A novel fully automated MRI-based deep-learning method for classification of 1p/19q co-deletion status in brain gliomas

Artificial intelligence in neuro-oncology

Artificial intelligence (AI) describes the application of computer algorithms to the solution of problems that have traditionally required human intelligence. Although formal work in AI has been slowly advancing for almost 70 years, developments in the last decade, and particularly in the last year, have led to an explosion of AI applications in multiple fields. Neuro-oncology has not escaped this trend. Given the expected integration of AI-based methods to neuro-oncology practice over the coming years, we set to provide an overview of existing technologies as they are applied to the neuropathology and neuroradiology of brain tumors. We highlight current benefits and limitations of these technologies and offer recommendations on how to appraise novel AI-tools as they undergo consideration for integration into clinical workflows.

Uncertainty-Aware Deep Learning Classification of Adamantinomatous Craniopharyngioma from Preoperative MRI

Diagnosis of adamantinomatous craniopharyngioma (ACP) is predominantly determined through invasive pathological examination of a neurosurgical biopsy specimen. Clinical experts can distinguish ACP from Magnetic Resonance Imaging (MRI) with an accuracy of 86%, and 9% of ACP cases are diagnosed this way. Classification using deep learning (DL) provides a solution to support a non-invasive diagnosis of ACP through neuroimaging, but it is still limited in implementation, a major reason being the lack of predictive uncertainty representation. We trained and tested a DL classifier on preoperative MRI from 86 suprasellar tumor patients across multiple institutions. We then applied a Bayesian DL approach to calibrate our previously published ACP classifier, extending beyond point-estimate predictions to predictive distributions. Our original classifier outperforms random forest and XGBoost models in classifying ACP. The calibrated classifier underperformed our previously published results, indicating that the original model was overfit. Mean values of the predictive distributions were not informative regarding model uncertainty. However, the variance of predictive distributions was indicative of predictive uncertainty. We developed an algorithm to incorporate predicted values and the associated uncertainty to create a classification abstention mechanism. Our model accuracy improved from 80.8% to 95.5%, with a 34.2% abstention rate. We demonstrated that calibration of DL models can be used to estimate predictive uncertainty, which may enable clinical translation of artificial intelligence to support non-invasive diagnosis of brain tumors in the future.

Assessment of artificial intelligence (AI) reporting methodology in glioma MRI studies using the Checklist for AI in Medical Imaging (CLAIM)

PURPOSE

The Checklist for Artificial Intelligence in Medical Imaging (CLAIM) is a recently released guideline designed for the optimal reporting methodology of artificial intelligence (AI) studies. Gliomas are the most common form of primary malignant brain tumour and numerous outcomes derived from AI algorithms such as grading, survival, treatment-related effects and molecular status have been reported. The aim of the study is to evaluate the AI reporting methodology for outcomes relating to gliomas in magnetic resonance imaging (MRI) using the CLAIM criteria.

METHODS

A literature search was performed on three databases pertaining to AI augmentation of glioma MRI, published between the start of 2018 and the end of 2021 RESULTS: A total of 4308 articles were identified and 138 articles remained after screening. These articles were categorised into four main AI tasks: grading (n= 44), predicting molecular status (n= 50), predicting survival (n= 25) and distinguishing true tumour progression from treatment-related effects (n= 10). The average CLAIM score was 20/42 (range: 10-31). Studies most consistently reported the scientific background and clinical role of their AI approach. Areas of improvement were identified in the reporting of data collection, data management, ground truth and validation of AI performance.

CONCLUSION

AI may be a means of producing high-accuracy results for certain tasks in glioma MRI; however, there remain issues with reporting quality. AI reporting guidelines may aid in a more reproducible and standardised approach to reporting and will aid in clinical integration.

Swin Transformer Improves the IDH Mutation Status Prediction of Gliomas Free of MRI-Based Tumor Segmentation

Background: Deep learning (DL) could predict isocitrate dehydrogenase (IDH) mutation status from MRIs. Yet, previous work focused on CNNs with refined tumor segmentation. To bridge the gap, this study aimed to evaluate the feasibility of developing a Transformer-based network to predict the IDH mutation status free of refined tumor segmentation. Methods: A total of 493 glioma patients were recruited from two independent institutions for model development (TCIA; N = 259) and external test (AHXZ; N = 234). IDH mutation status was predicted directly from T2 images with a Swin Transformer and conventional ResNet. Furthermore, to investigate the necessity of refined tumor segmentation, seven strategies for the model input image were explored: (i) whole tumor slice; (ii-iii) tumor mask and/or not edema; (iv-vii) tumor bounding box of 0.8, 1.0, 1.2, 1.5 times. Performance comparison was made among the networks of different architectures along with different image input strategies, using area under the curve (AUC) and accuracy (ACC). Finally, to further boost the performance, a hybrid model was built by incorporating the images with clinical features. Results: With the seven proposed input strategies, seven Swin Transformer models and seven ResNet models were built, respectively. Based on the seven Swin Transformer models, an averaged AUC of 0.965 (internal test) and 0.842 (external test) were achieved, outperforming 0.922 and 0.805 resulting from the seven ResNet models, respectively. When a bounding box of 1.0 times was used, Swin Transformer (AUC = 0.868, ACC = 80.7%), achieved the best results against the one that used tumor segmentation (Tumor + Edema, AUC = 0.862, ACC = 78.5%). The hybrid model that integrated age and location features into images yielded improved performance (AUC = 0.878, Accuracy = 82.0%) over the model that used images only. Conclusions: Swin Transformer outperforms the CNN-based ResNet in IDH prediction. Using bounding box input images benefits the DL networks in IDH prediction and makes the IDH prediction free of refined glioma segmentation feasible.

Prediction of glioma-subtypes: comparison of performance on a DL classifier using bounding box areas versus annotated tumors

Background

For brain tumors, identifying the molecular subtypes from magnetic resonance imaging (MRI) is desirable, but remains a challenging task. Recent machine learning and deep learning (DL) approaches may help the classification/prediction of tumor subtypes through MRIs. However, most of these methods require annotated data with ground truth (GT) tumor areas manually drawn by medical experts. The manual annotation is a time consuming process with high demand on medical personnel. As an alternative automatic segmentation is often used. However, it does not guarantee the quality and could lead to improper or failed segmented boundaries due to differences in MRI acquisition parameters across imaging centers, as segmentation is an ill-defined problem. Analogous to visual object tracking and classification, this paper shifts the paradigm by training a classifier using tumor bounding box areas in MR images. The aim of our study is to see whether it is possible to replace GT tumor areas by tumor bounding box areas (e.g. ellipse shaped boxes) for classification without a significant drop in performance.

Method

In patients with diffuse gliomas, training a deep learning classifier for subtype prediction by employing tumor regions of interest (ROIs) using ellipse bounding box versus manual annotated data. Experiments were conducted on two datasets (US and TCGA) consisting of multi-modality MRI scans where the US dataset contained patients with diffuse low-grade gliomas (dLGG) exclusively.

Results

Prediction rates were obtained on 2 test datasets: 69.86% for 1p/19q codeletion status on US dataset and 79.50% for IDH mutation/wild-type on TCGA dataset. Comparisons with that of using annotated GT tumor data for training showed an average of 3.0% degradation (2.92% for 1p/19q codeletion status and 3.23% for IDH genotype).

Conclusion

Using tumor ROIs, i.e., ellipse bounding box tumor areas to replace annotated GT tumor areas for training a deep learning scheme, cause only a modest decline in performance in terms of subtype prediction. With more data that can be made available, this may be a reasonable trade-off where decline in performance may be counteracted with more data.

Magnetic Resonance Imaging Correlates of Immune Microenvironment in Glioblastoma

Background

Glioblastoma (GBM) is the most commonly occurring primary malignant brain tumor, and it carries a dismal prognosis. Focusing on the tumor microenvironment may provide new insights into pathogenesis, but no clinical tools are available to do this. We hypothesized that the infiltration of different leukocyte populations in the tumoral and peritumoral brain tissues may be measured by magnetic resonance imaging (MRI).

Methods

Pre-operative MRI was combined with immune phenotyping of intraoperative tumor tissue based on flow cytometry of myeloid cell populations that are associated with immune suppression, namely, microglia and bone marrow-derived macrophages (BMDM). These cell populations were measured from the central and marginal areas of the lesion identified intraoperatively with 5-aminolevulinic acid-guided surgery. MRI features (volume, mean and standard deviation of signal intensity, and fractality) were derived from all MR sequences (T1w, Gd+ T1w, T2w, FLAIR) and ADC MR maps and from different tumor areas (contrast- and non-contrast-enhancing tumor, necrosis, and edema). The principal components of MRI features were correlated with different myeloid cell populations by Pearson's correlation.

Results

We analyzed 126 samples from 62 GBM patients. The ratio between BMDM and microglia decreases significantly from the central core to the periphery. Several MRI-derived principal components were significantly correlated (p <0.05, r range: [-0.29, -0.41]) with the BMDM/microglia ratio collected in the central part of the tumor.

Conclusions

We report a significant correlation between structural MRI clinical imaging and the ratio of recruited vs. resident macrophages with different immunomodulatory activities. MRI features may represent a novel tool for investigating the microenvironment of GBM.

Combining Radiology and Pathology for Automatic Glioma Classification

Subtype classification is critical in the treatment of gliomas because different subtypes lead to different treatment options and postoperative care. Although many radiological- or histological-based glioma classification algorithms have been developed, most of them focus on single-modality data. In this paper, we propose an innovative two-stage model to classify gliomas into three subtypes (i.e., glioblastoma, oligodendroglioma, and astrocytoma) based on radiology and histology data. In the first stage, our model classifies each image as having glioblastoma or not. Based on the obtained non-glioblastoma images, the second stage aims to accurately distinguish astrocytoma and oligodendroglioma. The radiological images and histological images pass through the two-stage design with 3D and 2D models, respectively. Then, an ensemble classification network is designed to automatically integrate the features of the two modalities. We have verified our method by participating in the MICCAI 2020 CPM-RadPath Challenge and won 1st place. Our proposed model achieves high performance on the validation set with a balanced accuracy of 0.889, Cohen’s Kappa of 0.903, and an F1-score of 0.943. Our model could advance multimodal-based glioma research and provide assistance to pathologists and neurologists in diagnosing glioma subtypes. The code has been publicly available online at https://github.com/Xiyue-Wang/1st-in-MICCAI2020-CPM.

Multimodal MRI Image Decision Fusion-Based Network for Glioma Classification

Purpose

Glioma is the most common primary brain tumor, with varying degrees of aggressiveness and prognosis. Accurate glioma classification is very important for treatment planning and prognosis prediction. The main purpose of this study is to design a novel effective algorithm for further improving the performance of glioma subtype classification using multimodal MRI images.

Method

MRI images of four modalities for 221 glioma patients were collected from Computational Precision Medicine: Radiology-Pathology 2020 challenge, including T1, T2, T1ce, and fluid-attenuated inversion recovery (FLAIR) MRI images, to classify astrocytoma, oligodendroglioma, and glioblastoma. We proposed a multimodal MRI image decision fusion-based network for improving the glioma classification accuracy. First, the MRI images of each modality were input into a pre-trained tumor segmentation model to delineate the regions of tumor lesions. Then, the whole tumor regions were centrally clipped from original MRI images followed by max-min normalization. Subsequently, a deep learning-based network was designed based on a unified DenseNet structure, which extracts features through a series of dense blocks. After that, two fully connected layers were used to map the features into three glioma subtypes. During the training stage, we used the images of each modality after tumor segmentation to train the network to obtain its best accuracy on our testing set. During the inferring stage, a linear weighted module based on a decision fusion strategy was applied to assemble the predicted probabilities of the pre-trained models obtained in the training stage. Finally, the performance of our method was evaluated in terms of accuracy, area under the curve (AUC), sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), etc.

Results

The proposed method achieved an accuracy of 0.878, an AUC of 0.902, a sensitivity of 0.772, a specificity of 0.930, a PPV of 0.862, an NPV of 0.949, and a Cohen's Kappa of 0.773, which showed a significantly higher performance than existing state-of-the-art methods.

Conclusion

Compared with current studies, this study demonstrated the effectiveness and superiority in the overall performance of our proposed multimodal MRI image decision fusion-based network method for glioma subtype classification, which would be of enormous potential value in clinical practice.

Predicting 1p/19q co-deletion status from magnetic resonance imaging using deep learning in adult-type diffuse lower-grade gliomas: a discovery and validation study

Determination of 1p/19q co-deletion status is important for the classification, prognostication, and personalized therapy in diffuse lower-grade gliomas (LGG). We developed and validated a deep learning imaging signature (DLIS) from preoperative magnetic resonance imaging (MRI) for predicting the 1p/19q status in patients with LGG. The DLIS was constructed on a training dataset (n = 330) and validated on both an internal validation dataset (n = 123) and a public TCIA dataset (n = 102). The receiver operating characteristic (ROC) analysis and precision recall curves (PRC) were used to measure the classification performance. The area under ROC curves (AUC) of the DLIS was 0.999 for training dataset, 0.986 for validation dataset, and 0.983 for testing dataset. The F1-score of the prediction model was 0.992 for training dataset, 0.940 for validation dataset, and 0.925 for testing dataset. Our data suggests that DLIS could be used to predict the 1p/19q status from preoperative imaging in patients with LGG. The imaging-based deep learning has the potential to be a noninvasive tool predictive of molecular markers in adult diffuse gliomas.

Brain tumor IDH, 1p/19q, and MGMT molecular classification using MRI-based deep learning: an initial study on the effect of motion and motion correction

Purpose: Deep learning has shown promise for predicting the molecular profiles of gliomas using MR images. Prior to clinical implementation, ensuring robustness to real-world problems, such as patient motion, is crucial. The purpose of this study is to perform a preliminary evaluation on the effects of simulated motion artifact on glioma marker classifier performance and determine if motion correction can restore classification accuracies. Approach: T2w images and molecular information were retrieved from the TCIA and TCGA databases. Simulated motion was added in the k-space domain along the phase encoding direction. Classifier performance for IDH mutation, 1p/19q co-deletion, and MGMT methylation was assessed over the range of 0% to 100% corrupted k-space lines. Rudimentary motion correction networks were trained on the motion-corrupted images. The performance of the three glioma marker classifiers was then evaluated on the motion-corrected images. Results: Glioma marker classifier performance decreased markedly with increasing motion corruption. Applying motion correction effectively restored classification accuracy for even the most motion-corrupted images. For isocitrate dehydrogenase (IDH) classification, 99% accuracy was achieved, exceeding the original performance of the network and representing a new benchmark in non-invasive MRI-based IDH classification. Conclusions: Robust motion correction can facilitate highly accurate deep learning MRI-based molecular marker classification, rivaling invasive tissue-based characterization methods. Motion correction may be able to increase classification accuracy even in the absence of a visible artifact, representing a new strategy for boosting classifier performance.

Molecular Biology in Treatment Decision Processes-Neuro-Oncology Edition

Computational approaches including machine learning, deep learning, and artificial intelligence are growing in importance in all medical specialties as large data repositories are increasingly being optimised. Radiation oncology as a discipline is at the forefront of large-scale data acquisition and well positioned towards both the production and analysis of large-scale oncologic data with the potential for clinically driven endpoints and advancement of patient outcomes. Neuro-oncology is comprised of malignancies that often carry poor prognosis and significant neurological sequelae. The analysis of radiation therapy mediated treatment and the potential for computationally mediated analyses may lead to more precise therapy by employing large scale data. We analysed the state of the literature pertaining to large scale data, computational analysis, and the advancement of molecular biomarkers in neuro-oncology with emphasis on radiation oncology. We aimed to connect existing and evolving approaches to realistic avenues for clinical implementation focusing on low grade gliomas (LGG), high grade gliomas (HGG), management of the elderly patient with HGG, rare central nervous system tumors, craniospinal irradiation, and re-irradiation to examine how computational analysis and molecular science may synergistically drive advances in personalised radiation therapy (RT) and optimise patient outcomes.

Introduction to Deep Learning in Clinical Neuroscience

The use of deep learning (DL) is rapidly increasing in clinical neuroscience. The term denotes models with multiple sequential layers of learning algorithms, architecturally similar to neural networks of the brain. We provide examples of DL in analyzing MRI data and discuss potential applications and methodological caveats.Important aspects are data pre-processing, volumetric segmentation, and specific task-performing DL methods, such as CNNs and AEs. Additionally, GAN-expansion and domain mapping are useful DL techniques for generating artificial data and combining several smaller datasets.We present results of DL-based segmentation and accuracy in predicting glioma subtypes based on MRI features. Dice scores range from 0.77 to 0.89. In mixed glioma cohorts, IDH mutation can be predicted with a sensitivity of 0.98 and specificity of 0.97. Results in test cohorts have shown improvements of 5-7% in accuracy, following GAN-expansion of data and domain mapping of smaller datasets.The provided DL examples are promising, although not yet in clinical practice. DL has demonstrated usefulness in data augmentation and for overcoming data variability. DL methods should be further studied, developed, and validated for broader clinical use. Ultimately, DL models can serve as effective decision support systems, and are especially well-suited for time-consuming, detail-focused, and data-ample tasks.

Development and Validation of an Efficient MRI Radiomics Signature for Improving the Predictive Performance of 1p/19q Co-Deletion in Lower-Grade Gliomas

Simple Summary

Low-grade gliomas (LGG) with the 1p/19q co-deletion mutation have been proven to have a better survival prognosis and response to treatment than individuals without the mutation. Identifying this mutation has a vital role in managing LGG patients; however, the current diagnostic gold standard, including the brain-tissue biopsy or the surgical resection of the tumor, remains highly invasive and time-consuming. We proposed a model based on the eXtreme Gradient Boosting (XGBoost) classifier to detect 1p/19q co-deletion mutation using non-invasive medical images. The performance of our model achieved 87% and 82.8% accuracy on the training and external test set, respectively. Significantly, the prediction was based on only seven optimal wavelet radiomics features extracted from brain Magnetic Resonance (MR) images. We believe that this model can address clinicians in the rapid diagnosis of clinical 1p/19q co-deletion mutation, thereby improving the treatment prognosis of LGG patients.

Abstract

The prognosis and treatment plans for patients diagnosed with low-grade gliomas (LGGs) may significantly be improved if there is evidence of chromosome 1p/19q co-deletion mutation. Many studies proved that the codeletion status of 1p/19q enhances the sensitivity of the tumor to different types of therapeutics. However, the current clinical gold standard of detecting this chromosomal mutation remains invasive and poses implicit risks to patients. Radiomics features derived from medical images have been used as a new approach for non-invasive diagnosis and clinical decisions. This study proposed an eXtreme Gradient Boosting (XGBoost)-based model to predict the 1p/19q codeletion status in a binary classification task. We trained our model on the public database extracted from The Cancer Imaging Archive (TCIA), including 159 LGG patients with 1p/19q co-deletion mutation status. The XGBoost was the baseline algorithm, and we combined the SHapley Additive exPlanations (SHAP) analysis to select the seven most optimal radiomics features to build the final predictive model. Our final model achieved an accuracy of 87% and 82.8% on the training set and external test set, respectively. With seven wavelet radiomics features, our XGBoost-based model can identify the 1p/19q codeletion status in LGG-diagnosed patients for better management and address the drawbacks of invasive gold-standard tests in clinical practice.

Quantitative MRI-based radiomics for noninvasively predicting molecular subtypes and survival in glioma patients

Gliomas can be classified into five molecular groups based on the status of IDH mutation, 1p/19q codeletion, and TERT promoter mutation, whereas they need to be obtained by biopsy or surgery. Thus, we aimed to use MRI-based radiomics to noninvasively predict the molecular groups and assess their prognostic value. We retrospectively identified 357 patients with gliomas and extracted radiomic features from their preoperative MRI images. Single-layered radiomic signatures were generated using a single MR sequence using Bayesian-regularization neural networks. Image fusion models were built by combing the significant radiomic signatures. By separately predicting the molecular markers, the predictive molecular groups were obtained. Prognostic nomograms were developed based on the predictive molecular groups and clinicopathologic data to predict progression-free survival (PFS) and overall survival (OS). The results showed that the image fusion model incorporating radiomic signatures from contrast-enhanced T1-weighted imaging (cT1WI) and apparent diffusion coefficient (ADC) achieved an AUC of 0.884 and 0.669 for predicting IDH and TERT status, respectively. cT1WI-based radiomic signature alone yielded favorable performance in predicting 1p/19q status (AUC = 0.815). The predictive molecular groups were comparable to actual ones in predicting PFS (C-index: 0.709 vs. 0.722, P = 0.241) and OS (C-index: 0.703 vs. 0.751, P = 0.359). Subgroup analyses by grades showed similar findings. The prognostic nomograms based on grades and the predictive molecular groups yielded a C-index of 0.736 and 0.735 in predicting PFS and OS, respectively. Accordingly, MRI-based radiomics may be useful for noninvasively detecting molecular groups and predicting survival in gliomas regardless of grades.

MRI-Based Deep-Learning Method for Determining Glioma MGMT Promoter Methylation Status

BACKGROUND AND PURPOSE

O6-Methylguanine-DNA methyltransferase (MGMT) promoter methylation confers an improved prognosis and treatment response in gliomas. We developed a deep learning network for determining MGMT promoter methylation status using T2 weighted Images (T2WI) only.

MATERIALS AND METHODS

Brain MR imaging and corresponding genomic information were obtained for 247 subjects from The Cancer Imaging Archive and The Cancer Genome Atlas. One hundred sixty-three subjects had a methylated MGMT promoter. A T2WI-only network (MGMT-net) was developed to determine MGMT promoter methylation status and simultaneous single-label tumor segmentation. The network was trained using 3D-dense-UNets. Three-fold cross-validation was performed to generalize the performance of the networks. Dice scores were computed to determine tumor-segmentation accuracy.

RESULTS

The MGMT-net demonstrated a mean cross-validation accuracy of 94.73% across the 3 folds (95.12%, 93.98%, and 95.12%, [SD, 0.66%]) in predicting MGMT methylation status with a sensitivity and specificity of 96.31% [SD, 0.04%] and 91.66% [SD, 2.06%], respectively, and a mean area under the curve of 0.93 [SD, 0.01]. The whole tumor-segmentation mean Dice score was 0.82 [SD, 0.008].

CONCLUSIONS

We demonstrate high classification accuracy in predicting MGMT promoter methylation status using only T2WI. Our network surpasses the sensitivity, specificity, and accuracy of histologic and molecular methods. This result represents an important milestone toward using MR imaging to predict prognosis and treatment response.