapy response in high-grade osteosarcoma. Cancer Imaging 20(1):1–12. 10.1186/s40644-019-0283-8
60. Tibshirani R The lasso method for variable selection in the Cox model Stat Med 1997 16 4 385 395 10.1002/(sici)1097-0258(19970228)16:4<385::aid-sim380>3.0.co;2-3 9044528
Tibshirani R (1997) The lasso method for variable selection in the Cox model. Stat Med 16(4):385–395. 10.1002/(sici)1097-0258(19970228)16:4%3c385::aid-sim380%3e3.0.co;2-39044528
61. Zheng BH Liu LZ Zhang ZZ Radiomics score: a potential prognostic imaging feature for postoperative survival of solitary HCC patients BMC Cancer 2018 18 1 1 12 10.1186/s12885-018-5024-z 29291726
Zheng BH, Liu LZ, Zhang ZZ et al (2018) Radiomics score: a potential prognostic imaging feature for postoperative survival of solitary HCC patients. BMC Cancer 18(1):1–12. 10.1186/s12885-018-5024-z29291726
62. Kaplan EL Meier P Nonparametric estimation from incomplete observations J Am Stat Assoc 1958 53 282 457 481 10.2307/2281868
Kaplan EL, Meier P (1958) Nonparametric estimation from incomplete observations. J Am Stat Assoc 53(282):457–481. 10.2307/2281868
63. Mantel N Evaluation of survival data and two new rank order statistics arising in its consideration Cancer Chemother Rep 1966 50 3 163 70 5910392
Mantel N (1966) Evaluation of survival data and two new rank order statistics arising in its consideration. Cancer Chemother Rep 50(3):163–705910392
64. Yang Q Li N Zhao Z MRI cross-modality image-to-image translation Sci Rep 2020 10 1 1 18 10.1038/s41598-020-60520-6 31913322
Yang Q, Li N, Zhao Z et al (2020) MRI cross-modality image-to-image translation. Sci Rep 10(1):1–18. 10.1038/s41598-020-60520-631913322
65. Kazemifar S Barragán Montero AM Souris K Dosimetric evaluation of synthetic CT generated with GANs for MRI-only proton therapy treatment planning of brain tumors J Appl Clin Med Phys 2020 21 5 76 86 10.1002/acm2.12856 32216098
Kazemifar S, Barragán Montero AM, Souris K et al (2020) Dosimetric evaluation of synthetic CT generated with GANs for MRI-only proton therapy treatment planning of brain tumors. J Appl Clin Med Phys 21(5):76–86. 10.1002/acm2.1285632216098
66. Nakamura M Megumi N Keiho I Geometric and dosimetric impact of 3D generative adversarial network-based metal artifact reduction algorithm on VMAT and IMPT for the head and neck region Radiat Oncol 2021 16 1 1 9 10.1186/s13014-021-01827-0 33402192
Nakamura M, Megumi N, Keiho I et al (2021) Geometric and dosimetric impact of 3D generative adversarial network-based metal artifact reduction algorithm on VMAT and IMPT for the head and neck region. Radiat Oncol 16(1):1–9. 10.1186/s13014-021-01827-033402192
67. Zhang J Yao K Liu P A radiomics model for preoperative prediction of brain invasion in meningioma non-invasively based on MRI: A multicentre study EBioMedicine 2020 58 102933 10.1016/j.ebiom.2020.102933 32739863
Zhang J, Yao K, Liu P et al (2020) A radiomics model for preoperative prediction of brain invasion in meningioma non-invasively based on MRI: A multicentre study. EBioMedicine 58:102933. 10.1016/j.ebiom.2020.10293332739863
68. Maynard J Okuchi S Wastling S World Health Organization grade II/III glioma molecular status: Prediction by MRI morphologic features and apparent diffusion coefficient Radiology 2020 296 1 111 121 10.1148/radiol.2020191832 32315266
Maynard J, Okuchi S, Wastling S et al (2020) World Health Organization grade II/III glioma molecular status: Prediction by MRI morphologic features and apparent diffusion coefficient. Radiology 296(1):111–121. 10.1148/radiol.202019183232315266
69. Yang Y Han Y Hu X An improvement of survival stratification in glioblastoma patients via combining subregional radiomics signatures Front Neurosci 2021 15 559 10.3389/fnins.2021.683452
Yang Y, Han Y, Hu X et al (2021) An improvement of survival stratification in glioblastoma patients via combining subregional radiomics signatures. Front Neurosci 15:559. 10.3389/fnins.2021.683452
70. Kickingereder P, Neuberger U, Bonekamp D, et al. Radiomic subtyping improves disease stratification beyond key molecular, clinical, and standard imaging characteristics in patients with glioblastoma. Neuro-oncology, 20(6): 848–857.
71. Prasanna P, Patel J, Partovi S, et al. Radiomic features from the peritumoral brain parenchyma on treatment-naive multi-parametric MR imaging predict long versus short-term survival in glioblastoma multiforme: preliminary findings. European radiology. 27:4188–4197
72. Chiu FY Le NQK Chen CY A multiparametric mri-based radiomics analysis to efficiently classify tumor subregions of glioblastoma: a pilot study in machine learning J Clin Med 2021 10 9 2030 10.3390/jcm10092030 34068528
Chiu FY, Le NQK, Chen CY (2021) A multiparametric mri-based radiomics analysis to efficiently classify tumor subregions of glioblastoma: a pilot study in machine learning. J Clin Med 10(9):2030. 10.3390/jcm1009203034068528
73. Yan J Zhang B Zhang S Quantitative MRI-based radiomics for noninvasively predicting molecular subtypes and survival in glioma patients NPJ Precision Oncol 2021 5 1 1 9 10.1038/s41698-021-00205-z
Yan J, Zhang B, Zhang S et al (2021) Quantitative MRI-based radiomics for noninvasively predicting molecular subtypes and survival in glioma patients. NPJ Precision Oncol 5(1):1–9. 10.1038/s41698-021-00205-z
74. Priya S, Liu Y, Ward C, et al. Machine learning based differentiation of glioblastoma from brain metastasis using MRI derived radiomics. Scientific reports. 11(1): 10478. 10.1038/s41598-021-90032-w
75. Lasocki A Buckland ME Drummond KJ Conventional MRI features can predict the molecular subtype of adult grade 2–3 intracranial diffuse gliomas Neuroradiology 2022 64 12 2295 2305 10.1007/s00234-022-02975-0 35606654
Lasocki A, Buckland ME, Drummond KJ et al (2022) Conventional MRI features can predict the molecular subtype of adult grade 2–3 intracranial diffuse gliomas. Neuroradiology 64(12):2295–230535606654
76. Han Z, Zhang H, Schiele B, & Xiao J. 2018 SAGAN: Self-Attention Generative Adversarial Networks. In International Conference on Machine Learning (ICML)
77. Ergen B Baykara M Texture based feature extraction methods for content based medical image retrieval systems Biomed Mater Eng 2014 24 6 3055 3062 10.3233/BME-141127 25227014
Ergen B, Baykara M (2014) Texture based feature extraction methods for content based medical image retrieval systems. Biomed Mater Eng 24(6):3055–3062. 10.3233/BME-14112725227014
78. Fave X Mackin D Yang J Can radiomics features be reproducibly measured from CBCT images for patients with non-small cell lung cancer? Med Phys 2015 42 12 6784 6797 10.1118/1.4934826 26632036
Fave X, Mackin D, Yang J et al (2015) Can radiomics features be reproducibly measured from CBCT images for patients with non-small cell lung cancer? Med Phys 42(12):6784–6797. 10.1118/1.493482626632036
79. Chaddad A Daniel P Niazi T Radiomics evaluation of histological heterogeneity using multiscale textures derived from 3D wavelet transformation of multispectral images Front Oncol 2018 4 8 96 10.3389/fonc.2018.00096
Chaddad A, Daniel P, Niazi T (2018) Radiomics evaluation of histological heterogeneity using multiscale textures derived from 3D wavelet transformation of multispectral images. Front Oncol 4(8):96. 10.3389/fonc.2018.00096
