Final Report Summary - DEBRA (Detection of Brain Abnormality)
The project DEBRA includes the development and implementation of medical imaging techniques for measuring cerebrovascular disease (CVD) from multi-parametric magnetic resonance (MR) images by employing advanced segmentation methods based on pattern classification and statistical modeling. CVD, both clinical and subclinical, is a very significant health problem, especially in view of the increasing aging population. It is also highly prevalent in diabetic populations, therefore measuring disease burden, progression, and response to treatments is very important for patient management in diabetes. Magnetic resonance imaging (MRI) is currently the most widely used way to characterise in vivo the type and extent of brain lesions. The goal of DEBRA is to develop an automated computer-assisted method for segmenting brain lesions that will provide more stable measurements compared to relatively more conventional methods and therefore increase the sensitivity of detecting subtle effects. While the development of tools mainly targets the measurement of CVD, DEBRA's methodology also aims to hold widespread potential for applications in other neuroimaging studies involving abnormality detection.
The main components of the project include the collection of data (clinical and simulated), the application of advanced three-dimensional (3D) image analysis methods and validation. The methodology of DEBRA is based on the idea of incorporating a-priori knowledge about characteristics of normal data and apply semi-supervised abnormality detection. In abnormality detection a test sample is characterised as anomalous if it does not comply with the probability distribution obtained from normal data. Two main research directions were followed both based on semi-supervised learning. In the first approach (for which most of the time was dedicated) the variation of the healthy anatomy was modelled without assuming statistical independence of nearby voxel intensities, whereas in the second approach, statistical independence was assumed (voxel-wise analysis). Anatomical structure is characterised by spatial continuity and follows specific patterns, thus the second approach (which most of the researchers follow) is a simplification. Since it does not take into respect the considerable spatial structure it cannot be easily applied in the detection of structural abnormalities. On the other hand, the first approach provides the challenge of high-dimensional data (dimensionality = number of voxels) and the lack of enough samples to statistically model the whole brain image. In both approaches, the challenges were addressed and corresponding results were successfully published. The methods were applied in the segmentation of white matter lesions, brain infarction and cortical dysplasia. As a side direction, we also developed an automatic method for segmenting or classifying other types of brain pathology, such as brain neoplasms in multiparametric MRI.
Data:
The first research direction involved clinical data consisted of axial FLAIR scans obtained from elderly individuals with diabetes from the ACCORD-MIND neuroimaging study, and provided by University of Pennsylvania. Some of these images which included white matter and periventricular lesions were selected for ground truth definition (performed by a clinical expert). Out of the rest of the images (without definition of lesion extent), the ones without lesions were used as part of the training set (healthy controls). Simulated data were also used and included three types of pathology:
(a) white matter and periventricular lesions;
(b) infarcts; and
(c) dysplasia.
The second research direction involved data consisting of transaxial T1w and FLAIR scans obtained from different imaging sites. Different sets of data were selected for testing the method including lesions with or without necrosis. The first set consisted of diabetic patients with white matter lesions and necrotic infarcts. The second set involved subjects selected from an aging population. The abnormality masks used for assessing the method were manually delineated by a single rater for the first dataset and by two independent raters for the second dataset.
Results:
The abnormalities detected by the main approach were compared against the region of simulated pathology for simulated data and expert-defined lesion mask for real data. The average (over all test data) rank order correlation (ROC) performance curves showed that the proposed method performs better than segmentation based on univariate statistics (z-score) for both simulated and real datasets. Also the method demonstrated improvement in image segmentation over two-group analysis performed by statistical parametric mapping (SPM), which is a package widely used for the analysis of brain imaging data sequences. The second research direction also achieved satisfactory results. The segmentation of white matter lesions (hyperintense lesions) approached the accuracy of manual segmentation in the cases of images with high lesion load. The detection of necrotic infarcts was only slightly worse than the performance of human experts.
Potential impact:
The cornerstone of DEBRA is the creation of a high-dimensional pattern classification framework that showed to identify subtle and spatially complex patterns of brain pathology that might have high predictive power in various clinical research studies. Most of the methods developed for brain lesion segmentation either require manual interaction such as samples labelling or seeds selection, apply a priori knowledge, or target specific tissue (and not structural) abnormalities. The developed image analysis methods avoid the tedious training phase and have wide spread applicability for different types of pathology. They can either be used as stand-alone tools that highlight potential foci of pathology, thus being useful for expediting the screening process. Thus, DEBRA has high impact, both clinical and subclinical. In addition the project had main impact in Dr Zacharaki's professional and social integration, as well as on scientific knowledge transfer. She has developed collaborations and discussed ideas for future research paths with faculty members from different departments in University of Patras, leading to the submission of different projects and the acceptance of one project under the National Strategic Reference Framework (NSRF) and one supported by the Seventh Framework Programme (FP7). This will allow the continuation of her research career.
The main components of the project include the collection of data (clinical and simulated), the application of advanced three-dimensional (3D) image analysis methods and validation. The methodology of DEBRA is based on the idea of incorporating a-priori knowledge about characteristics of normal data and apply semi-supervised abnormality detection. In abnormality detection a test sample is characterised as anomalous if it does not comply with the probability distribution obtained from normal data. Two main research directions were followed both based on semi-supervised learning. In the first approach (for which most of the time was dedicated) the variation of the healthy anatomy was modelled without assuming statistical independence of nearby voxel intensities, whereas in the second approach, statistical independence was assumed (voxel-wise analysis). Anatomical structure is characterised by spatial continuity and follows specific patterns, thus the second approach (which most of the researchers follow) is a simplification. Since it does not take into respect the considerable spatial structure it cannot be easily applied in the detection of structural abnormalities. On the other hand, the first approach provides the challenge of high-dimensional data (dimensionality = number of voxels) and the lack of enough samples to statistically model the whole brain image. In both approaches, the challenges were addressed and corresponding results were successfully published. The methods were applied in the segmentation of white matter lesions, brain infarction and cortical dysplasia. As a side direction, we also developed an automatic method for segmenting or classifying other types of brain pathology, such as brain neoplasms in multiparametric MRI.
Data:
The first research direction involved clinical data consisted of axial FLAIR scans obtained from elderly individuals with diabetes from the ACCORD-MIND neuroimaging study, and provided by University of Pennsylvania. Some of these images which included white matter and periventricular lesions were selected for ground truth definition (performed by a clinical expert). Out of the rest of the images (without definition of lesion extent), the ones without lesions were used as part of the training set (healthy controls). Simulated data were also used and included three types of pathology:
(a) white matter and periventricular lesions;
(b) infarcts; and
(c) dysplasia.
The second research direction involved data consisting of transaxial T1w and FLAIR scans obtained from different imaging sites. Different sets of data were selected for testing the method including lesions with or without necrosis. The first set consisted of diabetic patients with white matter lesions and necrotic infarcts. The second set involved subjects selected from an aging population. The abnormality masks used for assessing the method were manually delineated by a single rater for the first dataset and by two independent raters for the second dataset.
Results:
The abnormalities detected by the main approach were compared against the region of simulated pathology for simulated data and expert-defined lesion mask for real data. The average (over all test data) rank order correlation (ROC) performance curves showed that the proposed method performs better than segmentation based on univariate statistics (z-score) for both simulated and real datasets. Also the method demonstrated improvement in image segmentation over two-group analysis performed by statistical parametric mapping (SPM), which is a package widely used for the analysis of brain imaging data sequences. The second research direction also achieved satisfactory results. The segmentation of white matter lesions (hyperintense lesions) approached the accuracy of manual segmentation in the cases of images with high lesion load. The detection of necrotic infarcts was only slightly worse than the performance of human experts.
Potential impact:
The cornerstone of DEBRA is the creation of a high-dimensional pattern classification framework that showed to identify subtle and spatially complex patterns of brain pathology that might have high predictive power in various clinical research studies. Most of the methods developed for brain lesion segmentation either require manual interaction such as samples labelling or seeds selection, apply a priori knowledge, or target specific tissue (and not structural) abnormalities. The developed image analysis methods avoid the tedious training phase and have wide spread applicability for different types of pathology. They can either be used as stand-alone tools that highlight potential foci of pathology, thus being useful for expediting the screening process. Thus, DEBRA has high impact, both clinical and subclinical. In addition the project had main impact in Dr Zacharaki's professional and social integration, as well as on scientific knowledge transfer. She has developed collaborations and discussed ideas for future research paths with faculty members from different departments in University of Patras, leading to the submission of different projects and the acceptance of one project under the National Strategic Reference Framework (NSRF) and one supported by the Seventh Framework Programme (FP7). This will allow the continuation of her research career.
