Deep Neural Network Modeling for Brain Tumor Classification Using Magnetic Resonance Spectroscopic Imaging

doi:10.21203/rs.3.rs-3951768/v1

Download PDF

Article

Deep Neural Network Modeling for Brain Tumor Classification Using Magnetic Resonance Spectroscopic Imaging

https://doi.org/10.21203/rs.3.rs-3951768/v1

This work is licensed under a CC BY 4.0 License

Version 1

posted

You are reading this latest preprint version

This study is motivated by the intricate and expert-demanding nature of magnetic resonance spectroscopy imaging (MRSI) data processing, particularly in the context of brain tumor examinations. Traditional approaches often involve complex manual procedures, requiring substantial expertise. In response, we explore the application of deep neural networks directly on raw MRSI data in the time domain. With brain tumors posing significant health concerns, the imperative for early and accurate detection is paramount for effective treatment. While conventional MRI methods face limitations in rapid and accurate spatial evaluation of diffusive gliomas, accuracy and efficiency are compromised. In contrast, MRSI emerges as a promising tool, offering insights into tissue chemical composition and metabolic alterations. Our proposed model, leveraging deep neural networks, is specifically designed for spectral time series analysis and classification tasks. Trained on a dataset comprising synthetic and real MRSI data from brain tumor patients, the model aims to distinguish MRSI voxels indicative of pathologies from healthy ones. Our results demonstrate the model's robustness in domain transformation, seamlessly adapting from synthetic spectra to in vivo data through a fine-tuning process. Successful classification of MRSI voxels of glioma from healthy tissues underscores the model's potential in clinical applications, signifying a transformative impact on diagnostic and prognostic evaluations in brain tumor examinations. Ongoing research endeavors are directed towards validating these integrated approaches across larger datasets, with the ultimate goal of establishing standardized guidelines and further enhancing their clinical utility.

Health sciences/Biomarkers/Diagnostic markers

Biological sciences/Cancer/Cancer imaging

Biological sciences/Cancer/Head and neck cancer

Brain tumors present a significant public health concern, necessitating early detection and classification for effective treatment. While conventional magnetic resonance imaging (MRI) is widely employed for this purpose, it faces limitations in accuracy and efficiency ^1,2. Magnetic resonance spectroscopic imaging (MRSI), an advanced technique revealing insights into tissue chemical composition, holds promise as a valuable tool in diagnostic and treatment assessment of gliomas ^3–5. Despite proven value in research and clinical applications, MRSI encounters challenges in implementation and analysis. Lower signal-to-noise ratio (SNR) during data acquisition affects metabolite quantification accuracy, and the technique's complex steps demand expertise. The trade-off between spatial and spectral resolution further complicates MRSI. Technological developments aim to address these challenges, yet widespread accessibility to advanced protocols remains limited. Standardized analysis pipelines, software tools, and a unified data format are not entirely optimized; however, there have been some concerted efforts, particularly through consortium initiatives ⁶. Overcoming these challenges holds the potential to significantly enhance the effectiveness of MRSI in brain tumor examination.

Deep learning, a subset of machine learning algorithms, have exhibited significant success across diverse medical imaging applications, including brain tumor diagnosis and classification ^7,8. The recent effectiveness of deep learning, particularly convolutional neural networks (CNNs), in MRSI is promising ^9–14. These algorithms possess the capability to autonomously learn pertinent features from data, eliminating the need for manual feature extraction and streamlining the analysis pipeline, thereby enhancing prediction accuracy. Deep learning models have proven successful in tasks such as classifying spectra based on clinical assessment quality ¹¹ and distinguishing mutation statuses in gliomas based on MRI images ¹⁵. The automation of MRSI data analysis is anticipated to save time and broaden accessibility for clinics with limited exposure to conventional vendor-provided MRSI protocols.

Our proposed model employs a CNN architecture specifically crafted for spectral time series analysis and classification tasks. The CNN is trained on a dataset encompassing synthetic spectroscopic data and authentic data from patients with brain tumors, aiming to characterize tumor-specific features in the raw MR spectra and distinguish them from spectra of healthy tissues. Our investigation aimed to validate the clinical value of MRSI raw data in the time domain by circumventing several post-processing steps inherent in traditional pipelines. To this end, three distinct modeling approaches were implemented: (i) training a model (‘Model 1’) on generated synthetic MR spectra and phantom data to assess success in domain transformation from raw data; (ii) training a classification model (‘Model 2’) on in vivo MRSI data to classify gliomas from healthy MR spectra.

2.1 Simulated brain MR spectra

Dataset simulation followed the approach by Lee and Kim ¹⁴, constructing ground truth target spectra from a spectral basis set encompassing key brain metabolites. The metabolites included in the basis set were aspartate (Asp), creatine (Cr), gamma-aminobutyric acid (GABA), glutamate (Glu), glutamine (Gln), glutathione (GSH), glycine (Gly), glycerophosphocholine (GPC), glycerophosphoethanolamine (GPE), Myo-inositol (Ins), lactate (Lac), N-acetyl-aspartate (NAA), N-acetyl-aspartyl glutamate (NAAG), phosphocholine (PCh), phosphocreatine (PCr), phosphorylethanolamine (PE), scyllo-Inositol (Scy), serine (Ser), and taurine (Tau).

For each simulated spectrum, concentrations for each metabolite were randomly selected within the range of metabolite concentrations in normal adult brain according to the literature ^14,16 based on their intensity in the spectrum. Augmentation was then introduced to the spectra through phase modulation, Lorentzian line broadening, and white noise. These modulated spectra served as training input for the deep learning model. Augmentation not only increased the dataset's size but also mitigated model overfitting, rendering the synthetic spectra more realistic. The white noise, with magnitude [0.05 1], aimed to achieve an SNR ranging from 1 to 3. Subsequently, after Fourier transformation, the simulated spectra were sampled to 1024 points and normalized to the range [0,1] to enhance efficiency during model training. A total of 50,000 MR spectra were simulated, with 80% (40,000) allocated for model training, and 10% (5,000) reserved for validation and testing.

2.2 Brain phantom experiments

An in-house made structural-metabolic phantom was assessed, featuring six tubes each with a 25 mm diameter, symmetrically positioned within a larger cylindrical container measuring 110 mm in diameter. These six tubes were filled with buffered solutions (pH = 7) of brain metabolites: 6 mM of NAA, 8 mM of glutamate, 1 mM of GABA, 4 mM of creatine, 5 mM of choline, 8 mM of Myo-inositol, and 4 mM of lactate. Each small tube contain slightly different concentrations and combination of the above-mentioned metabolites. To expedite T1-relaxation, the tube solutions were doped with Magnevist at a concentration of 1 ml/L. Contrarily, the larger cylindrical container was filled with an aqueous solution devoid of metabolites and remained free from Magnevist doping. MRSI data were acquired using a 3 Tesla Philips MRI system (Ingenia, Best, the Netherlands) with following acquisition parameters: PRESS sequence, TR/TE = 2000/35, matrix size of 32×32, and voxel size of 1×1×5 mm³. MR spectra were augmented with phase and intensity modulations, Lorentzian line broadening, and by adding white noise, which resulted in total 10,000 MR spectra all with 1024 complex data points.

2.3 Human subjects

Training dataset: Multi-institutional clinical data from healthy volunteers (n = 21) and glioma patients (n = 4) including MR spectroscopic data sourced openly from Hingerl et al. ¹⁷ and Archibald et al. ¹⁸, and from national/international hospital partners were used in the current study. The data spans various magnetic field strengths, including 7 and 3 Tesla, across Philips and Siemens MRI systems, and involves different acquisition sequences. Hingerl L. and collegues ¹⁷ acquried 3-dimentional free induction decay (FID)-MRSI from glioma (n = 4) and healthy subjects (n = 6) using 7 T (Magnetom, Siemens Healthcare, Erlangen, Germany, maximum gradient strength = 40 mT/m, maximum slew rate = 200 mT/m/ms) with following acquisition parameters: TR/TE = 280/1.3 milliseconds; flip angle of 30°, 2778 Hz spectral bandwidth, and one average. Archibald J. and colleagues ¹⁸ obtained MRSI data from 15 healthy subjects using a 3 Tesla Philips (Achieva scanner, Best, the Netherlands). The acquisition involved a single-channel Transmit-Receive (T/R) head coil and utilized the Point RESolved Spectroscopy (PRESS) sequence with parameters TE/TR = 22/4000 ms, 32 averages, scan time = 3:12. A total of 16 non-water suppressed spectra were acquired as a baseline, and these spectra were utilized in the current study. In total, we prepared about 20,000 augmented in vivo MR spectra for training and validation.

Test dataset: MRSI data was also acquired from 40 patients with glioma (grades 2–4, 20 females, ages 25–74 years) from a national hospital. Anatomical MRI and proton 2D MRSI were acquired using PET/MRI systems (Siemens Biograph mMR, software version Syngo MR, Erlangen, Germany). Standard MRI sequences were acquired, including pre- and post-contrast enhanced (ce) 3D T1 magnetization prepared rapid gradient echo imaging (MPRAGE), 3D FLAIR, and axial T2-weighted imaging. MRSI data parameters included: 2-dimentional PRESS sequence, TR/TE = 1700/30 ms, flip angle of 90˚, three averages, matrix size of 16 × 16, vector size of 1024, and slice thickness of 16 mm.

Ethical approval

for the study was granted by the Regional Ethics Committee (REC Central Norway, reference numbers 2016/279 and 2018/2243), aligning with the principles of the Helsinki Declaration and national guidelines. All patients signed written informed consent.

2.4 Architecture of Neural Network

In this study, two models were trained to perform parallel tasks: the transformation of time-series free induction decay signals into frequency domain MR spectral representation and the classification of glioma versus healthy data. The primary objective was to develop a model capable of transforming time domain complex series into frequency domain MR spectra, while simultaneously identifying anomalies and excluding non-usable time series data. For model training and validation, a dataset comprising 80,000 MR spectra with 1024 time points were utilized, created through a combination of simulated spectra and data from a phantom containing crucial brain metabolites. The ground truth was established using standard spectral processing techniques, including phase modulation and Fourier Transformation in MATLAB (MathWorks, Natick, MA, USA). Notably, the data were not water suppressed, implying that the water signal would be significantly larger than that of the metabolites of interest, assuming imperfect water suppressions in the clinical data.

For Model 1, we developed a CNN inspired by the AUTOMAP network ¹⁹, customized for one-dimensional data (see Fig. 1). The model comprises three fully connected layers with hyperbolic tangent activation, succeeded by two convolutional layers with ReLU activation ²⁰, and a deconvolution layer. Training utilized mean-squared-error loss and the Adam optimizer optimizer ²¹, employing a batch size of 20, a learning rate of 5e^− 6, and 200 epochs. Notably, the real and imaginary parts of the data were trained separately within the same model.

Next, we aimed to train a CNN model (Model 2) for distinguishing healthy from glioma voxels. Figure 2 illustrates the architecture of the CNN model inspired by DenseNet ²², tailored for one-dimensional time series data. DenseNet's distinctive densely connected structure enables the training of deep models without excessive computational demands, minimizing overfitting. The model incorporates ReLU activation ²⁰, and to adapt it for binary classification a sigmoid activation were appended at the model's conclusion.

2.3 Training

For training the classification model (Model 2), the training dataset underwent partitioning MR spectra from either glioma or healthy tissues into distinct training (80%) and validation datasets (20%), with strict measures to prevent the model from exposure to the test dataset during training. The test dataset for evaluating the performance of the classification model was reserved from one institute (40 patients with gliomas), and their data were not used in the training phases of either Model 1 or Model 2. The training dataset encompasses 20,000 MR spectra from healthy subjects and patients with glioma ¹⁷. Each spectrum, comprising real and imaginary components, featuring 1024 time points. To distinguish between glioma and healthy tissues, each spectrum is labeled using one-hot encoding.

For training the classification model, binary cross entropy served as the loss function, and optimization was achieved using the Adam optimizer ²¹. A batch size of 200, a learning rate of 5e^− 5, and 100 epochs were employed during the training process. The model takes in raw time domain data, featuring separate channels for the real and imaginary components of the complex spectra. The final layer of the model is a fully connected layer with 1024×2 nodes, reconstructing the complex frequency spectra. The implementation utilized TensorFlow 2.5 ²³ and Keras ²⁴, and the training process occurred on a fourth-generation MacBook Pro with an Apple M1 chip, featuring an 8-core CPU, 8-core GPU, 16-core Neural Engine, and 16.0 GB RAM.

2.4 Evaluation

During each epoch, the training set underwent processing through the model to generate frequency domain spectra. Subsequently, the mean squared error (MSE) was calculated, and the model's weights were updated through backpropagation. This iterative process continued throughout the epochs, refining the model's performance over time.

$$\text{M}\text{S}\text{E} = \frac{1}{N}\sum _{i=1}^{N}{\left({y}_{i}-{\widehat{y}}_{i}\right)}^{2}$$

To evaluate Model 1, where the true frequency domains were known, the test dataset was processed, and the MSE was calculated to assess performance and monitor for overfitting. Following the last epoch, the weights were finalized, and the test dataset underwent evaluation to provide a comprehensive assessment of the final model.

In evaluating the performance of Model 2, we leveraged Receiver Operating Characteristic (ROC) curves, a widely accepted metric in medical imaging and classification tasks. The ROC curves provided a comprehensive assessment of the model's ability to discriminate between different classes, in the context of voxels containing brain tumor tissue (“Glioma”) versus healthy (non-tumor) tissue (“Healthy”) using MRSI raw data. The utilization of ROC analysis enhances the robustness of our findings, offering insights into the model's sensitivity and specificity, crucial factors in clinical decision-making and diagnostic accuracy.

Also, we utilized confusion matrices to further assess the performance of our model. The confusion matrix allowed for a detailed examination of the model's classification outcomes, providing insights into true positives, true negatives, false positives, and false negatives. This comprehensive evaluation strategy enhances the reliability of our findings, offering a nuanced understanding of the model's precision, recall, and overall classification accuracy. The combined use of ROC curves and confusion matrices strengthens the robustness of our results, contributing to a thorough characterization of the CNN's performance in the context of glioma vs. non-tumor tissue classification through in vivo MRSI data.

To investigate and visualize how our deep learning classification model (‘Model 2’) predictes tumors from healthy non-tumor tissues, we utilized Grad-CAM (Gradient-weighted Class Activation Mapping) ²⁵. Grad-CAM helps us understand and interpret the model’s decisions by highlighting crucial areas in the input data that significantly affect the model’s output. This involves performing a backward pass to compute the gradients of the target class score concerning the final convolutional layer's feature maps. We prepared 50 consecutive MR spectra for each group of healthy and glioma tissue classes and averaging them to enhance SNR.

3.1. Domain transformation

To conduct domain transformation task, synthetic, phantom, and in vivo MRSI data were employed. The trained model underwent evaluation on an unseen test dataset (greater than 10,000 MR spectra). Notably, on the test dataset, the model demonstrated an excellent fit with the original MR spectra in the frequency domain, as indicated by the MSE (SD) depicted in Fig. 3.

3.2. Densely deep learning model classified MR spectra obtained from gliomas versus healthy voxels

The ROC curve for the test dataset, depicted in Fig. 4a, illustrates our model's ability to balance sensitivity and specificity, emphasizing the discriminative power of the model in terms of the Area Under the Curve (AUC). The AUC of 0.97 achieved for the commendable predictive power of our model in distinguishing glioma from healthy voxels. Next, we evaluated the model’s performance in classifying gliomas from healthy MR spectra using a confusion matrix. Our test dataset comprised a well-balanced representation of glioma and healthy cases to ensure a comprehensive assessment of the model’s classification accuracy. The confusion matrix, depicted in Fig. 4b, provide four interpretations of (i) true positives (54.70%, dark gray box): the number of instances where the model correctly classified glioma voxels; false negatives (1.22%, white box): instances where the model incorrectly predicted healthy when the actual class was glioma; false positives (1.17%, white box): instances where the model incorrectly predicted glioma when the actual class was healthy; true negatives (42.91.8%, light gray box): The number of instances where the model correctly classified healthy cases. The overall sensitivity and specificity of the classification model were 97.8% and 97.3%, respectively.

The proposed model, designed for spectral time series analysis exhibited promising outcomes. Trained on a diverse dataset encompassing synthetic spectroscopic data, phantom data, and real data from brain tumor patients, the model demonstrated its efficacy in discerning the presence of tumor lesions. The three distinct modeling approaches implemented for validation—utilizing generated synthetic MR spectra, fine-tuning on in vivo data, and direct training on in vivo data for glioma classification—contributed to a comprehensive evaluation of the model's performance. The example of MR spectra derived from patients with glioma are depicted in Fig. 5.

The experiments revealed that our model (model 2, the classification model), based on the time domain analysis of MRSI raw data, classified glioma tissue from non-tumor healthy tissue with high accuracy. The model's success in domain transformation from raw data, as demonstrated in the synthetic MR spectra and phantom data experiments, provided a foundation for its subsequent performance on in vivo data. The fine-tuning process further optimized the model's weights, enhancing its adaptability to real-world, clinical scenarios.

3.3. Grad-CAM derived saliency maps depicted metabolite’s contribution in the classification task

Grad-CAM analysis was subsequently applied to visualize the decision-making process of the model. Due to the one-hot encoding, the model evaluates whether a spectrum is healthy or glioma separately, with each decision illustrated in Fig. 6. The model demonstrated a particular interest in the area near the choline peak for glioma evaluation, with activation also observed around 2.25 ppm. Conversely, for assessing healthy spectra, a larger portion of the spectrum was activated, particularly focusing on areas around NAA.

Our study's findings highlight the significant potential of deep learning in the realm of magnetic resonance spectroscopic imaging for brain tumor classification. The demonstrated success of our proposed model in distinguishing tumor tissue from non-tumor healthy tissue based on raw MRSI data in the time domain underscores the promise of automated, data-driven approaches to improve accuracy and streamline the analysis pipeline.

Recently, deep learning models have emerged as valuable tools, addressing challenges in the trade-off between sufficient SNR and higher spatial resolution in MRSI. These models also contribute incremental steps toward automating processing and analysis. In a notable study by Hatami and colleagues in 2018 ¹⁰, CNNs outperformed standard quantification methods and simpler machine learning models, such as random forests. Gurbani et al. ¹² implemented an autoencoder approach for spectral fitting, predicting parameters to capture metabolite peaks and employing wavelet reconstruction to generate a baseline. Additionally, the same group developed a model to identify and discard MR spectra of poor quality by recognizing common artifacts ¹¹. While this approach reduces the workload by removing anomalous spectra, the processing and analysis of acceptable spectra still require expert intervention. Hu et al. ¹³ utilized Long Short-Term Memory (LSTM) in their CNN for denoising data, demonstrating less loss of low-intensity data compared to other deep learning approaches. Lee and Kim ¹⁴ employed a CNN to predict spectra containing only metabolites based on noisy input MR spectra. This model can be viewed as a form of pseudo-fitting, where metabolite concentrations can be determined by line fitting on the predicted spectra or estimated using simplified methods.

Koonjoo and colleagues ²⁶ investigated the use of the AUTOMAP-deep learning framework designed specifically for transforming lower SNR-images acquired at ultra-low-field. They successfully illustrated the utility of the AUTOMAP-based image reconstruction approach across diverse low-field MRI datasets obtained in the mT regime, demonstrating significant improvements in both SNR and image quality. In our study, we adapted AUTOMAP for one-dimensional time series MR data, incorporating a blend of low and high SNR MR spectra. However, during the initial experiments, utilizing solely synthetic data, we observed a tendency of our training with the AUTOMAP-derived structure to overfit the data. To address this issue, we introduced more variation in augmentation and incorporating phantom and in vivo MRSI data into the training process. In contrast, the DenseNet model yielded more robust results that generalized well to unseen in vivo MRSI data. While the DenseNet network effectively performed the classification, a comprehensive evaluation of the generated frequency domain spectra is crucial, necessitating a comparison to standard methods. In future research, we aim to assess the model's performance on other multi-center clinical MRSI data and compare it with alternative computational modeling approaches.

An integral aspect of evaluating and interpreting a model involves identifying the primary input features influencing its classification decisions. Certain compounds identifiable in spectra are associated with specific disease progressions, and our goal is to determine whether this holds true for our model. Additionally, our analysis might uncover compounds crucial for specific disease conditions previously unknown to be involved. In this study, we proposed an extension of AUTOMAP's data reconstruction framework to DenseNet classification problems and presented preliminary results. While the combined models demonstrate precision in classification, achieving a clearer output from Grad-CAM and stronger highlighting of areas an expert would focus on is desirable for enhanced model interpretability. These factors may suggest potential data leakage, where the model recognizes differences between the classes irrelevant to the task. To address this, more in vivo subject data is required to control for such factors, and transfer learning could be employed to guide the model toward the areas of interest.

Considering the influence of SNR on the model's effectiveness, the logical next step involves refining the model by initially excluding low-quality spectra from the training set. This refinement is intended to strengthen the model's resilience, leading to a more precise and dependable classification. The creation of a resilient classification model offers substantial potential for streamlining the analysis of MRSI data, particularly in the diagnosis and monitoring of diverse pathologies, such as gliomas in this specific instance.

The comparison of the CNN's results with existing methodologies and traditional pipelines in brain tumor analysis using MRSI data, highlighted the potential of deep learning algorithms. The automated analysis pipeline, enabled by our model, demonstrates efficiency gains and accessibility improvements, which are crucial for clinical settings with limited exposure to advanced MRSI protocols. To provide clinical benefits, our deep learning approach can further extend to classify tumor grades based on 'raw' MRSI data in the future. This could include the classification of isocitrate hydrogenase (IDH)-mutated gliomas versus IDH wild-type gliomas, or grade 2 versus grade 3 gliomas; however, this task requires additional data from other institutions for validation.

In conclusion, the demonstrated robustness in domain transformation, spanning from synthetic spectra to in vivo data and refined through the fine-tuning process, underscores the adaptability of our model to diverse datasets and real-world scenarios. The model's efficacy in distinguishing glioma from healthy MR spectra further positions it as a valuable tool with potential clinical applications.

The integration of MR spectroscopy with deep learning algorithms holds immense promise for improving the accuracy of in vivo investigation of lower-grade gliomas. As these techniques progress and gain broader acceptance in clinical settings, they have the potential to streamline the diagnostic and prognostic evaluation of glioma patients, providing valuable guidance for clinical decision-making.

However, further research is imperative to validate the performance of these integrated approaches across larger, multi-center datasets and to assess their impact on patient outcomes. Standardizing data collection, evaluation criteria, and reporting guidelines will be crucial to ensure the generalizability and reliability of these methods in diverse clinical scenarios. This ongoing exploration and refinement are essential steps toward establishing the clinical utility of integrated MRSI and deep learning in the management of gliomas, particularly in examining and distinguishing lower-grade (grade 2 versus 3) gliomas in future applications.

Author Contribution

E.B.B. conceived and conducted the experiments, analyzed the data, and drafted the manuscript; M.E. contributed to the study concept, the experimental design, funding acquisition, supervised the project, and revised the manuscript; K.J., A.K., L.E. contributed to the medical data collection, preparation, and interpretation, revised the manuscript, validated the results; J.T.G. provided project resources, validated results, and revised the manuscript. All authors reviewed and editted the manuscript.

Guzman-De-Villoria, J. A., Mateos-Perez, J. M., Fernandez-Garcia, P., Castro, E. & Desco, M. Added value of advanced over conventional magnetic resonance imaging in grading gliomas and other primary brain tumors. Cancer Imaging 14, 35, doi:10.1186/s40644-014-0035-8 (2014).
Jenkinson, M. D., Du Plessis, D. G., Walker, C. & Smith, T. S. Advanced MRI in the management of adult gliomas. Br J Neurosurg 21, 550-561, doi:10.1080/02688690701642020 (2007).
Chronaiou, I., Stensjoen, A. L., Sjobakk, T. E., Esmaeili, M. & Bathen, T. F. Impacts of MR spectroscopic imaging on glioma patient management. Acta Oncol 53, 580-589, doi:10.3109/0284186X.2014.891046 (2014).
Maudsley, A. A. et al. Advanced magnetic resonance spectroscopic neuroimaging: Experts' consensus recommendations. NMR Biomed 34, e4309, doi:10.1002/nbm.4309 (2021).
Andronesi, O. C. et al. Early changes in glioblastoma metabolism measured by MR spectroscopic imaging during combination of anti-angiogenic cediranib and chemoradiation therapy are associated with survival. NPJ Precis Oncol 1, doi:10.1038/s41698-017-0020-3 (2017).
Near, J. et al. Preprocessing, analysis and quantification in single-voxel magnetic resonance spectroscopy: experts' consensus recommendations. NMR Biomed 34, e4257, doi:10.1002/nbm.4257 (2021).
Koh, D. M. et al. Artificial intelligence and machine learning in cancer imaging. Commun Med (Lond) 2, 133, doi:10.1038/s43856-022-00199-0 (2022).
Hosny, A., Parmar, C., Quackenbush, J., Schwartz, L. H. & Aerts, H. Artificial intelligence in radiology. Nat Rev Cancer 18, 500-510, doi:10.1038/s41568-018-0016-5 (2018).
Lundervold, A. S. & Lundervold, A. An overview of deep learning in medical imaging focusing on MRI. Z Med Phys 29, 102-127, doi:10.1016/j.zemedi.2018.11.002 (2019).
Hatami, N., Sdika, M. & Ratiney, H. Magnetic Resonance Spectroscopy Quantification using Deep Learning. arXiv, doi:arXiv:1806.07237 (2018).
Gurbani, S. S. et al. A convolutional neural network to filter artifacts in spectroscopic MRI. Magn Reson Med 80, 1765-1775, doi:10.1002/mrm.27166 (2018).
Gurbani, S. S., Sheriff, S., Maudsley, A. A., Shim, H. & Cooper, L. A. D. Incorporation of a spectral model in a convolutional neural network for accelerated spectral fitting. Magn Reson Med 81, 3346-3357, doi:10.1002/mrm.27641 (2019).
Hu, W. et al. Denoising Single Voxel Magnetic Resonance Spectroscopy with Deep Learning on Repeatedly Sampled In Vivo Data. doi:10.48550/arXiv.2101.11442 (2021).
Lee, H. H. & Kim, H. Intact metabolite spectrum mining by deep learning in proton magnetic resonance spectroscopy of the brain. Magn Reson Med 82, 33-48, doi:10.1002/mrm.27727 (2019).
Bangalore Yogananda, C. G. et al. A novel fully automated MRI-based deep-learning method for classification of IDH mutation status in brain gliomas. Neuro Oncol 22, 402-411, doi:10.1093/neuonc/noz199 (2020).
Perry, T. L., Hansen, S., Berry, K., Mok, C. & Lesk, D. Free amino acids and related compounds in biopsies of human brain. J Neurochem 18, 521-528, doi:10.1111/j.1471-4159.1971.tb11980.x (1971).
Hingerl, L. et al. Clinical High-Resolution 3D-MR Spectroscopic Imaging of the Human Brain at 7 T. Invest Radiol 55, 239-248, doi:10.1097/RLI.0000000000000626 (2020).
Archibald, J. et al. Metabolite activity in the anterior cingulate cortex during a painful stimulus using functional MRS. Sci Rep 10, 19218, doi:10.1038/s41598-020-76263-3 (2020).
Zhu, B., Liu, J. Z., Cauley, S. F., Rosen, B. R. & Rosen, M. S. Image reconstruction by domain-transform manifold learning. Nature 555, 487-492, doi:10.1038/nature25988 (2018).
Agarap, A. F. Deep Learning using Rectified Linear Units (ReLU). arXiv (2019).
Kingma, D. P. & Ba, J. L. Adam: A Method for Stochastic Optimization. arXiv (2014).
Huang, G., Liu, Z., Van der Maaten, L. & Weinberger, K. Q. Densely Connected Convolutional Networks. arXiv: 1608.06993 (2016).
TensorFlow v. 2.5.0 (https://www.tensorflow.org/, 2022).
Keras v. 2.5.0 (https://keras.io, 2015).
Selvaraju, R. R. et al. Grad-CAM: Visual Explanations from Deep Networks via Gradient-Based Localization. International Journal of Computer Vision 128, doi:10.1007/s11263-019-01228-7 (2019).
Koonjoo, N., Zhu, B., Bagnall, G. C., Bhutto, D. & Rosen, M. S. Boosting the signal-to-noise of low-field MRI with deep learning image reconstruction. Sci Rep 11, 8248, doi:10.1038/s41598-021-87482-7 (2021).

No competing interests reported.

Download PDF

Version 1

posted

You are reading this latest preprint version

Deep Neural Network Modeling for Brain Tumor Classification Using Magnetic Resonance Spectroscopic Imaging

Status:

Version 1

Abstract

Figures

1 Introduction

2 Materials and Methods

2.1 Simulated brain MR spectra

2.2 Brain phantom experiments

2.3 Human subjects

2.4 Architecture of Neural Network

2.3 Training

2.4 Evaluation

3 Results

3.1. Domain transformation

3.2. Densely deep learning model classified MR spectra obtained from gliomas versus healthy voxels

3.3. Grad-CAM derived saliency maps depicted metabolite’s contribution in the classification task

4 Discussion

5 Conclusion and Future Perspectives

Declarations

Author Contribution

References

Additional Declarations

Status:

Version 1