Arvid Lundervold, BSc, MD, PhDProfessor of Medical Information Technology and Physiology, University of Bergen, Norway - https://www.uib.no/en/persons/Arvid.Lundervold- https://scholar.google.com/citations?user=HqmyBUUAAAAJ
In this paper we focus on motion correction of contrast enhanced kidney MRI time series, which is... more In this paper we focus on motion correction of contrast enhanced kidney MRI time series, which is an important step towards accurate assessment of regional renal function. Due to respiratory motion and pulsations, the organ of interest undergoes complex movement and deformation, which disturb further renal function analysis. We propose geometric movement correction by image registration. We have compared rigid and nonrigid registration methods as well as registration of whole images and registration limited to ROI that defines the organ under investigation. The obtained results show that image registration methods benefit to renal function analysis, i.e. to the assessment of intensity time courses. Furthermore, the comparison of the registration methods shows benefits of ROI limited methods and eventual problems of nonrigid methods.
Computer-aided diagnosis of Alzheimer's disease (AD) is a rapidly developing field of neuroim... more Computer-aided diagnosis of Alzheimer's disease (AD) is a rapidly developing field of neuroimaging with strong potential to be used in practice. In this context, assessment of models' robustness to noise and imaging protocol differences together with post-processing and tuning strategies are key tasks to be addressed in order to move towards successful clinical applications. In this study, we investigated the efficacy of Random Forest classifiers trained using different structural MRI measures, with and without neuroanatomical constraints in the detection and prediction of AD in terms of accuracy and between-cohort robustness. From The ADNI database, 185 AD, and 225 healthy controls (HC) were randomly split into training and testing datasets. 165 subjects with mild cognitive impairment (MCI) were distributed according to the month of conversion to dementia (4-year follow-up). Structural 1.5-T MRI-scans were processed using Freesurfer segmentation and cortical reconstruction....
The goal of this prospective study was to evaluate multispectral analysis techniques for automati... more The goal of this prospective study was to evaluate multispectral analysis techniques for automatic recognition of uterine cancer in magnetic resonance (MR) imaging. The first part of this study was an open training phase in which the statistical parameters of the various normal and pathological tissue types were estimated. This was followed by a test phase that was done as a blind experiment. Results from an extensive pathological examination of the surgically removed organs served as the reference for the diagnosis and various geometric measurements of the lesions. A radiological examination of the MR images was also performed. All malignant test tumors were correctly or close to correctly classified. However, parts of normal endometrium and other mucosal linings were also classified as adenocarcinomas. In addition, parts of some of the malignant tumors were classified as normal endometrium. The geometrical extension of the tumor and its relationship to the surroundings were slightly better predicted than those obtained by the radiologist. The results indicate that it is possible to differentiate and determine the local extension of some types of uterine malignancies based on the information present in MR images. o 1992 Academic press, Inc. 1, INTRODUCTION Tissue classification of benign and malignant tumor types in vivo using magnetic resonance (MR) images has been the aim of several recent research projects (8, 14, 20). If possible, such a reliable determination would have significant impact on present diagnostic and surgical procedures for malignant tumor diseases. However, the results so far have not been very promising (8, 14, 16); but see also (19, 21). Furthermore, to the best of our knowledge no prospective evaluation of the obtained classification results has been published. In this paper we have applied multispectral analysis techniques traditionally used in remote sensing to do a prospective statistical classification study of malignant uterine corpus tumors (for a similar approach, see (27,28)). The recording of the multispectral MR images was part of the routine diagnostic examination of uterine tumors. Malignant tumors of the uterine corpus were studied because most of them are surgically removed, making an extensive pathological examination possible. They are 55
Section for Medical Image Analysis and Pattern Recognition University of Bergen, Arstadveien 19, ... more Section for Medical Image Analysis and Pattern Recognition University of Bergen, Arstadveien 19, N-5009 Bergen, Norway ... In this study we demonstrate an improved differen-tiation of the most common tissue types in the human brain and surrounding structures by quantitative ...
Magnetic resonance imaging (MRI) of the brain, followed by automated segmentation of the corpus c... more Magnetic resonance imaging (MRI) of the brain, followed by automated segmentation of the corpus callosum (CC) in midsagittal sections have important applications in both clinical neurology and neurocognitive research since the size and shape of the CC are shown to be correlated to sex, age, neurodegenerative diseases and various lateralized behavior in man. Moreover, whole head, multispectral 3D MRI recordings enable voxel-based tissue classification and estimation of total brain volumes, in addition to CC morphometric parameters. We propose a new algorithm that uses both multispectral MRI measurements (intensity values) and prior information about shape (CC template) to segment CC in midsagittal slices with very little user interaction. The algorithm has been tested on a sample of 10 subjects scanned with multispectral 3D MRI, collected for a study of dyslexia, with very good agreement between the manually traced ("true") CC outline and the detected CC outline. We conclude that the proposed method for CC segmentation is promising for clinical use when multispectral MR images are recorded. 10 multispectral 3D whole-brain acquisitions and a comparison to manual tracings of the corpus callosum, are reported in Section 4. Section 5 concludes the paper with a
The authors demonstrate an improved differentiation of the most common tissue types in the human ... more The authors demonstrate an improved differentiation of the most common tissue types in the human brain and surrounding structures by quantitative validation using multispectral analysis of magnetic resonance images. This is made possible by a combination of a special training technique and an increase in the number of magnetic resonance channel images with different pulse acquisition parameters. The authors give a description of the tissue-specific multivariate statistical distributions of the pixel intensity values and discuss how their properties may be explored to improve the statistical modeling further. A statistical method to estimate the tissue-specific longitudinal and transverse relaxation times is also given. It is concluded that multispectral analysis of magnetic resonance images is a valuable tool to recognize the most common normal tissue types in the brain and surrounding structures.
Medical image registration can be formulated as a tissue deformation problem, where parameter est... more Medical image registration can be formulated as a tissue deformation problem, where parameter estimation methods are used to obtain the inverse deformation. However, there is limited knowledge about the ability to recover an unknown deformation. The main objective of this study is to estimate the quality of a restored deformation field obtained from image registration of dynamic MR sequences. Methods: We investigate the behavior of forward deformation models of various complexities. Further, we study the accuracy of restored inverse deformations generated by image registration. Results: We found that the choice of 1) heterogeneous tissue parameters and 2) a poroelastic (instead of elastic) model had significant impact on the forward deformation. In the image registration problem, both 1) and 2) were found not to be significant. Here, the presence of image features were dominating the performance. We also found that existing algorithms will align images with high precision while at the same time obtain a deformation field with a relative error of 40%. Conclusion: Image registration can only moderately well restore the true deformation field. Still, estimation of volume changes instead of deformation fields can be fairly accurate and may represent a proxy for variations in tissue characteristics. Volume changes remain essentially unchanged under choice of discretization and the prevalence of pronounced image features. Significance: We suggest that image registration of high-contrast MR images has potential to be used as a tool to produce imaging biomarkers sensitive to pathology affecting tissue stiffness.
Background Tumor burden assessment is essential for radiation therapy (RT), treatment response ev... more Background Tumor burden assessment is essential for radiation therapy (RT), treatment response evaluation, and clinical decision-making. However, manual tumor delineation remains laborious and challenging due to radiological complexity. The objective of this study was to investigate the feasibility of the HD-GLIO tool, an ensemble of pre-trained deep learning models based on the nnUNet-algorithm, for tumor segmentation, response prediction, and its potential for clinical deployment. Methods We analyzed the predicted contrast-enhanced (CE) and non-enhancing (NE) HD-GLIO output in 49 multi-parametric MRI examinations from 23 grade-4 glioma patients. The volumes were retrospectively compared to corresponding manual delineations by 2 independent operators, before prospectively testing the feasibility of clinical deployment of HD-GLIO-output to a RT setting. Results For CE, median Dice scores were 0.81 (95% CI 0.71–0.83) and 0.82 (95% CI 0.74–0.84) for operator-1 and operator-2, respecti...
BackgroundGlioblastoma (GBM) is an aggressive malignant brain tumor where median survival is appr... more BackgroundGlioblastoma (GBM) is an aggressive malignant brain tumor where median survival is approximately 15 months after best available multimodal treatment. Recurrence is inevitable, largely due to O6 methylguanine DNA methyltransferase (MGMT) that renders the tumors resistant to temozolomide (TMZ). We hypothesized that pretreatment with bortezomib (BTZ) 48 hours prior to TMZ to deplete MGMT levels would be safe and tolerated by patients with recurrent GBM harboring unmethylated MGMT promoter. The secondary objective was to investigate whether 26S proteasome blockade may enhance differentiation of cytotoxic immune subsets to impact treatment responses measured by radiological criteria and clinical outcomes.MethodsTen patients received intravenous BTZ 1.3 mg/m2 on days 1, 4, and 7 during each 4th weekly TMZ‐chemotherapy starting on day 3 and escalated from 150 mg/m2 per oral 5 days/wk via 175 to 200 mg/m2 in cycles 1, 2, and 3, respectively. Adverse events and quality of life were...
What has happened in machine learning lately, and what does it mean for the future of medical ima... more What has happened in machine learning lately, and what does it mean for the future of medical image analysis? Machine learning has witnessed a tremendous amount of attention over the last few years. The current boom started around 2009 when so-called deep artificial neural networks began outperforming other established models on a number of important benchmarks. Deep neural networks are now the state-of-the-art machine learning models across a variety of areas, from image analysis to natural language processing, and widely deployed in academia and industry. These developments have a huge potential for medical imaging technology, medical data analysis, medical diagnostics and healthcare in general, slowly being realized. We provide a short overview of recent advances and some associated challenges in machine learning applied to medical image processing and image analysis. As this has become a very broad and fast expanding field we will not survey the entire landscape of applications, but put particular focus on deep learning in MRI. Our aim is threefold: (i) give a brief introduction to deep learning with pointers to core references; (ii) indicate how deep learning has been applied to the entire MRI processing chain, from acquisition to image retrieval, from segmentation to disease prediction; (iii) provide a starting point for people interested in experimenting and perhaps contributing to the field of deep learning for medical imaging by pointing out good educational resources, state-of-the-art open-source code, and interesting sources of data and problems related medical imaging.
Skull stripping in brain imaging is the removal of the parts of images corresponding to non-brain... more Skull stripping in brain imaging is the removal of the parts of images corresponding to non-brain tissue. Fast and accurate skull stripping is a crucial step for numerous medical brain imaging applications, e.g. registration, segmentation and feature extraction, as it eases subsequent image processing steps. In this work, we propose and compare two novel skull stripping methods based on 2D and 3D convolutional neural networks trained on a large, heterogeneous collection of 2777 clinical 3D T1-weighted MRI images from 1681 healthy subjects. We investigated the performance of the models by testing them on 927 images from 324 subjects set aside from our collection of data, in addition to images from an independent, large brain imaging study: the IXI dataset (n = 556). Our models achieved mean Dice scores higher than 0.978 and Jaccard indices higher than 0.957 on all tests sets, making predictions on new unseen brain MR images in approximately 1.4s for the 3D model and 12.4s for the 2D ...
2016 Signal Processing: Algorithms, Architectures, Arrangements, and Applications (SPA), 2016
This paper describes a pilot study aimed at evaluating the usefulness of texture parameters compu... more This paper describes a pilot study aimed at evaluating the usefulness of texture parameters computed from blood pharmacokinetic maps as biomarkers of cancers. The maps are extracted from DCE MRI images which, in turn, visualize blood perfusion in tissue by means of a contrast medium. Each voxel of a DCE image is characterised by a curve in time domain (enhancement signal). The shape of such curves is characteristic to different tissues and is quantified by fitting a proposed parameterized signal model. Different-modality MR images were acquired for a small number of patients, prior to diagnosed cancer surgical extraction. Six images were collected for each patient, four of them were distributions of blood pharmacokinetic model parameters and two — ADC and VIBE MR images. The extracted carcinoma tissue for each patient was investigated by pathologists. This histological grade provided labels to images for supervised data exploration. Both unsupervised (clustering) and supervised (LDA) experiments demonstrated the image texture can be a promising biomarker to reflect correlations between tumour tissue appearance at structural and functional MRI and the corresponding histological characteristics.
A reduction or reversal of the normal leftward asymmetry of the planum temporale (PT) has been cl... more A reduction or reversal of the normal leftward asymmetry of the planum temporale (PT) has been claimed to be typical of dyslexia, although some recent studies have challenged this view. In a population-based study of 20 right-handed dyslexic boys and 20 matched controls, we have measured the PT and the adjacent planum parietale (PP) region in sagittal magnetic resonance images. For the PT, mean left and right areas and asymmetry coefficients were compared. Since a PP area often could not be identified in one or both hemispheres, a qualitative comparison was used for this region. The total planar area (sum of PT and PP) was also compared between the two groups. A dichotic listening (DL) test with consonant-vowel syllables was administered to assess functional asymmetry of language. The results showed a mean leftward PT asymmetry in both the dyslexic and the control group, with no significant difference for the degree of PT asymmetry. Planned comparisons revealed however, a trend towards smaller left PT in the dyslexic group. In control children, but not in the dyslexic children, a significant correlation between PT asymmetry and reading was observed. A mean leftward asymmetry was also found for the total planar area, with no difference between the groups for the degree of asymmetry. Significantly fewer dyslexic children than control children showed a rightward asymmetry for the PP region. Both groups showed a normal right ear advantage on the DL task, with no significant difference for DL asymmetry. No significant correlation was observed between PT asymmetry and DL asymmetry. The present populationbased study adds to recent reports of normal PT asymmetry in dyslexia, but indicates that subtle morphological abnormalities in the left planar area may be present in this condition.
In the present study, we investigated differences between dyslexic and normal reading children in... more In the present study, we investigated differences between dyslexic and normal reading children in asymmetry of the planum temporale area in the upper posterior part of the temporal lobe and dichotic listening performance to consonant-vowel syllables. The current study was an extension of previous studies in our laboratory on the same participants, now including also girls and left-handers. There were 20 boys and 3 girls in the dyslexic group and 19 boys and 4 girls in the normal reading group. The age of the participants was 10-12 years for both groups. The participants were screened from a population of 950 students in the fourth school grade in the greater Bergen district. The planum temporale area was measured in sagittal magnetic resonance (MR) images. Mean left and right area and asymmetry index were compared between the groups. Dichotic presentations of consonant-vowel syllables made it possible to separately probe left and right hemisphere phonological function, and to correlate this with planum temporale area. The results showed a significantly larger left than right planum temporale area for both groups. However, while the right planum temporale area was similar for the dyslexic and control groups, the left planum temporale was significantly (one-tailed t-test) smaller in the dyslexic group. Both groups also showed a significant right ear advantage to the consonant-vowel syllables in the dichotic listening test. The relation between planum temporale and dichotic listening asymmetry showed a significant correlation for the dyslexic group only, indicating a positive relation between brain structure and function in dyslexic children. The results are discussed in terms of important subject characteristics with regard to brain markers of dyslexia.
Cellular multiparametric magnetic resonance imaging (MRI) provided an in vivo visualisation of ne... more Cellular multiparametric magnetic resonance imaging (MRI) provided an in vivo visualisation of neural stem cells' (NSCs) tropism for gliomas in the rat brain. NSCs were magnetically labelled in vitro with the bimodal gadolinium-based contrast agent, Gadolinium Rhodamine Dextran (GRID), and injected into the contralateral hemisphere to the developing tumour. Contrast-tonoise measurements showed that GRID-labelled cells induced a signal attenuation on both T2-, T2*-weighted images, and a modest signal gain on T1-weighted images. Tumour development and progression were longitudinally monitored in vivo by serial MR scanning. Measurements of tumour volume and tumour progression over time in terms of tumour doubling time showed a tendency towards a reduced tumour growth in NSC-treated animals. MR findings of migration and infiltration of tumours by labelled NSCs were corroborated with immunohistopathology, where labelled cells were detected in the corpus callosum at the tumour border and dispersed in the solid tumour tissue. Immunohistopathology also revealed that macrophages invaded the tumour tissue and in some cases engulfed GRID-labelled stem cells. No significant difference in macrophage recruitment between NSC-treated and vehicle-treated animals were detected, indicating that magnetically labelled NSC do not increase macrophage invasion of tumour tissue. Our findings demonstrate that cellular multiparametric MRI provides a valuable tool for in vivo dynamic monitoring of tumour-directed neural stem cell migration as well as therapeutic efficacy.
In this paper we focus on motion correction of contrast enhanced kidney MRI time series, which is... more In this paper we focus on motion correction of contrast enhanced kidney MRI time series, which is an important step towards accurate assessment of regional renal function. Due to respiratory motion and pulsations, the organ of interest undergoes complex movement and deformation, which disturb further renal function analysis. We propose geometric movement correction by image registration. We have compared rigid and nonrigid registration methods as well as registration of whole images and registration limited to ROI that defines the organ under investigation. The obtained results show that image registration methods benefit to renal function analysis, i.e. to the assessment of intensity time courses. Furthermore, the comparison of the registration methods shows benefits of ROI limited methods and eventual problems of nonrigid methods.
Computer-aided diagnosis of Alzheimer's disease (AD) is a rapidly developing field of neuroim... more Computer-aided diagnosis of Alzheimer's disease (AD) is a rapidly developing field of neuroimaging with strong potential to be used in practice. In this context, assessment of models' robustness to noise and imaging protocol differences together with post-processing and tuning strategies are key tasks to be addressed in order to move towards successful clinical applications. In this study, we investigated the efficacy of Random Forest classifiers trained using different structural MRI measures, with and without neuroanatomical constraints in the detection and prediction of AD in terms of accuracy and between-cohort robustness. From The ADNI database, 185 AD, and 225 healthy controls (HC) were randomly split into training and testing datasets. 165 subjects with mild cognitive impairment (MCI) were distributed according to the month of conversion to dementia (4-year follow-up). Structural 1.5-T MRI-scans were processed using Freesurfer segmentation and cortical reconstruction....
The goal of this prospective study was to evaluate multispectral analysis techniques for automati... more The goal of this prospective study was to evaluate multispectral analysis techniques for automatic recognition of uterine cancer in magnetic resonance (MR) imaging. The first part of this study was an open training phase in which the statistical parameters of the various normal and pathological tissue types were estimated. This was followed by a test phase that was done as a blind experiment. Results from an extensive pathological examination of the surgically removed organs served as the reference for the diagnosis and various geometric measurements of the lesions. A radiological examination of the MR images was also performed. All malignant test tumors were correctly or close to correctly classified. However, parts of normal endometrium and other mucosal linings were also classified as adenocarcinomas. In addition, parts of some of the malignant tumors were classified as normal endometrium. The geometrical extension of the tumor and its relationship to the surroundings were slightly better predicted than those obtained by the radiologist. The results indicate that it is possible to differentiate and determine the local extension of some types of uterine malignancies based on the information present in MR images. o 1992 Academic press, Inc. 1, INTRODUCTION Tissue classification of benign and malignant tumor types in vivo using magnetic resonance (MR) images has been the aim of several recent research projects (8, 14, 20). If possible, such a reliable determination would have significant impact on present diagnostic and surgical procedures for malignant tumor diseases. However, the results so far have not been very promising (8, 14, 16); but see also (19, 21). Furthermore, to the best of our knowledge no prospective evaluation of the obtained classification results has been published. In this paper we have applied multispectral analysis techniques traditionally used in remote sensing to do a prospective statistical classification study of malignant uterine corpus tumors (for a similar approach, see (27,28)). The recording of the multispectral MR images was part of the routine diagnostic examination of uterine tumors. Malignant tumors of the uterine corpus were studied because most of them are surgically removed, making an extensive pathological examination possible. They are 55
Section for Medical Image Analysis and Pattern Recognition University of Bergen, Arstadveien 19, ... more Section for Medical Image Analysis and Pattern Recognition University of Bergen, Arstadveien 19, N-5009 Bergen, Norway ... In this study we demonstrate an improved differen-tiation of the most common tissue types in the human brain and surrounding structures by quantitative ...
Magnetic resonance imaging (MRI) of the brain, followed by automated segmentation of the corpus c... more Magnetic resonance imaging (MRI) of the brain, followed by automated segmentation of the corpus callosum (CC) in midsagittal sections have important applications in both clinical neurology and neurocognitive research since the size and shape of the CC are shown to be correlated to sex, age, neurodegenerative diseases and various lateralized behavior in man. Moreover, whole head, multispectral 3D MRI recordings enable voxel-based tissue classification and estimation of total brain volumes, in addition to CC morphometric parameters. We propose a new algorithm that uses both multispectral MRI measurements (intensity values) and prior information about shape (CC template) to segment CC in midsagittal slices with very little user interaction. The algorithm has been tested on a sample of 10 subjects scanned with multispectral 3D MRI, collected for a study of dyslexia, with very good agreement between the manually traced ("true") CC outline and the detected CC outline. We conclude that the proposed method for CC segmentation is promising for clinical use when multispectral MR images are recorded. 10 multispectral 3D whole-brain acquisitions and a comparison to manual tracings of the corpus callosum, are reported in Section 4. Section 5 concludes the paper with a
The authors demonstrate an improved differentiation of the most common tissue types in the human ... more The authors demonstrate an improved differentiation of the most common tissue types in the human brain and surrounding structures by quantitative validation using multispectral analysis of magnetic resonance images. This is made possible by a combination of a special training technique and an increase in the number of magnetic resonance channel images with different pulse acquisition parameters. The authors give a description of the tissue-specific multivariate statistical distributions of the pixel intensity values and discuss how their properties may be explored to improve the statistical modeling further. A statistical method to estimate the tissue-specific longitudinal and transverse relaxation times is also given. It is concluded that multispectral analysis of magnetic resonance images is a valuable tool to recognize the most common normal tissue types in the brain and surrounding structures.
Medical image registration can be formulated as a tissue deformation problem, where parameter est... more Medical image registration can be formulated as a tissue deformation problem, where parameter estimation methods are used to obtain the inverse deformation. However, there is limited knowledge about the ability to recover an unknown deformation. The main objective of this study is to estimate the quality of a restored deformation field obtained from image registration of dynamic MR sequences. Methods: We investigate the behavior of forward deformation models of various complexities. Further, we study the accuracy of restored inverse deformations generated by image registration. Results: We found that the choice of 1) heterogeneous tissue parameters and 2) a poroelastic (instead of elastic) model had significant impact on the forward deformation. In the image registration problem, both 1) and 2) were found not to be significant. Here, the presence of image features were dominating the performance. We also found that existing algorithms will align images with high precision while at the same time obtain a deformation field with a relative error of 40%. Conclusion: Image registration can only moderately well restore the true deformation field. Still, estimation of volume changes instead of deformation fields can be fairly accurate and may represent a proxy for variations in tissue characteristics. Volume changes remain essentially unchanged under choice of discretization and the prevalence of pronounced image features. Significance: We suggest that image registration of high-contrast MR images has potential to be used as a tool to produce imaging biomarkers sensitive to pathology affecting tissue stiffness.
Background Tumor burden assessment is essential for radiation therapy (RT), treatment response ev... more Background Tumor burden assessment is essential for radiation therapy (RT), treatment response evaluation, and clinical decision-making. However, manual tumor delineation remains laborious and challenging due to radiological complexity. The objective of this study was to investigate the feasibility of the HD-GLIO tool, an ensemble of pre-trained deep learning models based on the nnUNet-algorithm, for tumor segmentation, response prediction, and its potential for clinical deployment. Methods We analyzed the predicted contrast-enhanced (CE) and non-enhancing (NE) HD-GLIO output in 49 multi-parametric MRI examinations from 23 grade-4 glioma patients. The volumes were retrospectively compared to corresponding manual delineations by 2 independent operators, before prospectively testing the feasibility of clinical deployment of HD-GLIO-output to a RT setting. Results For CE, median Dice scores were 0.81 (95% CI 0.71–0.83) and 0.82 (95% CI 0.74–0.84) for operator-1 and operator-2, respecti...
BackgroundGlioblastoma (GBM) is an aggressive malignant brain tumor where median survival is appr... more BackgroundGlioblastoma (GBM) is an aggressive malignant brain tumor where median survival is approximately 15 months after best available multimodal treatment. Recurrence is inevitable, largely due to O6 methylguanine DNA methyltransferase (MGMT) that renders the tumors resistant to temozolomide (TMZ). We hypothesized that pretreatment with bortezomib (BTZ) 48 hours prior to TMZ to deplete MGMT levels would be safe and tolerated by patients with recurrent GBM harboring unmethylated MGMT promoter. The secondary objective was to investigate whether 26S proteasome blockade may enhance differentiation of cytotoxic immune subsets to impact treatment responses measured by radiological criteria and clinical outcomes.MethodsTen patients received intravenous BTZ 1.3 mg/m2 on days 1, 4, and 7 during each 4th weekly TMZ‐chemotherapy starting on day 3 and escalated from 150 mg/m2 per oral 5 days/wk via 175 to 200 mg/m2 in cycles 1, 2, and 3, respectively. Adverse events and quality of life were...
What has happened in machine learning lately, and what does it mean for the future of medical ima... more What has happened in machine learning lately, and what does it mean for the future of medical image analysis? Machine learning has witnessed a tremendous amount of attention over the last few years. The current boom started around 2009 when so-called deep artificial neural networks began outperforming other established models on a number of important benchmarks. Deep neural networks are now the state-of-the-art machine learning models across a variety of areas, from image analysis to natural language processing, and widely deployed in academia and industry. These developments have a huge potential for medical imaging technology, medical data analysis, medical diagnostics and healthcare in general, slowly being realized. We provide a short overview of recent advances and some associated challenges in machine learning applied to medical image processing and image analysis. As this has become a very broad and fast expanding field we will not survey the entire landscape of applications, but put particular focus on deep learning in MRI. Our aim is threefold: (i) give a brief introduction to deep learning with pointers to core references; (ii) indicate how deep learning has been applied to the entire MRI processing chain, from acquisition to image retrieval, from segmentation to disease prediction; (iii) provide a starting point for people interested in experimenting and perhaps contributing to the field of deep learning for medical imaging by pointing out good educational resources, state-of-the-art open-source code, and interesting sources of data and problems related medical imaging.
Skull stripping in brain imaging is the removal of the parts of images corresponding to non-brain... more Skull stripping in brain imaging is the removal of the parts of images corresponding to non-brain tissue. Fast and accurate skull stripping is a crucial step for numerous medical brain imaging applications, e.g. registration, segmentation and feature extraction, as it eases subsequent image processing steps. In this work, we propose and compare two novel skull stripping methods based on 2D and 3D convolutional neural networks trained on a large, heterogeneous collection of 2777 clinical 3D T1-weighted MRI images from 1681 healthy subjects. We investigated the performance of the models by testing them on 927 images from 324 subjects set aside from our collection of data, in addition to images from an independent, large brain imaging study: the IXI dataset (n = 556). Our models achieved mean Dice scores higher than 0.978 and Jaccard indices higher than 0.957 on all tests sets, making predictions on new unseen brain MR images in approximately 1.4s for the 3D model and 12.4s for the 2D ...
2016 Signal Processing: Algorithms, Architectures, Arrangements, and Applications (SPA), 2016
This paper describes a pilot study aimed at evaluating the usefulness of texture parameters compu... more This paper describes a pilot study aimed at evaluating the usefulness of texture parameters computed from blood pharmacokinetic maps as biomarkers of cancers. The maps are extracted from DCE MRI images which, in turn, visualize blood perfusion in tissue by means of a contrast medium. Each voxel of a DCE image is characterised by a curve in time domain (enhancement signal). The shape of such curves is characteristic to different tissues and is quantified by fitting a proposed parameterized signal model. Different-modality MR images were acquired for a small number of patients, prior to diagnosed cancer surgical extraction. Six images were collected for each patient, four of them were distributions of blood pharmacokinetic model parameters and two — ADC and VIBE MR images. The extracted carcinoma tissue for each patient was investigated by pathologists. This histological grade provided labels to images for supervised data exploration. Both unsupervised (clustering) and supervised (LDA) experiments demonstrated the image texture can be a promising biomarker to reflect correlations between tumour tissue appearance at structural and functional MRI and the corresponding histological characteristics.
A reduction or reversal of the normal leftward asymmetry of the planum temporale (PT) has been cl... more A reduction or reversal of the normal leftward asymmetry of the planum temporale (PT) has been claimed to be typical of dyslexia, although some recent studies have challenged this view. In a population-based study of 20 right-handed dyslexic boys and 20 matched controls, we have measured the PT and the adjacent planum parietale (PP) region in sagittal magnetic resonance images. For the PT, mean left and right areas and asymmetry coefficients were compared. Since a PP area often could not be identified in one or both hemispheres, a qualitative comparison was used for this region. The total planar area (sum of PT and PP) was also compared between the two groups. A dichotic listening (DL) test with consonant-vowel syllables was administered to assess functional asymmetry of language. The results showed a mean leftward PT asymmetry in both the dyslexic and the control group, with no significant difference for the degree of PT asymmetry. Planned comparisons revealed however, a trend towards smaller left PT in the dyslexic group. In control children, but not in the dyslexic children, a significant correlation between PT asymmetry and reading was observed. A mean leftward asymmetry was also found for the total planar area, with no difference between the groups for the degree of asymmetry. Significantly fewer dyslexic children than control children showed a rightward asymmetry for the PP region. Both groups showed a normal right ear advantage on the DL task, with no significant difference for DL asymmetry. No significant correlation was observed between PT asymmetry and DL asymmetry. The present populationbased study adds to recent reports of normal PT asymmetry in dyslexia, but indicates that subtle morphological abnormalities in the left planar area may be present in this condition.
In the present study, we investigated differences between dyslexic and normal reading children in... more In the present study, we investigated differences between dyslexic and normal reading children in asymmetry of the planum temporale area in the upper posterior part of the temporal lobe and dichotic listening performance to consonant-vowel syllables. The current study was an extension of previous studies in our laboratory on the same participants, now including also girls and left-handers. There were 20 boys and 3 girls in the dyslexic group and 19 boys and 4 girls in the normal reading group. The age of the participants was 10-12 years for both groups. The participants were screened from a population of 950 students in the fourth school grade in the greater Bergen district. The planum temporale area was measured in sagittal magnetic resonance (MR) images. Mean left and right area and asymmetry index were compared between the groups. Dichotic presentations of consonant-vowel syllables made it possible to separately probe left and right hemisphere phonological function, and to correlate this with planum temporale area. The results showed a significantly larger left than right planum temporale area for both groups. However, while the right planum temporale area was similar for the dyslexic and control groups, the left planum temporale was significantly (one-tailed t-test) smaller in the dyslexic group. Both groups also showed a significant right ear advantage to the consonant-vowel syllables in the dichotic listening test. The relation between planum temporale and dichotic listening asymmetry showed a significant correlation for the dyslexic group only, indicating a positive relation between brain structure and function in dyslexic children. The results are discussed in terms of important subject characteristics with regard to brain markers of dyslexia.
Cellular multiparametric magnetic resonance imaging (MRI) provided an in vivo visualisation of ne... more Cellular multiparametric magnetic resonance imaging (MRI) provided an in vivo visualisation of neural stem cells' (NSCs) tropism for gliomas in the rat brain. NSCs were magnetically labelled in vitro with the bimodal gadolinium-based contrast agent, Gadolinium Rhodamine Dextran (GRID), and injected into the contralateral hemisphere to the developing tumour. Contrast-tonoise measurements showed that GRID-labelled cells induced a signal attenuation on both T2-, T2*-weighted images, and a modest signal gain on T1-weighted images. Tumour development and progression were longitudinally monitored in vivo by serial MR scanning. Measurements of tumour volume and tumour progression over time in terms of tumour doubling time showed a tendency towards a reduced tumour growth in NSC-treated animals. MR findings of migration and infiltration of tumours by labelled NSCs were corroborated with immunohistopathology, where labelled cells were detected in the corpus callosum at the tumour border and dispersed in the solid tumour tissue. Immunohistopathology also revealed that macrophages invaded the tumour tissue and in some cases engulfed GRID-labelled stem cells. No significant difference in macrophage recruitment between NSC-treated and vehicle-treated animals were detected, indicating that magnetically labelled NSC do not increase macrophage invasion of tumour tissue. Our findings demonstrate that cellular multiparametric MRI provides a valuable tool for in vivo dynamic monitoring of tumour-directed neural stem cell migration as well as therapeutic efficacy.
Uploads
Papers by Arvid Lundervold