Brainlesion: Glioma, Multiple Sclerosis, Stroke and Traumatic Brain Injuries

In this paper we propose a new generative model for simultaneous brain parcellation and white matter lesion segmentation from multi-contrast magnetic resonance images. The method combines an existing whole-brain segmentation technique with a novel spatial lesion model based on a convolutional restricted Boltzmann machine. Unlike current state-of-the-art lesion detection techniques based on discriminative modeling, the proposed method is not tuned to one specific scanner or imaging protocol, and simultaneously segments dozens of neuroanatomical structures. Experiments on a public benchmark dataset in multiple sclerosis indicate that the method’s lesion segmentation accuracy compares well to that of the current state-of-the-art in the field, while additionally providing robust whole-brain segmentations.

The accurate segmentation of lesions in magnetic resonance images of stroke patients is important, for example, for comparing the location of the lesion with functional areas and for determining the optimal strategy for patient treatment. Manual labeling of each lesion turns out to be time-intensive and costly, making an automated method desirable. Standard approaches for brain parcellation make use of spatial atlases that represent prior information about the spatial distribution of different tissue types and of anatomical structures of interest. Different from healthy tissue, however, the spatial distribution of a stroke lesion varies considerably, limiting the use of such brain image segmentation approaches for stroke lesion analysis, and for integrating brain parcellation with stroke lesion segmentation. We propose to amend the standard atlas-based generative image segmentation model by a spatial atlas of stroke lesion occurrence by making use of information about the vascular territories. As the territories of the major arterial trees often coincide with the location and extensions of large stroke lesions, we use 3D maps of the vascular territories to form patient-specific atlases combined with outlier information from an initial run, following an iterative procedure. We find our approach to perform comparable to (or better than) standard approaches that amend the tissue atlas with a flat lesion prior or that treat lesion as outliers, and to outperform both for large heterogeneous lesions.

Traumatic brain injury (TBI) can cause widespread and long-lasting damage to white matter. Diffusion weighted imaging methods are uniquely sensitive to this disruption. Even so, traumatic injury often disrupts brain morphology as well, complicating the analysis of brain integrity and connectivity, which are typically evaluated with tractography methods optimized for analyzing normal healthy brains. To understand which fiber tracking methods show promise for analysis of TBI, we tested 9 different tractography algorithms for their classification accuracy and their ability to identify vulnerable areas as candidates for longitudinal follow-up in pediatric TBI participants and matched controls. Deterministic tractography models yielded the highest classification accuracies, but their limitations in areas of extensive fiber crossing suggested that they generated poor candidates for longitudinal follow-up. Probabilistic methods, including a method based on the Hough transform, yielded slightly lower accuracy, but generated follow-up candidate connections more coherent with the known neuropathology of TBI.

White-matter lesions are associated to several diseases, which can be characterized by neuroimaging biomarkers through lesion segmentation in MR images. We present a novel automated lesion segmentation method consisting of an unsupervised mixture model based extraction of candidate lesion voxels, which are subsequently classified by a random decision forest (RDF) using simple visual features like multi-sequence MR intensities sourced from connected voxel neighborhoods. The candidate lesion extraction prior to RDF training and classification balanced the number of non-lesion and lesion voxels and the number of non-lesion classes versus a lesion class. Thereby, the RDF established highly discriminating decision rules based on such simple visual features, which have the benefit of no computational overhead and easy extraction from the MR images. On MR images of 18 patients with multiple sclerosis the proposed method achieved the median Dice similarity of 0.73, sensitivity of 0.90 and positive predictive value of 0.61, which indicate accurate segmentation of white-matter lesions.

We report our methodological developments to investigate, in a multi-center study using mean diffusivity, the tissue damage caused by a severe traumatic brain injury (GSC \(<9\)) in the 10 days post-event. To assess the diffuse aspect of the injury, we fuse several atlases to parcel cortical, subcortical and WM structures into well identified regions where MD values are computed and compared to normative values. We used P-LOCUS to provide brain tissue segmentation and exclude voxels labeled as CSF, ventricles and hemorrhagic lesion and then automatically detect the lesion load. Preliminary results demonstrate that our method is coherent with expert opinion in the identification of lesions. We outline the challenges posed in automatic analysis for TBI.

Brain tumor segmentation and brain tumor growth assessment are inter-dependent and benefit from a joint evaluation. Starting from a generative model for multimodal brain tumor segmentation, we make use of a nonparametric growth model that is implemented as a conditional random field (CRF) including directed links with infinite weight in order to incorporate growth and inclusion constraints, reflecting our prior belief on tumor occurrence in the different image modalities. In this study, we validate this model to obtain brain tumor segmentations and volumetry in longitudinal image data. Moreover, we use the model to develop a probabilistic framework for estimating the likelihood of disease progression, i.e. tumor regrowth, after therapy. We present experiments for longitudinal image sequences with , , and flair images, acquired for ten patients with low and high grade gliomas.

Diffusion Tensor Imaging (DTI), the Fractional Anisotropy (FA) is used to measure the integrity of the white matter (WM); it is considered as a biomarker for stroke recovery. This measure is highly sensitive to applied pre-processing steps; in particular, the presence of a lesion may result into severe misregistration. In this paper, it is proposed to quantitatively assess the impact of large stroke lesions onto the registration process. To reduce this impact, a new registration algorithm, that localizes the lesion via Bayesian estimation, is proposed.

Immunohistochemistry is widely used as a gold standard to inspect tissues, characterize their structure and detect pathological alterations. As such, the joint analysis of histological images and other imaging modalities (MRI, PET) is of major interest to interpret these physical signals and establish their correspondence with the biological constitution of the tissues. However, it is challenging to provide a meaningful characterization of the signal specificity. In this paper, we propose an integrated method to quantitatively evaluate the discriminative power of imaging modalities. This method was validated using a macaque brain dataset containing: 3 immunohistochemically stained and 1 histochemically stained series, 1 photographic volume and 1 in vivo T2 weighted MRI. First, biological regions of interest (ROIs) were automatically delineated from histological sections stained for markers of interest and mapped on the target non-specific modalities through co-registration. These non-overlapping ROIs were considered ground truth for later classification. Voxels were evenly split in training and testing sets for a logistic regression model. The statistical significance of resulting accuracy scores was evaluated through null distribution simulations. Such an approach could be of major interest to assess relevant biological characteristics from various imaging modalities.

From clinical practice as well as research methods arises the need for accurate, reproducible and reliable segmentation of pathological areas from brain MR scans. This paper describes a set of hand-selected, voxel-based image features highly suitable for the tissue discrimination task. Embedded in a random decision forest framework, the proposed method was applied to sub-acute ischemic stroke (ISLES 2015 - SISS), acute ischemic stroke (ISLES 2015 - SPES) and glioma (BRATS 2015) segmentation with only minor adaptation. For all of these three challenges, our generic approach received high ranks, among them a second place. The outcome underlines the robustness of our features for segmentation in brain MR, while simultaneously stressing the necessity for highly specialized solution to achieve state-of-the-art performance.

In their most aggressive form, the mortality rate of gliomas is high. Accurate segmentation is important for surgery and treatment planning, as well as for follow-up evaluation. In this paper, we propose to segment brain tumors using a Deep Convolutional Neural Network. Neural Networks are known to suffer from overfitting. To address it, we use Dropout, Leaky Rectifier Linear Units and small convolutional kernels. To segment the High Grade Gliomas and Low Grade Gliomas we trained two different architectures, one for each grade. Using the proposed method it was possible to obtain promising results in the 2015 Multimodal Brain Tumor Segmentation (BraTS) data set, as well as the second position in the on-site challenge.

We present an approach for segmenting low- and high-grade gliomas in multimodal magnetic resonance imaging volumes. The proposed approach is based on a hybrid generative-discriminative model. Firstly, a generative approach based on an Expectation-Maximization framework that incorporates a glioma growth model is used to segment the brain scans into tumor, as well as healthy tissue labels. Secondly, a gradient boosting multi-class classification scheme is used to refine tumor labels based on information from multiple patients. Lastly, a probabilistic Bayesian strategy is employed to further refine and finalize the tumor segmentation based on patient-specific intensity statistics from the multiple modalities. We evaluated our approach in 186 cases during the training phase of the BRAin Tumor Segmentation (BRATS) 2015 challenge and report promising results. During the testing phase, the algorithm was additionally evaluated in 53 unseen cases, achieving the best performance among the competing methods.

In this work, we investigated the potential of a recently proposed parameter learning algorithm for Conditional Random Fields (CRFs). Parameters of a pairwise CRF are estimated via a stochastic subgradient descent of a max-margin learning problem. We compared the performance of our brain tumor segmentation method using parameter learning to a version using hand-tuned parameters. Preliminary results on a subset of the BRATS2015 training set show that parameter learning leads to comparable or even improved performance. In addition, we also performed experiments to study the impact of the composition of training data on the final segmentation performance. We found that models trained on mixed data sets achieve reasonable performance compared to models trained on stratified data.

Accurate Segmentation of Gliomas from Magnetic Resonance Images (MRI) is required for treatment planning and monitoring disease progression. As manual segmentation is time consuming, an automated method can be useful, especially in large clinical studies. Since Gliomas have variable shape and texture, automated segmentation is a challenging task and a number of techniques based on machine learning algorithms have been proposed. In the recent past, deep learning methods have been tested on various image processing tasks and found to outperform state of the art techniques. In our work, we consider stacked denoising autoencoder (SDAE), a deep neural network that reconstructs its input. We trained a three layer SDAE where the input layer was a concatenation of fixed size 3D patches (11\(\,\times \,\)11\(\,\times \,\)3 voxels/neurons) from multiple MRI sequences. The 2nd, 3rd and 4th layers had 3000, 1000 and 500 neurons respectively. Two different networks were trained one with high grade glioma (HGG) data and other with a combination of high grade and low grade gliomas (LGG). Each network was trained with 35 patients for pre-training and 21 patients for fine tuning. The predictions from the two networks were combined based on maximum posterior probability. For HGG data, the whole tumor dice score was .81, tumor core was .68 and active tumor was .64 (\(n=220\) patients). For LGG data, the whole tumor dice score was .72, tumor core was .42 and active tumor was .29 (\(n=54\) patients).

We consider the problem of fully automatic brain focal pathology segmentation, in MR images containing low and high grade gliomas and ischemic stroke lesion. We propose a Convolutional Neural Network (CNN) approach which is amongst the top performing methods while also being extremely computationally efficient, a balance that existing methods have struggled to achieve. Our CNN is trained directly on the image modalities and thus learns a feature representation directly from the data. We propose a cascaded architecture with two pathways: one which focuses on small details in gliomas and one on the larger context. We also propose a two-phase patch-wise training procedure allowing us to train models in a few hours. Fully exploiting the convolutional nature of our model also allows us to segment a complete brain image in 25 s to 3 min. Experimental results on BRain Tumor Segmentation challenges (BRATS’13, BRATS’15) and Ischemic Stroke Lesion Segmentation challenge (ISLES’15) reveal that our approach is among the most accurate in the literature, while also being computationally very efficient.

Automated methods for segmentation of ischemic stroke lesions could significantly reduce the workload of radiologists and speed up the beginning of patient treatment. In this paper, we present a method for subacute ischemic stroke lesion segmentation from multispectral magnetic resonance images (MRI). The method involves classification of voxels with a Random Forest algorithm and subsequent classification refinement with contextual clustering. In addition, we utilize the training data to build statistical group-specific templates and use them for calculation of individual voxel-wise differences from the global mean. Our method achieved a Dice coefficient of 0.61 for the leave-one-out cross-validated training data and 0.47 for the testing data of the ISLES challenge 2015.

Stroke is a common cause of sudden death and disability worldwide. In clinical practice, brain magnetic resonance (MR) scans are used to assess the stroke lesion presence. In this work, we have built a fully automatic stroke lesion segmentation system using 3D brain magnetic resonance (MR) data. The system contains a 3D registration framework and a 3D multi-random forest model trained from the data provided by the Ischemic Stroke Lesion Segmentation (ISLES) challenge of the 18th International Conference on Medical Image Computing and Computer Assisted Intervention. The preliminary test results show that the presented system is capable to detect stroke lesion from 3D brain MRI data.

Accurate segmentation of ischemic lesions is still a challenging task. In this paper, we propose a framework to extract ischemic lesions from multi-spectral MR images. In the proposed framework, MR images of each modality are first segmented into brain tissues and ischemic lesions by weighting suppressed fuzzy c-means. Preliminary lesion segmentation results are then fused among all the imaging modalities by majority voting. The fused segmentation results are finally refined by a three phase level set method. The level set formulation is defined on multi-spectral images with the capability of dealing with intensity inhomogeneities. The proposed framework has been applied to the MICCAI 2015 ISLES challenge. According to the ranking rules of the challenge, the proposed framework took the second place and the fourth place in sub-acute lesion segmentation and acute stroke estimation, respectively.

We present a novel fully-automated generative ischemic stroke lesion segmentation method that can be applied to individual patient images without need for a training data set. An Expectation Maximization-approach is used for estimating intensity models for both normal and pathological tissue. The segmentation is represented by a level-set that is iteratively updated to label voxels as either normal or pathological, based on which intensity model explains the voxels’ intensity the best. A convex level-set formulation is adopted, that eliminates the need for manual initialization of the level-set. The performance of the method for segmenting the ischemic stroke is summarized by an average Dice score of \(0.78\pm 0.08\) and \(0.53 \pm 0.26\) for the SPES and SISS 2015 training data set respectively and \(0.67\pm 0.24\) and \(0.37 \pm 0.33\) for the test data set.

Automated localisation and segmentation of stroke lesions in patients is of great interest to clinicians and researchers alike. We propose a supervised method based on cascaded extremely randomised trees for lesion segmentation, working on a per voxel basis in native subject space. The proposed pipeline is evaluated in the MICCAI Ischemic Stroke Lesion Segmentation (ISLES) challenge, both with nested cross-validation on the training data as well as on independent, multi-centre test data. We obtained good performance although inter-subject variability is large, and reached 3rd place in the SPES sub-challenge.

This paper presents an automated segmentation framework for ischemic stroke lesion segmentation in multi-spectral MRI images. The framework is based on a random forests (RF), which is an ensemble learning technique that generates several classifiers and combines their results in order to make decisions. In RF, we employ several meaningful features such as intensities, entropy, gradient etc. to classify the voxels in multi-spectral MRI images. The segmentation framework is validated on both training and testing data, obtained from MICCAI ISLES-2015 SISS challenge dataset. The performance of the framework is evaluated relative to the manual segmentation (ground truth). The experimental results demonstrate the robustness of the segmentation framework, and that it achieves reasonable segmentation accuracy for segmenting the sub-acute ischemic stroke lesion in multi-spectral MRI images.

We propose a fully automatic method for segmenting the ischemic penumbra, using image texture and spatial features and a modified Random Forest algorithm, which we call Segmentation Forests, which has been designed to adapt the original Random Forests algorithm of Breiman to the segmentation of medical images. The method was trained and tested on the SPES dataset, part of the ISLES MICCAI Grand Challenge. The method is fast, taking approximately six minutes to segment a new case, and yields convincing results. On the testing portion of the SPES dataset, the method achieved an average Dice coefficient of 0.82, with a standard deviation of 0.08.

In machine learning larger databases are usually associated with higher classification accuracy due to better generalization. This generalization may lead to non-optimal classifiers in some medical applications with highly variable expressions of pathologies. This paper presents a method for learning from a large training base by adaptively selecting optimal training samples for given input data. In this way heterogeneous databases are supported two-fold. First, by being able to deal with sparsely annotated data allows a quick inclusion of new data set and second, by training an input-dependent classifier. The proposed approach is evaluated using the SISS challenge. The proposed algorithm leads to a significant improvement of the classification accuracy.