Non-invasive differential diagnosis of brain tumours is currently based on the assessment of tumour vascularity through magnetic resonance imaging (MRI) coupled with dynamic susceptibility contrast (DSC). However, given its limited accuracy, reaching a definitive diagnosis often requires complex neurosurgical interventions that compromise the patients’ quality of life. We applied deep learning on DSC images from histology-confirmed patients with glioblastoma, metastasis or lymphoma, the three most common brain malignancies. The convolutional neural network trained on ~ 50,000 voxels from 40 patients provided intra-tumour probability maps that yielded clinical-grade diagnosis. Performance was tested in 400 additional cases and an external validation cohort (n = 128). The tool reached a three-way accuracy of 0.78, superior to standard diagnosis with cerebral blood volume (0.55) and percentage of signal recovery (0.59) perfusion metrics. Our open-access software, Brain Enhancing Region Radiological analysis (BERRY), demonstrates the potential of voxel-wise probability maps for differential diagnosis of brain tumours using standard-of-care MRI.
Article
An accessible deep learning tool for voxel-wise classification of brain malignancies from perfusion MRI
https://doi.org/10.21203/rs.3.rs-2362207/v1
This work is licensed under a CC BY 4.0 License
Journal Publication
published 01 Mar, 2024
Version 1
posted
You are reading this latest preprint version
Differential diagnosis between the most common brain malignancies, i.e., glioblastoma multiforme (GBM), brain metastasis from solid tumours and primary central nervous system lymphoma (PCNSL) represents a clinical unmet need, as each of these entities requires a distinct therapeutic approach1–3. While pathology evaluation of tumour samples remains the gold standard for diagnosis, it requires invasive neuro-surgical procedures, with a significant risk of complications, and eventually can be confounded by the use of prior medication, such as steroids4,5.
To overcome the need for surgery, magnetic resonance imaging (MRI) with intravenous contrast injection is being explored as a non-invasive support system for differential diagnosis of brain malignancies. GBM, brain metastasis and PCNSL represent up to 70% of all malignant brain tumours and more than 80% of contrast enhancing tumours within the brain6. Nevertheless, the enhancing patterns exhibit a high degree of similarity across these tumour types, making differential diagnosis challenging even for experienced neuroradiologists7–9.
The non-invasive characterization of brain tumours on MRI is an active subject of study10,11 that has gained momentum with the surge of machine learning techniques applied to imaging data. However, most studies to date have focused on identifying anatomical MRI sequences that differentiate between two specific tumour types10–14, thus limiting the generalizability and clinical utility of this approach. Dynamic susceptibility contrast perfusion-weighted imaging (DSC-PWI) is a quantitative MRI technique that enables the visualisation of vascular characteristics including vascular density and permeability, thereby providing useful information for differential diagnosis10,15,16. DSC-PWI consists of a temporal T2*-weighted acquisition during the administration of a vascular contrast bolus. The contrast agent causes an initial decrease in the T2*-weighted signal intensity, followed by the signal recovery during washout. In DSC-PWI every voxel in the image yields a unique time-intensity curve (TIC) that describes the temporal evolution of the T2*-weighted signal intensity, and reflects local tissue vascular properties.
The standard approach to analyse TICs is to derive metrics such as the relative cerebral blood volume (rCBV) and the percentage of signal recovery (PSR). The rCBV relates to the tumour vascular density with respect to normal tissue and the PSR reflects the vascular permeability17. Both parameters remain the main focus of DSC-PWI analyses for tasks like tumour grade stratification, differentiation status and treatment response10,18. However, the performance of these parameters differs greatly among diverse clinically-used DSC-PWI protocols17,19,20, which limits its use in routine clinical practice. This issue is exacerbated by a lack of standardised DSC-PWI workflows including contrast preload settings, imaging parameters, leakage correction and patient-specific conditions, all of which pose additional challenges to the generalizability of the technique and the establishment of reference rCBV/PSR values.
Voxel-by-voxel analyses of the full TICs can overcome these limitations and unlock the potential of DSC-PWI as a tool for differential diagnosis among the most common brain malignancies (GBM, brain metastasis and PCNSL). To test this, we developed and validated an innovative, comprehensive framework for differential diagnosis of GBM, brain metastasis and PCNSL, taking advantage of the full TIC DSC-PWI data. The developed Brain Enhancing Region Radiological analYsis (BERRY) app provides voxel-by-voxel signatures of tumour type and is based on training 1D deep convolutional neural networks (CNNs) with only a small number of pilot scans for a given DSC-PWI protocol. In the current study, we demonstrate the feasibility and accuracy of the method and show its superior performance compared to classifiers based on conventional rCBV and PSR metrics. We further designed a user-friendly interface to evaluate the potential of BERRY to minimise the use of invasive brain biopsies, and guide selection of the best treatment strategies in clinical practice.
Cohort and clinical characteristics
In this multi-center, retrospective study we analysed MRI data from patients with biopsy-confirmed GBM, brain metastasis or PCNSL. A total of 568 patients from three institutions (Bellvitge University Hospital, Spain; UC San Diego Health Center, USA; HT Medica Jaen, Spain) were included in the study. Eligibility criteria included: i) histologically confirmed diagnosis of GBM, brain metastasis or PCNSL, ii) diagnostic MR scan on 1.5T or 3T including DSC-PWI and contrast-enhanced T1-weighted imaging (CE-T1WI) acquired prior to any oncological treatment and iii) a minimum of 10 mm of diameter of enhancing tumour in the CE-T1WI. Four hundred and forty patients (45 for PCNSL, 95 for metastasis and 300 for GBM) diagnosed from 2007 to 2020 at Bellvitge University Hospital (Spain) with MRI available were included in the development cohort after image quality inspection and exclusion (Fig. 1a). Additional independent cohorts were included and processed for external validation: a) 80 patients from UC San Diego Health Center (USA) and HT Medica Jaen (Spain), b) 25 patients from Bellvitge University Hospital (Spain), acquired at magnet strength of 3T and c) 23 patients from the IvyGAP21 open database of patients with GBM and MRI scan with pre-bolus contrast administration, to account for the effect of different DSC-PWI protocols in the tool performance. Further information about the study cohorts and classification results can be found in Appendix A. Patient demographics and clinical characteristics per groups are shown in Supplementary Table S1. No statistically significant differences (p > 0.05) in terms of age and sex were observed between the three tumour types.
Development Of A Cnn For Brain Tumour Classification
We trained our CNN classifier on a development cohort where patients were randomly split into training and test sets. For the training set, we included 20 patients with PCNSL and 20 non-PCNSL (10 with GBM and 10 with metastasis). This provides a comparable number of voxels for each tumour type and each binary classification (i.e., PCSNL vs non-PCNSL; GBM vs metastasis for the non-PCNSL cases). The test set consisted of 25 patients with PCNSL, 85 with metastasis and 290 with GBM (Fig. 1). Approximately 50,000 TICs from voxels of the enhancing region in the training group were used to train the classifier. Each TIC corresponds to a specific spatial voxel of the enhancing tumour.
Berry Outperforms Standard Classifiers For Brain Tumour Diagnosis
Following a hierarchical classification approach, our CNN method, BERRY, successfully achieved three-way tumoral classification, outperforming the traditional perfusion metrics (i.e., rCBV and PSR) and standing out from simpler binary classifiers. Specifically, for the task of PCNSL diagnosis, BERRY achieved superior performance with an accuracy of 0.94 (CI: 0.93–0.94); while mean rCBV and mean PSR classified patients with accuracies of 0.72 (CI: 0.70–0.74) and 0.84 (CI: 0.83–0.85), respectively. In a second step, patients not classified as PCNSL were categorised as GBM or brain metastasis. BERRY differentiated GBM from metastases with an accuracy of 0.81 (CI: 0.79–0.82). By contrast, the performance of standard DSC-derived metrics was markedly lower: rCBV classification achieved an accuracy of 0.69 (CI: 0.67–0.71) and mean PSR of 0.65 (CI: 0.63–0.67). In Fig. 2 the area under the receiver-operating characteristic (ROC) curves of the binary classifiers and 3-way average ROC curves are shown for both the BERRY classifier and conventional rCBV/PSR. Lastly, we mimicked a real-world clinical scenario in which our diagnostic support system is confronted with a priori agnostic brain lesions comprised by the three most common conditions, in this case represented by our blinded test dataset. In this setting, BERRY achieved an accuracy of 0.78 (CI: 0.76–0.79), which is substantially better than the three-way accuracy achieved using mean rCBV (0.59, CI: 0.57–0.60) and mean PSR (0.55, CI: 0.53–0.56). Furthermore, the combination of rCBV and PSR into a logistic regression model also yielded poor performance (Supplementary Table S3). Additional sensitivity and specificity values can be found in Supplementary Table S2. These data underscore the potential of BERRY for differentiating among the three most common clinical diagnostic challenges in patients with enhancing brain lesions.
Voxel-wise Explainable Representation Of The Cnn Decision Process
BERRY provides spatial probability maps of tumour classification, which are then used to obtain a voxel proportion and a patient classification label. In Fig. 2A we present three examples per tumour type of the voxel-wise probability maps according to the BERRY classifier. The probability maps are shown overlaid onto the CE-T1W MRI for anatomical references. Overall, the tumour type probability maps are smooth and identify the tumour type with high confidence in most voxels, even when intra-tumour signal heterogeneity is seen in the contrast enhanced T1W scan. Voxels exhibiting a high probability of belonging to the incorrect tumour class tend to be located either in the boundary of the enhancing area, or around necrotic intra-tumoral spots. This potentially reflects partial volume (i.e., inclusion of signal from tumour and non-tumour areas within a voxel), or intra-tumoral heterogeneity. Voxel-wise classification allows for computing spatial distribution of the predictions within a tumour yielding relevant information about the prediction consistency and also about potential tumour heterogeneity, useful to guide interventions and for tumour spatial characterization.
Visual Interpretation Of The Cnn Classification
We further sought to implement Class Activation Mapping (CAM) to provide visual explanation of the BERRY classification network. The ScoreCAM22 method yields a normalised score of the contribution of every input to the final classification of a CNN. This allows us to identify the most discriminative timepoints for TIC differentiation. ScoreCAM spatial maps were obtained for each binary classification (Fig. 3A). The CNN mostly focuses on the bolus passage to classify the central tumour region (middle row for PCNSL vs non-PCNSL and lower row for GBM vs metastasis differentiation in Fig. 3A). In contrast, the bolus passage seems less important for some voxels in surrounding regions. This suggests that the CNN effectively considered the bolus passage as a discerning characteristic, but also that it provides additional tissue perfusion differences compared to the raw DSC-PWI signal (top row in Fig. 3A).
The average ScoreCAM values per tumour type and per CNN classifier can be found in Fig. 3B (upper row for PCNSL vs non-PCNSL and lower row for GBM vs metastasis differentiation). Overall, the sharper signal changes of the TICs, i.e., steep slopes during contrast arrival and washout, have a higher contribution score. This is especially true for GBM, with greater differences in these timepoints with respect to the other two tumour types (average TICs shown in black in Fig. 3B). For PCNSL and metastasis, the last part of the signal is also considered important, which can be expected given the overall higher signal magnitude reached in these cases. Importantly, applying 1D CNNs over TIC signals allows to analyse the local changes of the signal over time. In this regard, methods that only consider the signal magnitude of specific timepoints, such as PSR, or a derived measurement like rCBV, may overlook local TIC changes occurring over time that reflect specific physiological traits of the tumour.
A User-friendly Berry App
The BERRY app was successfully implemented at the participating institutions for validating the tool in external cohorts, as illustrated in Fig. 4. The tool requires approximately two minutes to process a new case and provides a classification outcome, in the form of i) voxel-wise tumour type probability maps, and ii) patient-wise tumour type. In addition, it shows the average TIC for the enhancing tumour and white matter, as well as a visualisation of the segmentation for the user to safely check the process. The mask can be automatically segmented from the enhancing tumour by BERRY or it can be provided by the user. The BERRY app provides a classification label with balanced sensitivity and specificity (Youden’s index) by default, but a given clinical scenario may require a different classification threshold. To that end, sensitivities and specificities for every threshold are displayed, and the default settings can be changed.
We present a novel voxel-wise method for analysing perfusion scans with CNNs and improve brain cancer diagnosis, built upon prior DSC-PWI signal normalization23. By applying this method, we were able to surpass the performance of previous models for non-invasive differential diagnosis of the most frequent malignant brain tumours (i.e., GBM, metastasis and PCNSL, representing up to 70% of all malignant tumours in the brain6), which is critical to define an optimal treatment approach.
Our deep learning framework takes advantage of the large amount of information provided by the thousands of voxel-wise TICs available in each individual DSC-PWI scan24, and achieves optimal performance through training with a limited number of scans from a few patients at fixed DSC protocol (on the order of 30–40 cases). Our approach is particularly appealing for medical imaging applications, where the design of robust deep learning methods is challenged by the limited number of scans available. Additionally, our method distinguishes between tumour types in a three-way classification task. This can be of particular relevance as a support tool for differential diagnosis in clinical practice, and is a considerable step forward as compared to current literature, which is dominated by binary classification studies10–14. This is particularly important when PCNSL is considered among the potential diagnoses. Corticosteroids are usually the first treatment of choice to reduce the neurological symptoms secondary to oedema in patients with malignant brain tumours. However, early stereotactic biopsy prior to corticosteroids administration is mandatory when a brain PCNSL is suspected by imaging, as medication with steroids can alter the histological pattern of PCNSL5. Moreover, PCNSL is highly sensitive to chemoradiotherapy instead of resection, which is contraindicated, as opposed to GBM or metastasis. Therefore, a reliable characterization of the tumour type by imaging is critical to devise the appropriate management of patients.
The BERRY app provides voxel-wise tumour type probability maps, which are then used to obtain a voxel proportion and a patient classification label. The default Youden’s index (trade-off between sensitivity and specificity) can be changed to the needs of different clinical scenarios. For instance, some clinical scenarios may require a very high specificity for suspected GBM and metastases with respect to PCNSL and, if all evidence supports it, an additional intervention for a biopsy could be prevented. Therefore, the voxel proportion can be adjusted in the app to match the user’s needs.
The presented method successfully achieved three-way tumoral classification, outperforming the traditional perfusion metrics and standing out from simpler binary classifiers. When tested, our method performed with accuracies of 0.94 for PCNSL identification, 0.81 for differentiation of GBM from metastasis and 0.78 for three-way classification.
Of note, the DSC protocol used for model development did not include contrast preload. Contrast preload is a common approach described in the literature to achieve a better estimate of the rCBV20. However, preload can be undesirable for a number of reasons. Firstly, it delivers a higher contrast dose to the patient. Secondly, it can introduce potential variability sources, affecting the TIC signal morphology. As countermeasures, leakage correction and acquisition parameters that minimise T1 effect, such as low flip angle, have been shown to effectively yield reliable rCBV estimates without preload20,25. In this study, rCBV was estimated with leakage correction, obtaining comparable results to those of PSR. The combination of both rCBV and PSR in logistic regression was explored for completeness, but it did not improve the results of the individual parameters.
A key feature of our CNN approach is the computation of voxel-wise spatial representations of perfusion curve characteristics, in the form of maps describing the probability of a voxel to belong to a specific tumour type. Such spatial probability maps provide an explainable representation of the CNN decision process and may enable further studies of intra-tumour heterogeneity, making them an appealing tool for integrative multi-omics research and also of potential clinical interest to plan surgical procedures. To our knowledge, this is one of the first studies applying deep learning to voxel-wise DSC TICs in neuro-oncological applications. A recent study26 used a deep autoencoder to derive a set of five descriptors of TICs that could differentiate between pairs of tumour types. However, the reconstructed timepoints from such a minimal set of descriptors, in contrast to the original signal, produce a smooth TIC morphology, which may be omitting relevant details for diagnostic applications.
We acknowledge that our model trained with DSC-PWI data without preload from 1.5T scanners can be a potential limitation. As discussed, preload can change the TIC morphology and it is possible that the performance of our model is hindered when deployed in preloaded data. Nevertheless, tests on all eligible external 3T scans with contrast preload of 23 patients with GBM from the IvyGAP21 dataset yielded 18 cases correctly classified as GBM (0.78 accuracy). Of note, internal centre tests on 3T scans (n = 25) resulted in 0.72 accuracy, showing an overall agreement in performance with mixed scans from external centres (n = 80) reaching 0.71 of accuracy. The results show that, albeit training on a small cohort, our method notably generalises in external populations, in contrast to reported conventional metrics. Importantly, the proposed method could accommodate different DSC-PWI protocols by training the models with just a few new imaging samples.
Regarding the segmented regions of interest, we used the automatic segmented masks revised by an experienced neuroradiologist as reference, but the variability in the segmentations from different neuroradiologists is to be explored. We believe that future segmentation methods could make the manual input minimal and it could be integrated in the proposed pipeline. To account for that in the online tool, the user can check the automatic segmentation obtained from thresholding and they can provide their own if needed. In the future, the algorithm could be adapted to include rich multi-parametric MRI protocols that include additional contrasts (e.g., diffusion MRI27), as these may provide orthogonal information on tumour microstructure that could improve the classification performance even further. Nevertheless, we proposed here a user-friendly tool that can be applied with just two MRI sequences (CE-T1WI for determining the area of interest and DSC-PWI for classification). BERRY allows for classifying the three most common enhancing brain tumours (i.e., GBM, metastases and PCNSL). Furthermore, the developed framework will allow expanding its use by training the model with a few cases of other less common tumours such as anaplastic astrocytoma and extra-axial tumours such as meningioma in the future.
In conclusion, the presented CNN framework for three-class brain tumour classification based on voxel-wise DSC-PWI signal analysis is feasible and outperforms classifiers built on conventional rCBV and PSR metrics. The method can be trained using a limited number of scans, which most centres are likely to have available, with notable generalisation to external data. Additionally, it provides voxel-wise maps of tumour type signatures that could be useful to visualise the CNN classification process, and for tumour spatial characterization. As a way to make this tool more accessible and eventually make an impact in clinical practice, the proposed method has been implemented on the user-friendly BERRY application that is made freely accessible at https://berry-app.vhio.org, in order to enhance study reproducibility and accelerate its adoption in future clinical studies.
The research ethics committee of Bellvitge University Hospital (Barcelona, Spain) approved the study and informed consent was waived. The confidential data from patients were anonymized and protected in accordance with national and European regulations.
Statistical analysis of patient characteristic distribution
Statistical tests were performed to compare: (i) patient age distribution between training and test sets of each tumour type (Welch’s t-test), as well as among tumour types (one-way ANOVA); (ii) patient sex distribution between training and test sets of each tumour type (Fisher’s exact test), as well as among tumour types (Chi-square test). Relations shown in Supplementary Table S1.
Data pre-processing
The MRI scans were performed in 1.5T Philips scanners (219 on Ingenia and 221 on Intera). The DSC-PWI acquisition parameters were: temporal sampling of 1.26-1.93 seconds, 40-60 timepoints, acquired during a bolus administration of gadolinium contrast agent (gadobutrol 0.1 mmol/kg) without contrast preload (more details in Appendix B of the Supplementary Material).
The DSC-PWI dynamic sequence was motion-corrected by rigid registration of all DSC volumes. Bias field correction28 was applied to the CE-T1WI, which was rigidly registered to the DSC-PWI. Brain region masking was obtained on the CE-T1WI using a hierarchical approach29. Segmentations of the enhancing tumour and contralateral normal-appearing white matter were first obtained by simple thresholding and afterwards revised by an experienced neuroradiologist (APE). The TICs reflecting the bolus passage in every voxel of the DSC-PWI sequence were extracted for the enhancing tumour and normal white matter regions. According to a previously presented normalisation method23, TICs from the enhancing tumour were normalised to the white matter. The minimum peak point of the TICs was retrieved, the curves were aligned to this point and TICs with points within the average plus and minus standard deviation were used for training the CNNs. Slicer30 (www.slicer.org) and Python 3.8 were used for segmentation, processing, training, inference and statistical tests.
CNN architecture
A CNN was designed with three 1D convolutional layers with kernel sizes [3,5,7], followed by a 10% dropout layer and max pooling layer (pool size 2), then concatenated into a dense layer with 100 nodes of rectified linear units and a final binary output layer with softmax activation, cross-entropy loss function and Adam optimizer. The CNN was built using Tensorflow v2 with Keras frontend. The CNN classifier receives a given TIC as input and outputs a binary probability.
Classification scheme
By applying the CNN classifier voxel-by-voxel, a probability map is obtained over the enhancing tumour region. These voxel-wise probabilities are then converted into a patient-wise classification as:
Above, 
is the number of voxels with a probability higher than 0.9 for one tumour type and 
is the number of voxels with a probability higher than 0.9 for the second tumour type on the binary classifier. The tumour type of each patient was inferred by applying Youden’s index (highest sum of specificity and sensitivity in the training set) to the voxel proportion above. In practice, a 3-class classifier differentiating between PCNSL, GMB and metastases was implemented by concatenating two 2-class classifiers. The first classifier distinguishes PCNSL from non-PCNSL cases, while the second classifier differentiates the non-PCNSL cases into GBM or metastasis (Fig. 1).
Classification performance and interpretation
Classification accuracy, sensitivity and specificity were obtained for 100 groups of 25 randomly-selected patients from the test cohort in order to obtain average classification performance and 95% confidence intervals (CI) for all classifiers, as reported in Supplementary Table S2.
In addition, the area under the ROC curve was obtained for binary classifications, which shows the trade-off between sensitivity and specificity of different classification thresholds (Fig. 2B). The thresholds were set by Youden’s index as described above, but the app allows users to change them to meet different clinical needs, as discussed. Average ROC curves were obtained for the 3-class problem (Fig. 2C).
Current standard metrics of DSC-PWI analyses were obtained to compare against our voxel-trained CNN. Mean PSR and mean rCBV were computed with Slicer, for which further details can be found in Supplementary Appendix C. Classification metrics and ROC curves were obtained for PSR and rCBV applying the same classification structure used for the CNN-based approach.
CNN interpretation
The CNN provides a tumour-type probability value from each TIC found in each voxel. The map inherently informs about the decision process of the CNN classifier towards one tumour type or another and, more importantly, about the confidence of the classification in spatial regions.
To further explore the features that the CNN associated with each tumour type, down to the individual timepoints of the DSC-PWI TIC signal, we applied a score-weighted visual explanation for CNNs (ScoreCAM22). On Fig. 3A, the importance score was scaled to sum 1 over all timepoints in every voxel, in order to see the spatial relative importance. On Fig 3B, the average importance score is shown in each timepoint for each tumour type, with the average tumour TIC from the training data overlayed in black, in order to see the temporal differences.
Development of the online app BERRY
The processing and classification pipeline was bundled into a Docker image which can run as a standalone application in any system (Fig. 4). For demonstrative purposes, the BERRY app is also available on the VHIO server through a web interface, so that it is accessible from anywhere using an internet connection. The user can input their anonymized DSC-PWI and CE-T1WI scans in raw DICOM, Nifti or NRRD formats and, optionally, their own segmentations. When the study is processed, the tool shows the average TIC, the result of the classification and the spatial probability map. The online tool can be accessed at https://berry-app.vhio.org for research purposes.
Data availability
Four out of five datasets used in this study, from the 4 participating sites, are not publicly available due to their containing information that could compromise the privacy of research participants. The IvyGAP21,31 dataset used as one of the validation cohorts is publicly available at The Cancer Imaging Archive32 (https://doi.org/10.7937/K9/TCIA.2016.XLwaN6nL), along with reference segmentations33,34 (https://doi.org/10.7937/9j41-7d44).
Code availability
The code integrating the app processing and classification pipeline can be publicly accessed at https://github.com/radiomicsvhio/berry-app. We relied on the open-source software dcm2niix (https://github.com/rordenlab/dcm2niix/) for DICOM conversion and Slicer30 (www.slicer.org) for image annotations and computing. The BERRY online app can be accessed at https://berry-app.vhio.org.
Equal contribution:
*AGR, APE and FG are co-first authors with equal contributions.
- Young RM, Jamshidi A, Davis G, Sherman JH. Current trends in the surgical management and treatment of adult glioblastoma. Ann Transl Med. 2015; 3(9):121.
- Hatiboglu MA, Wildrick DM, Sawaya R. The role of surgical resection in patients with brain metastases. Ecancermedicalscience. 2013; 7:308.
- Hoang-Xuan K, Bessell E, Bromberg J, et al. Diagnosis and treatment of primary CNS lymphoma in immunocompetent patients: guidelines from the European Association for Neuro-Oncology. The Lancet Oncology. 2015; 16(7):e322-e332.
- Dammers R, Haitsma IK, Schouten JW, Kros JM, Avezaat CJ, Vincent AJ. Safety and efficacy of frameless and frame-based intracranial biopsy techniques. Acta Neurochir (Wien). 2008; 150(1):23–29.
- Chiavazza C, Pellerino A, Ferrio F, Cistaro A, Soffietti R, Ruda R. Primary CNS Lymphomas: Challenges in Diagnosis and Monitoring. Biomed Res Int. 2018; 2018:3606970.
- Miller KD, Ostrom QT, Kruchko C, et al. Brain and other central nervous system tumor statistics, 2021. CA Cancer J Clin. 2021; 71(5):381–406.
- Leung D, Han X, Mikkelsen T, Nabors LB. Role of MRI in primary brain tumor evaluation. J Natl Compr Canc Netw. 2014; 12(11):1561–1568.
- Arita K, Miwa M, Bohara M, Moinuddin FM, Kamimura K, Yoshimoto K. Precision of preoperative diagnosis in patients with brain tumor - A prospective study based on "top three list" of differential diagnosis for 1061 patients. Surg Neurol Int. 2020; 11:55.
- Chakravorty A, Steel T, Chaganti J. Accuracy of percentage of signal intensity recovery and relative cerebral blood volume derived from dynamic susceptibility-weighted, contrast-enhanced MRI in the preoperative diagnosis of cerebral tumours. Neuroradiol J. 2015; 28(6):574–583.
- Cha S, Lupo JM, Chen MH, 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. 2007; 28(6):1078–1084.
- Fordham AJ, Hacherl CC, Patel N, et al. Differentiating Glioblastomas from Solitary Brain Metastases: An Update on the Current Literature of Advanced Imaging Modalities. Cancers (Basel). 2021; 13(12).
- Artzi M, Bressler I, Ben Bashat D. Differentiation between glioblastoma, brain metastasis and subtypes using radiomics analysis. J Magn Reson Imaging. 2019; 50(2):519–528.
- Bae S, An C, Ahn SS, et al. Robust performance of deep learning for distinguishing glioblastoma from single brain metastasis using radiomic features: model development and validation. Sci Rep. 2020; 10(1):12110.
- Qian Z, Li Y, Wang Y, et al. Differentiation of glioblastoma from solitary brain metastases using radiomic machine-learning classifiers. Cancer Lett. 2019; 451:128–135.
- Neska-Matuszewska M, Bladowska J, Sasiadek M, Zimny A. Differentiation of glioblastoma multiforme, metastases and primary central nervous system lymphomas using multiparametric perfusion and diffusion MR imaging of a tumor core and a peritumoral zone-Searching for a practical approach. PLoS One. 2018; 13(1):e0191341.
- Lee MD, Baird GL, Bell LC, Quarles CC, Boxerman JL. 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. 2019; 40(9):1445–1450.
- Bell LC, Hu LS, Stokes AM, McGee SC, Baxter LC, Quarles CC. Characterizing the Influence of Preload Dosing on Percent Signal Recovery (PSR) and Cerebral Blood Volume (CBV) Measurements in a Patient Population With High-Grade Glioma Using Dynamic Susceptibility Contrast MRI. Tomography. 2017; 3(2):89–95.
- Bell LC, Semmineh N, An H, et al. Evaluating the Use of rCBV as a Tumor Grade and Treatment Response Classifier Across NCI Quantitative Imaging Network Sites: Part II of the DSC-MRI Digital Reference Object (DRO) Challenge. Tomography. 2020; 6(2):203–208.
- Boxerman JL, Paulson ES, Prah MA, Schmainda KM. 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. 2013; 34(7):1364–1369.
- Paulson ES, Schmainda KM. Comparison of dynamic susceptibility-weighted contrast-enhanced MR methods: recommendations for measuring relative cerebral blood volume in brain tumors. Radiology. 2008; 249(2):601–613.
- Shah N, Feng X, Lankerovich M, Puchalski RB, Keogh B. Data from Ivy Glioblastoma Atlas Project (IvyGAP): The Cancer Imaging Archive; 2016.
- Wang H, Wang Z, Du M, et al. Score-CAM: Score-Weighted Visual Explanations for Convolutional Neural Networks. Paper presented at: 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW)2020.
- Pons-Escoda A, Garcia-Ruiz A, Naval-Baudin P, et al. Presurgical Identification of Primary Central Nervous System Lymphoma with Normalized Time-Intensity Curve: A Pilot Study of a New Method to Analyze DSC-PWI. AJNR Am J Neuroradiol. 2020; 41(10):1816–1824.
- Grussu F, Blumberg SB, Battiston M, et al. Feasibility of Data-Driven, Model-Free Quantitative MRI Protocol Design: Application to Brain and Prostate Diffusion-Relaxation Imaging. Frontiers in Physics. 2021; 9.
- Schmainda KM, Prah MA, Hu LS, 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. 2019; 40(4):626–633.
- Park JE, Kim HS, 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. 2020; 10(1):21485.
- Nilsson M, Englund E, Szczepankiewicz F, van Westen D, Sundgren PC. Imaging brain tumour microstructure. Neuroimage. 2018; 182:232–250.
- Tustison NJ, Avants BB, Cook PA, et al. N4ITK: improved N3 bias correction. IEEE Trans Med Imaging. 2010; 29(6):1310–1320.
- Pohl KM, Bouix S, Nakamura M, et al. A hierarchical algorithm for MR brain image parcellation. IEEE Trans Med Imaging. 2007; 26(9):1201–1212.
- Fedorov A, Beichel R, Kalpathy-Cramer J, et al. 3D Slicer as an image computing platform for the Quantitative Imaging Network. Magn Reson Imaging. 2012; 30(9):1323–1341.
- Puchalski RB, Shah N, Miller J, et al. An anatomic transcriptional atlas of human glioblastoma. Science. 2018; 360(6389):660–663.
- Clark K, Vendt B, Smith K, et al. The Cancer Imaging Archive (TCIA): maintaining and operating a public information repository. J Digit Imaging. 2013; 26(6):1045–1057.
- Data from the Multi-Institutional Paired Expert Segmentations and Radiomic Features of the Ivy GAP Dataset. The Cancer Imaging Archive (TCIA); 2020. https://wiki.cancerimagingarchive.net/pages/viewpage.action?pageId=70222827.
- Pati S, Verma R, Akbari H, et al. Reproducibility analysis of multi-institutional paired expert annotations and radiomic features of the Ivy Glioblastoma Atlas Project (Ivy GAP) dataset. Med Phys. 2020; 47(12):6039–6052.
There is NO Competing Interest.