Abstract
Current image processing methods for dynamic susceptibility contrast (DSC) magnetic resonance imaging (MRI) do not capture complex dynamic information of time-signal intensity curves. We investigated whether an autoencoder-based pattern analysis of DSC MRI captured representative temporal features that improves tissue characterization and tumor diagnosis in a multicenter setting. The autoencoder was applied to the time-signal intensity curves to obtain representative temporal patterns, which were subsequently learned by a convolutional neural network. This network was trained with 216 preoperative DSC MRI acquisitions and validated using external data (nā=ā43) collected with different DSC acquisition protocols. The autoencoder applied to time-signal intensity curves and clustering obtained nine representative clusters of temporal patterns, which accurately identified tumor and non-tumoral tissues. The dominant clusters of temporal patterns distinguished primary central nervous system lymphoma (PCNSL) from glioblastoma (AUC 0.89) and metastasis from glioblastoma (AUC 0.95). The autoencoder captured DSC time-signal intensity patterns that improved identification of tumoral tissues and differentiation of tumor type and was generalizable across centers.
Similar content being viewed by others
Introduction
Dynamic susceptibility contrast (DSC) MRI is the most commonly used perfusion technique in clinical practice1 and in clinical trial2, 3 for evaluation of brain tumor vascularity. Much of the discussion of DSC MRI has focused on measurements using relative cerebral blood volume (rCBV)4,5,6,7, and other parameters of relative recirculation8, 9 and vascular permeability10,11,12. However, these techniques are limited to fully capture dynamic information that DSC MRI retains. Every voxel in tumors exhibits different series of signal intensities according to time, and analysis of stationary and one-dimensional (scalar) values do not allow for a complete characterization of dynamic perfusion information. Also, model-based calculation of rCBV and relative recirculation requires sophisticated assumption and introduces potential sources of variances8, 13, 14 from different pulse sequence parameters, contrast media, model assumption from different hospital setting15. Unsupervised learning of time-series data, on the other hand, has a great potential to learn complex time-signal intensity data in DSC MRI16, 17, which can provide additional information that cannot be captured with current perfusion MRI processing.
Time-signal intensity patterns can be learned by an autoencoder neural network for high-dimensional time-series data16. This method uses an unsupervised learning algorithm to learn a compressed representation of the input data. The autoencoder has an encoder-decoder architecture that removes unnecessary noise and generates representations from a low-dimensional latent space by stacking multiple non-linear transformations16, 18. This architecture enables the modeling of complex non-linear functions, unlike linear transformations in principal component analysis19. Also, the low-dimensional latent space in the autoencoder can encode the most important features of the input data, which can then be used in data visualization and become a representation for a supervised learning model.
We hypothesized that the autoencoder would capture temporal patterns from DSC MRI data that are not accessible with current perfusion MRI processing, whichwould yield high diagnostic values in brain tumors with clinical relevance. We investigated whether autoencoder-derived pattern analysis of DSC MRI data captured representative temporal features and augmented tumor characterization in terms of heterogeneity and across multiple centers.
Methods
Patients
This retrospective study was approved by the Institutional Review Board of the Asan Medical Center (approval no. 2019-0594), and all methods were carried out in accordance with the institutionās guidelines and regulations. The requirement for informed consent was formally waived by the approving committee. The electronic database of the Department of Radiology at the University of Ulsan, College of Medicine was searched and records for patients between September 2013 and March 2019 were retrospectively reviewed. In total, 389 patients were identified. The inclusion process is shown in Fig.Ā 1. The inclusion criteria were patients with a single contrast-enhancing tumor revealed by preoperative anatomic MRI and DSC MRI. Exclusion criteria were as follows: (1) no pathologic confirmation; (2) patients who had a prior tumor treated with surgery, radiation, or chemotherapy; (3) multiple contrast-enhancing tumors; and (4) poor image quality. Twenty-three patients were excluded because of prior resection or therapy, and 30 patients were excluded because of multiple tumors. In addition, 50 patients were excluded because DSC MRI data were not available, and 10 patients were excluded because of poor image quality caused by artifacts. After applying the exclusion criteria, there were 276 remaining patients (mean age 60.4Ā years; male: female ratio 152: 124). Of these patients, 137 had glioblastoma, 80 had lymphoma, and 59 had brain metastasis. The autoencoder was developed using training, and internal validation set was formed based on a temporal split of the institutional data obtained in September 2013 and July 2018. The training set included 216 patients (glioblastoma:PCNSL:metastasisā=ā109:58:49). The internal validation set included 60 patients (glioblastoma:PCNSL:metastasisā=ā28:22:10).
To test the generalizability and performance of the algorithm20, an external validation set consisting of 43 patients from another tertiary medical center (Severance Hospital, Seoul, Korea) was included. A total of 28 patients with pathologically confirmed glioblastoma and 15 patients with PCNSL were used for external validation of the model.
Imaging data
All MRI sessions on enrolled patients at both institutions were performed on a 3Ā T unit (Achieva or Ingenia; Philips Medical Systems, Best, Netherlands) equipped with a 32-channel head coil. The brain-tumor imaging protocol included the following sequences: T2-weighted, fluid-attenuated inversion recovery (FLAIR), T1-weighted, diffusion-weighted, contrast-enhanced T1-weighted (CE-T1WI), and DSC perfusion MRI. The DSC data were collected using T2*-weighted gradient echo-planar images.
The DSC MRI protocol used for the training and internal validation sets began with a preload of 0.1Ā mmol/kg gadoterate meglumine (Dotarem; Guerbet), was given before the dynamic bolus. Then, the dynamic bolus was administered as a standard dose of 0.1Ā mmol/kg gadoterate meglumine delivered at a rate of 4Ā mL/s by an MRI-compatible power injector (Spectris; Medrad). The bolus of contrast material was followed by a 20Ā mL bolus of saline, which was injected at the same rate. The pulse sequence parameters included the following: repetition time (TR)/echo time (TE), 1724/40Ā ms; flip angle, 35Ā°; field of view (FOV), 240Ā mm; slice thickness/gap, 5/2Ā mm; matrix, 128āĆā128; total acquisition time, 1Ā min and 50Ā s. Dynamic acquisition was performed with a temporal resolution of 1.7Ā s, and 60 dynamics were acquired. Each timepoint included 20 slices, covering the whole brain in the same orientation as that used for the conventional sequences.
In the external validation set, DSC images were also obtained with a preload of gadobutrol (0.1Ā mmol/kg Gadovist, Bayer Schering Pharma) given before the dynamic bolus. Then, a 0.1Ā mmol/kg dynamic bolus of gadobutrol was injected at a rate of 4Ā mL/s, followed by a bolus injection of saline (total 20Ā mL at 4Ā mL/s). The pulse sequence parameters included the following: TR/TE, 1600/30Ā ms; flip angle, 40Ā°; FOV, 220Ā mm; slice thickness/gap, 5/1.8Ā mm; 128āĆā128 matrix. The dynamic acquisition was performed with a temporal resolution of 1.6Ā s, and 60 dynamics were acquired.
Labeling of brain tissue and brain tumor
The CE-T1WI and FLAIR images were co-registered to the DSC images to define brain regions. The co-registration process was performed in Statistical Parametric Mapping (SPM12) and used affine transformations with six degrees of freedom and 4th degree B-spline interpolation. The probability template matching in SPM12 was also used to segment and label brain regions21 into gray matter, white matter, and cerebral spinal fluid (CSF). Non-brain tissues were excluded in this step.
Tumor segmentation was performed semi-automatically by a neuroradiologist (2Ā years of experience in neuro-oncological imaging) using the CE-T1WI and FLAIR data. Necrosis and the contrast-enhancing solid portion were selected on the 3D CE-T1WI. The FLAIR images were used to select edema/non-enhancing tumor portions using a segmentation threshold and region-growing segmentation algorithm that was implemented using MITK software (www.mitk.org; German Cancer Research Center, Heidelberg, Germany)22. All segmented images were checked by an expert radiologist (6Ā years of experience in neuro-oncological imaging). For glioblastoma cases, contrast-enhancing tumors, non-enhancing tumors, and necrosis were all segmented. For lymphoma and metastasis cases, contrast-enhancing tumors were segmented. In total, there were nine pre-defined different tissue types in the co-registered DSC images: three types of contrast-enhancing tumors (glioblastoma, lymphoma, and metastasis), non-enhancing tumors, edema, gray matter, white matter, CSF, and necrosis.
Calculation of autoencoder representations
DSC time-signal intensity curves were calculated and the scale was normalized based on the median value of white matter for each brain. Then, outliers were removed from the time-series data by excluding data that wereāĀ±ā1 standard deviation outside the mean signal. After this process, DSC time-signal intensity curves included for each patient were approximately 4.6āĆā104 time-signal intensity curves (mean 46,342, standard deviation, 12,765).
Representations from high-dimensional DSC time-signal intensity data were trained using a single-layer sparse autoencoder.
Each autoencoder consisted of two parts: the encoder and the decoder (Supplementary Fig. S1). Each time-signal intensity curve of 60 time points formed a 60-element input vector. The encoder was a one-dimensional convolutional layer using a rectified linear unit (ReLu) and a one-dimensional max pooling layer. The kernel size of the convolution layer was 2 and the filter size 32; the output size of the convolution layer was kept the same by applying the padding parameter as āsameā. A dropout layer with 0.7 probability was introduced after the batch-normalization layer to reduce overfitting. The number of latent spaces was set to five. The decoder reconstructed data with the same size (60 time points) as the input data through two fully-connected layers. Mean squared error was used to calculate the loss between the output and input signal. The learning rate was 1āĆā10ā5 and the batch size for training was set to 10,000.
The autoencoder-derived representations captured the highest variance from DSC time-signal intensity data and all subjects under the constraint that variance was uncorrelated with each representation. These representations thus reflected various aspects of the DSC time-signal intensity curve patterns including baseline signal, depth of signal decrease, slopes of signal decrease and recovery, and percentage recovery. We trained the number of representations that accounted for more than 99% of the overall variance in the perfusion signal across individuals.
Then, k-means clustering was applied to the autoencoder-derived representations. A total of nine clusters were identified from the autoencoder representations, and the autoencoder representations in the latent space were decoded into decoded time-signal intensity curves representing temporal patterns. The nine clusters of decoded time-signal intensity curves (temporal patterns) were used to characterize the nine pre-defined tissue types, which consisted of three types of contrast-enhancing tumors, non-enhancing tumors, edema, gray matter, white matter, CSF, and necrosis. For this, information on the anatomic position using an xāy-z coordinate system was obtained from the decoded time-signal intensity curves and subsequently used for anatomic mapping. The nine clusters of temporal patterns were mapped onto anatomic images for visualization and to distinguish tumoral regions from normal-appearing brain tissue.
Supervised learning using decoded time-signal intensity curves in brain tumor
Whether the decoded time-signal intensity patterns provided clinically relevant information for brain tumor characterization and differential diagnosis in terms of intratumoral perfusion heterogeneity was investigated. For this analysis, a deep convolutional neural network (CNN) was used to learn the representative time-signal intensity patterns characterizing specific pathologic tumor types. Because there was no voxel-based confirmation of tumor and non-tumor tissue, it was assumed that two dominant patterns exhibiting the highest proportions within the contrast-enhancing lesion represented specific pathologic tumor types. Within the CEL, the proportion of each cluster was calculated. Among them, two clusters with the highest proportions were selected as ādominant clustersā and the decoded time-signal intensity curves from the two dominant clusters were subsequently used as input for CNN. Thus, the input for CNN was all 1āĆā60 input vectors obtained from the decoded time-signal intensity curves from the two dominant clusters. A CNN was designed with 3 convolutional blocks, 2 fully connected layers, and softmax layer for classification. The input data size was 60-element input vector from 60 time points of each time-signal intensity curve. The convolutional blocks consisted of one-dimensional convolutional layer combined by a max-pooling operation, batch-normalization, and a drop-out layer (Supplementary Fig. S2). The kernel size of convolution layer was 2 and the number of feature maps was 32, and the output size of convolution layer was kept same by applying padding parameter as āsameā. Max pooling layer down-sampled the representations by 50% for each repetition and batch-normalization was introduced, followed by drop-out layer with 0.7 of probability. Two fully-connected layers using ReLU were applied to make the same size representatives with the number of classes and then softmax layer was applied to convert these representative values as probabilities. Softmax-cross-entropy cost function was used to calculate the loss between labels and predictions. The learning rate was 1āĆā10ā8 and the batch size for training was set to 1,000. The deep-learning architecture was implemented by using the Tensorflow1.14.
Statistical analysis
Studentās t-tests and the Chi-square tests were used to assess differences in demographic data and the prevalence of each classification category between the training and validation sets.
Tumor differentiation
To test diagnostic performance, the area under the curve (AUC), from receiver operating characteristic curve analysis, was calculated. The optimal thresholds of the AUCs were determined by maximizing the sum of the sensitivity and specificity values calculated for differentiation of brain tumors. To differentiate PCNSL from glioblastoma and brain metastasis from glioblastoma, the diagnostic performance was calculated with glioblastoma as a reference diagnosis. The accuracy, sensitivity, and specificity of correctly diagnosing glioblastoma from other tumors were defined as
where TP is a true positive, TN a true negative, FP a false positive, and FN a false negative.
Comparison to conventional approach
The diagnostic performance of the autoencoder analysis of DSC time-signal intensity patterns was compared with that of the conventional approach using mean rCBV.
P values less than 0.05 were considered to indicate significant differences. Statistical analyses were performed using R statistical software (version 3.6.3, R Core Team, Vienna, Austria).
Results
Patient baseline demographics and clinical characteristics are summarized in Table 1. The training set included 216 patients (109 glioblastoma, 58 PCNSL, and 49 metastasis), and the internal validation set had 60 patients (28 glioblastoma, 22 PCNSL, and 10 metastasis). The external validation set had 43 patients (28 glioblastomas and 15 PCNSLs). There were no significant differences in sex across the sets or in sex among patients in either the training or validation sets.
Encoded representations and clusters of temporal patterns for tissue characterization
The extraction of autoencoder-derived representations, cluster visualization, and decoded time-signal intensity patterns are shown in Fig.Ā 2.
The pre-defined tissue types corresponding to each pattern are shown in Table 2. In normally appearing brain tissue, the dominant clusters were cluster 6 (meanāĀ±āstandard deviation, 24.1āĀ±ā0.4%) in gray matter and cluster 4 (34.4āĀ±ā3.8%) in white matter. The dominant clusters in CSF and necrosis were different, although they were bright, exhibiting high T2 signal intensity with high baseline. The dominant cluster in CSF was cluster 5 (21.6āĀ±ā4.7%), wheres in necrosis it was cluster 1 (76.8āĀ±ā1.0%). Using the autoencoder-derived representations, the differentiation of CEL tumors from other tissue types was 91.1% accurate in the training set and 94.6% accurate in the internal validation set.
FigureĀ 3 shows the nine clustered representative temporal patterns and how the patterns were spatially mapped in the representative patients. The nine clusters of temporal patterns differed in terms of baseline signal, depth of signal decrease, slope of signal decrease and recovery, and percentage recovery. For example, clusters 4 and cluster 6 represented normal brain tissue with intermediate baseline signal intensity and a small depth of signal decrease.
Dominant clusters of temporal patterns in brain tumor
The dominant clusters of temporal patterns for glioblastoma were cluster 5 (32.6%) and cluster 3 (26.7%), as shown in Fig.Ā 4a. These patterns had an intermediate baseline, acute slope of signal decrease, and a large depth of signal decrease. The relative recovery was small, and the pattern distribution was heterogeneous.
The dominant clusters for PCNSL were the same as for glioblastoma, namely cluster 5 (37.0%) and cluster 3 (27.7%). However, the frequency of cluster 3 was slightly higher in PCNSL than in glioblastoma. FigureĀ 4b shows the dominant clusters of temporal patterns for PCNSL, which had an intermediate baseline, less acute signal decrease slope, and small depth of signal decrease. The relative recovery was high. The cluster distribution was more homogeneous in PCNSL than in glioblastoma.
The dominant clusters for metastasis were cluster 8 (40.6%) and cluster 5 (16.9%), and the corresponding temporal patterns are shown in Fig.Ā 4c. Metastasis was characterized by a high baseline and had the most acute signal decrease slope. It also showed a large depth of signal decrease, and the relative recovery was low. Compared with PCNSL, the cluster distribution for metastasis was heterogeneous.
Tumor differentiation and external validation
Table 3 summarizes the diagnostic performance results. Using the dominant clusters of time-signal intensity curves, the CNN distinguished PCNSL from glioblastoma in the training set with an AUC of 0.921 (95% confidence interval [CI] 0.860ā0.951), sensitivity of 93.6%, and specificity of 81.0%. On the internal validation set, the CNN distinguished PCNSL from glioblastoma with an AUC of 0.93 (95% CI 0.821ā0.983), sensitivity of 85.7%, and specificity of 90.9%. External validation of the CNN gave an AUC of 0.89 (95% CI 0.75ā0.97), sensitivity of 95.2%, and specificity of 76.5%.
When distinguishing metastasis from glioblastoma, the CNN had an AUC of 0.89 (95% CI 0.83ā0.94), sensitivity of 81.6%, and specificity of 84.4% on the training set. Internal validation gave an AUC of 0.95 (95% CI 0.83ā0.99), sensitivity of 90.0%, and specificity of 85.7%. There was no external validation set available for metastasis.
For distinguishing metastasis from PCNSL, the CNN had an AUC of 0.80 (95% CI 0.71ā0.87), sensitivity of 74.1%, and specificity of 81.6% on the training set. Internal validation gave an AUC of 0.90 (95% CI 0.75ā0.98), sensitivity of 81.8%, and specificity of 90.0%. There was no external validation set available for metastasis.
The assignments of each tissue type in the external validation set are shown in Supplementary Table S1. The assignment of tissue types was similar to that of normally appearing brain tissue located in cluster 4 (44.0%), whereas the tumoral tissues were located in cluster 5 (49.5%) and cluster 3 (29.3%). In the external validation set, the accuracy of differentiating CEL tumors from other tissue types according to cluster frequencies was 86.7%.
Comparison to conventional rCBV approach
Clustering analysis based on autoencoder-derived temporal patterns had significantly better diagnostic performance when distinguishing PCNSL from glioblastoma than the rCBV approach (AUC 0.79, 95% CI 0.65ā0.90) (Delongās test, Pā<ā0.001). In the external validation set, cluster analysis and rCBV had similar diagnostic performance (AUC 0.92, 95% CI 0.88ā0.98) (Pā=ā0.56) (Table 3).
The clustering analysis of autoencoder-derived temporal patterns also had significantly better diagnostic performance than rCBV for distinguishing metastasis from glioblastoma (AUC 0.81, 95% CI 0.65ā0.92; Delongās test, Pā<ā0.001) and metastasis from PCNSL (AUC 0.59, 95% CI 0.41ā0.76; Delongās test, Pā<ā0.001).
Discussion
In this study, a clustering analysis of autoencoder-derived temporal pattern that used information from complex time-signal intensity curves in the whole brain highlighted how the spatial heterogeneity of DSC imaging can be used for tissue characterization. The high-dimensional information included in DSC MRI obtained from all voxels and patients was reduced into nine representative clusters of temporal patterns. These patterns were functionally clustered to separate tissue types and were 91% accurate when used to differentiate between tumor and non-tumor. The differentiation of glioblastoma from metastasis, which was difficult to be achieved using the scalar-based parameters of rCBV methods, was improved when it was based on the dominant clusters of temporal patterns. These dominant clusters offered high diagnostic performance during tumor diagnosis. Furthermore, this autoencoder method does not require model-based calculation of the DSC signal and external validation. The autoencoder effectively learned the complex DSC time-signal intensity curves, and compared with other perfusion methods, provided additional information on intratumoral perfusion heterogeneity and generalizability across centers, which has clinical significance in perfusion analysis.
The autoencoder-derived temporal pattern enabled the use of comprehensive information in the signal-intensity time curves of DSC MRI, including baseline signal intensity, slope and depth of signal decrease, and percentage recovery. Leakage-corrected rCBV reflects an early phase of information by using signal drop and recovery7, while percentage signal recovery23, derived from the ĪR2* curves obtained from DSC perfusion imaging, reflects a delayed phase of information. These information snapshots, as well as one-dimensional parameters, are affected by the complex interplay of technical (e.g., pulse sequence, contrast agent dose) and pathologic (e.g., vascular permeability, tumor cell volume fraction) factors8, 13. For example, the accuracy of rCBV depends on acquisition parameters such as TE and flip angle14. The calculation of leakage correction is software dependent24, 25, and the percentage signal recovery is also influenced by contrast agent preload and pulse sequence parameters8. Indeed, when we compared the average rCBV values, the rCBVs of glioblastoma were 5.52āĀ±ā2.33 (meanāĀ±āstandard deviation) vs. 7.34āĀ±ā1.51 in the internal and external validation sets, respectively. The average rCBV of PCNSL was 3.77āĀ±ā1.52 vs. 3.38āĀ±ā1.34 in the internal and external validation sets, respectively. The high diagnostic performance of rCBV for differentiating between glioblastoma and PCNSL comes from the higher rCBV values in glioblastoma. The higher rCBV values may be a result of the shorter TE (30Ā ms) in the external validation set compared with the TE of 40Ā ms, which would increase the baseline signal and induce a large signal drop26. These factors influencing the model-based assumptions are a barrier to the general use of DSC MRI. The autoencoder-derived temporal pattern used in this study did not require complex model-based assumptions, which might lead to broader use of the technique. Also, the method did not need correction based on center or scanner type, which illustrates its potential for use in multicenter DSC MRI temporal pattern analysis.
This method provides an effective way of learning from a large amount of DSC MRI data. Because a single DSC perfusion image can have approximately 3.2āĆā105 time-signal intensity curves, achieving supervised learning in all voxels is challenging. Previous studies utilizing DSC perfusion imaging relied on either averaged time-signal intensity curves within a given ROI27 or manually selected single or several voxels in the tissue of interest28, 29. However, these methods do not accurately represent tumor heterogeneity and require discarding spatial information included in DSC time-signal intensity data. The autoencoder-derived temporal pattern method used entire labeled tissues with a probability matching template, and the spatial mapping of autoencoder features captured relevant spatial information in the clustering method. Spatial heterogeneity is a hallmark of cancer30. The visualization of clusters revealed such differences in tumors, e.g., glioblastoma exhibited more intratumoral heterogeneity than PCNSL. Quantifying heterogeneity in representative time-signal intensity patterns will be useful to the investigation of tumor biology.
The autoencoder learned temporal patterns that were subsequently used for supervised learning of tumor diagnosis. In the differentiation of glioblastoma from PCNSL, our method showed high diagnostic performance, with AUCs of 0.93 and 0.89 in the internal and external validation sets, respectively. Also, for distinguishing metastasis from glioblastoma, an AUC of 0.95 was shown in the internal validation. For differentiating metastasis from PCNSL, our method showed an AUC of 0.90 in the internal validation. Machine learning using quantified maps from diffusion-weighted or perfusion-weighted imaging has been applied to the differentiation of the above three diseases31, 32, but our autoencoder approach has strengths in that it captures intratumoral heterogeneity by utilizing time-signal intensity curves. Furthermore, the diagnostic performance was high in our method compared with a maximum accuracy of 69.2%31 amongst three diseases. Glioblastoma and metastasis can be distinguished from PCNSL based on high rCBV and low percentage of signal recovery relative to the low rCBV and high percentage signal recovery seen in PCNSL33. However, neither of these parameters are helpful distinguishing metastasis from glioblastoma, which remains inarguably one of the most difficult distinctions in clinical practice. Indeed, metastatic tumors showed distinct differences from glioblastoma and PCNSL in the dominant cluster, which had a very low relative signal recovery. Furthermore, PCNSL showed the most homogeneous cluster distributions of temporal patterns, whereas metastasis showed vastly different clusters according to the varied cell types. Differentiating the primary tumors of PCNSL and glioblastoma from the secondary brain tumors of metastasis is an important task in neuro-oncology because treatment options differ vastly34, 35. The performance of the autoencoder might benefit tumoral diagnoses based on differences in temporal patterns.
This study has several limitations. First, the diagnostic performance of the CNN dropped in the external validation, possibly as a result of overfitting. Although the external validation used a similar DSC acquisition based on the consensus recommendation3 and standardized assessment of rCBV2, the time-signal intensity curves can nonetheless be affected by technical differences. Thus, the curves might need to be further adjusted if the contrast agent preload and image acquisition protocols are very different. Second, there was no external validation for metastasis because DSC MRI is not commonly performed on patients with suspected brain metastases. Finally, tumors exhibited spatial heterogeneity although dominant clusters were selected to represent tumors. Further sophisticated method to explain entire spatial heterogeneity of DSC time-signal intensity will be needed in the future.
In conclusion, the clustering of autoencoder-derived temporal patterns provided information that was not otherwise accessible with rCBV perfusion methods and showed generalizability across different centers.
Data availability
The datasets generated during and/or analyzed during the current study are available from the corresponding author on reasonable request.
References
Thust, S. C. et al. Glioma imaging in Europe: a survey of 220 centres and recommendations for best clinical practice. Eur. Radiol. 28, 3306ā3317 (2018).
Anzalone, N. et al. Brain gliomas: multicenter standardized assessment of dynamic contrast-enhanced and dynamic susceptibility contrast MR images. Radiology 287, 933ā943 (2018).
Ellingson, B. M. et al. Consensus recommendations for a standardized Brain Tumor Imaging Protocol in clinical trials. Neuro Oncol. 17, 1188ā1198 (2015).
Boxerman, J. L., Schmainda, K. M. & Weisskoff, R. M. Relative cerebral blood volume maps corrected for contrast agent extravasation significantly correlate with glioma tumor grade, whereas uncorrected maps do not. AJNR Am. J. Neuroradiol. 27, 859ā867 (2006).
Paulson, E. S. & Schmainda, K. M. Comparison of dynamic susceptibility-weighted contrast-enhanced MR methods: recommendations for measuring relative cerebral blood volume in brain tumors. Radiology 249, 601ā613 (2008).
Schmainda, K. M. et al. Dynamic susceptibility contrast MRI measures of relative cerebral blood volume as a prognostic marker for overall survival in recurrent glioblastoma: results from the ACRIN 6677/RTOG 0625 multicenter trial. Neuro Oncol. 17, 1148ā1156 (2015).
Aronen, H. J. et al. Cerebral blood volume maps of gliomas: comparison with tumor grade and histologic findings. Radiology 191, 41ā51 (1994).
Boxerman, J. L., Paulson, E. S., Prah, M. A. & Schmainda, K. M. The effect of pulse sequence parameters and contrast agent dose on percentage signal recovery in DSC-MRI: implications for clinical applications. AJNR Am. J. Neuroradiol. 34, 1364ā1369 (2013).
Lee, M. D., Baird, G. L., Bell, L. C., Quarles, C. C. & Boxerman, J. L. Utility of percentage signal recovery and baseline signal in DSC-MRI optimized for relative CBV measurement for differentiating glioblastoma, lymphoma, metastasis, and meningioma. AJNR Am. J. Neuroradiol. 40, 1445ā1450 (2019).
Lee, B. et al. Clinical value of vascular permeability estimates using dynamic susceptibility contrast MRI: improved diagnostic performance in distinguishing hypervascular primary CNS lymphoma from glioblastoma. AJNR Am. J. Neuroradiol. 39, 1415ā1422 (2018).
Taoka, T. et al. Application of histogram analysis for the evaluation of vascular permeability in glioma by the K2 parameter obtained with the dynamic susceptibility contrast method: Comparisons with Ktrans obtained with the dynamic contrast enhance method and cerebral blood volume. Magn. Reson. Imaging 34, 896ā901 (2016).
Ho, C. Y., Cardinal, J. S., Kamer, A. P., Lin, C. & Kralik, S. F. Contrast leakage patterns from dynamic susceptibility contrast perfusion MRI in the grading of primary pediatric brain tumors. AJNR Am. J. Neuroradiol. 37, 544ā551 (2016).
Leu, K., Boxerman, J. L. & Ellingson, B. M. Effects of MRI protocol parameters, preload injection dose, fractionation strategies, and leakage correction algorithms on the fidelity of dynamic-susceptibility contrast MRI estimates of relative cerebral blood volume in gliomas. AJNR Am. J. Neuroradiol. 38, 478ā484 (2017).
Semmineh, N. B. et al. Optimization of acquisition and analysis methods for clinical dynamic susceptibility contrast MRI using a population-based digital reference object. AJNR Am. J. Neuroradiol. 39, 1981ā1988 (2018).
Schmainda, K. M. et al. Moving toward a consensus DSC-MRI Protocol: validation of a low-flip angle single-dose option as a reference standard for brain tumors. AJNR Am. J. Neuroradiol. 40, 626ā633 (2019).
Hinton, G. E. & Salakhutdinov, R. R. Reducing the dimensionality of data with neural networks. Science 313, 504ā507 (2006).
Ranzato, M. A., Huang, F. J., Boureau, Y.-L. & LeCun, Y. Unsupervised learning of invariant feature hierarchies with applications to object recognition. In 2007 IEEE conference on computer vision and pattern recognition 1ā8 (IEEE).
Lei, N. et al. A Geometric Understanding of Deep Learning. Engineering 6, 361ā374 (2020).
Jolliffe, I. T. & Cadima, J. Principal component analysis: a review and recent developments. Philos. Trans. A 374, 20150202 (2016).
Park, S. H. & Han, K. Methodologic guide for evaluating clinical performance and effect of artificial intelligence technology for medical diagnosis and prediction. Radiology 286, 800ā809 (2018).
Ashburner, J. & Friston, K. J. J. N. Unified Segment 26, 839ā851 (2005).
Nolden, M. et al. The Medical Imaging Interaction Toolkit: challenges and advances: 10 years of open-source development. Int. J. Comput. Assist. Radiol. Surg. 8, 607ā620 (2013).
Cha, S. et al. Differentiation of glioblastoma multiforme and single brain metastasis by peak height and percentage of signal intensity recovery derived from dynamic susceptibility-weighted contrast-enhanced perfusion MR imaging. AJNR Am. J. Neuroradiol. 28, 1078ā1084 (2007).
Kelm, Z. S. et al. Variability and accuracy of different software packages for dynamic susceptibility contrast magnetic resonance imaging for distinguishing glioblastoma progression from pseudoprogression. J. Med. Imaging 2, 026001 (2015).
Korfiatis, P. et al. Dynamic susceptibility contrast-MRI quantification software tool: development and evaluation. Tomography 2, 448ā456 (2016).
Smith, M. R., Lu, H. & Frayne, R. Signal-to-noise ratio effects in quantitative cerebral perfusion using dynamic susceptibility contrast agents. Magn. Reson. Med. 49, 122ā128 (2003).
Choi, K. S., Choi, S. H. & Jeong, B. Prediction of IDH genotype in gliomas with dynamic susceptibility contrast perfusion MR imaging using an explainable recurrent neural network. Neuro-Oncology 21, 1197ā1209 (2019).
Akbari, H. et al. Pattern analysis of dynamic susceptibility contrast-enhanced MR imaging demonstrates peritumoral tissue heterogeneity. Radiology 273, 502ā510 (2014).
Bakas, S. et al. In vivo detection of EGFRvIII in glioblastoma via perfusion magnetic resonance imaging signature consistent with deep peritumoral infiltration: the phi-index. Clin. Cancer Res. 23, 4724ā4734 (2017).
OāConnor, J. P. et al. Imaging intratumor heterogeneity: role in therapy response, resistance, and clinical outcome. Clin. Cancer Res. 21, 249ā257 (2015).
Swinburne, N. C. et al. Machine learning for semi-automated classification of glioblastoma, brain metastasis and central nervous system lymphoma using magnetic resonance advanced imaging. Ann. Transl. Med. 7, 232 (2019).
Yun, J. et al. Radiomic features and multilayer perceptron network classifier: a robust MRI classification strategy for distinguishing glioblastoma from primary central nervous system lymphoma. Sci. Rep. 9, 5746 (2019).
Xu, W., Wang, Q., Shao, A., Xu, B. & Zhang, J. The performance of MR perfusion-weighted imaging for the differentiation of high-grade glioma from primary central nervous system lymphoma: a systematic review and meta-analysis. PLoS ONE 12, e0173430 (2017).
Haldorsen, I. S., Espeland, A. & Larsson, E. M. Central nervous system lymphoma: characteristic findings on traditional and advanced imaging. AJNR Am. J. Neuroradiol. 32, 984ā992 (2011).
Hunt, M. A., Jahnke, K., Murillo, T. P. & Neuwelt, E. A. Distinguishing primary central nervous system lymphoma from other central nervous system diseases: a neurosurgical perspective on diagnostic dilemmas and approaches. Neurosurg. Focus 21, E3 (2006).
Acknowledgements
This research was supported by the Basic Science Research Program through the National Research Foundation of Korea (NRF) funded by the Ministry of Education, Science and Technology (grant number: NRF-2020R1A2B5B01001707 and NRF-2020R1A2C4001748) and by a grant (2020IP0097) from the Asan Institute for Life Sciences, Asan Medical Center, Seoul, Korea.
Author information
Authors and Affiliations
Contributions
All authors reviewed the manuscript. J.E.P. contributed to conceptual design and writing the manuscript. H.S.K. contributed to editing the manuscript, conceptual design and project integrity. J.K.L. contributed to autoencoder analysis and deep learning analysis. E.N.J. contributed to image pre-processing and template segmentation. W.H.S. contributed to conceptual feedback and software support. S.Y.P. contributed to statistical analysis. Y.H.K. contributed to database construction and data provision. J.H.K. contributed to clinical/oncologic database curation and oversight.
Corresponding author
Ethics declarations
Competing interests
The authors declare no competing interests.
Additional information
Publisher's note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Supplementary Information
Rights and permissions
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/.
About this article
Cite this article
Park, J.E., Kim, H.S., Lee, J. et al. Deep-learned time-signal intensity pattern analysis using an autoencoder captures magnetic resonance perfusion heterogeneity for brain tumor differentiation. Sci Rep 10, 21485 (2020). https://doi.org/10.1038/s41598-020-78485-x
Received:
Accepted:
Published:
DOI: https://doi.org/10.1038/s41598-020-78485-x
Comments
By submitting a comment you agree to abide by our Terms and Community Guidelines. If you find something abusive or that does not comply with our terms or guidelines please flag it as inappropriate.