Search

Despite the pervasive growth of deep neural networks in medical image analysis, methods to monitor and assess network outputs, such as segmentation or regression, remain limited. In this paper, we introduce SMOCAM (SMOoth Conditional Attention Mask), an optimization method that reveals the specific regions of the input image taken into account by the prediction of a trained neural network. We developed SMOCAM explicitly to perform saliency analysis for complex regression tasks in 3D medical imagery. Our formulation optimises an 3D-attention mask at a given layer of a convolutional neural network (CNN). Unlike previous attempts, our method is relatively fast (40s per output) and is suitable for large data such as 3D MRI. We applied SMOCAM on a CNN that predicts Brain morphometry from 3D MRI which was trained using more than 5000 3D brain MRIs. We show that SMOCAM highlights neural network’s limitations when cases are underrepresented and in cases with large volume asymmetry.

HDR-MRI is a sophisticated non-linear image fusion technique for MRI which enhances image quality by fusing multiple contrast types into a single composite image. It offers improved outcomes in automatic segmentation and potentially in diagnostic power, but the existing technique is slow and requires accurate image co-registration in order to function reliably. In this work, a lightweight fully convolutional neural network architecture is developed with the goal of approximating HDR-MRI images in real-time. The resulting Contrast Enhancement Network (CEN) is capable of performing near-perfect ( $\text{SSIM} =0.98$ ) 2D approximations of HDR-MRI in 10ms and full 3D approximations in 1s, running two orders of magnitude faster than the original implementation. It is also able to perform the approximation ( $\text{SSIM} =0.93$ ) with only two of the three contrasts required to generate the original HDR-MRI image, while requiring no image co-registration.

Different modalities of magnetic resonance imaging (MRI) can indicate tumor-induced tissue changes from different perspectives, thus benefit brain tumor segmentation when they are considered together. Meanwhile, it is always interesting to examine the diagnosis potential from single modality, considering the cost of acquiring multi-modality images. Clinically, T1-weighted MRI is the most commonly used MR imaging modality, although it may not be the best option for contouring brain tumor. In this paper, we investigate whether synthesizing FLAIR images from T1 could help improve brain tumor segmentation from the single modality of T1. This is achieved by designing a 3D conditional Generative Adversarial Network (cGAN) for FLAIR image synthesis and a local adaptive fusion method to better depict the details of the synthesized FLAIR images. The proposed method can effectively handle the segmentation task of brain tumors that vary in appearance, size and location across samples.

Accurate segmentation of hip joint cartilage from magnetic resonance (MR) images provides a basis for obtaining morphometric data of articular cartilages for investigation of pathoanatomical conditions such as osteoarthritis. In this paper, we present an automated MR-based cartilage segmentation method using an ensemble of neural networks for the individual femoral and acetabular cartilage plates of the hip joint. The segmentation is performed in two stages with different image resolution levels for segmentation of the combined hip cartilage and separation of the individual cartilage plates, respectively. Neural networks used in both stages are trained in an over-complete manner using 20 training MR images with manual labeled images. Compared with expert manual segmentations, the automated method achieved mean Dice's similarity coefficients of 0.805, 0.766 and 0.712 for segmentation of the combined, femoral and acetabular cartilage volumes in a set of 26 testing MR images.

Quantitative medical imaging often utilizes intensity normalization based on a signal from a neighboring region. The choice of this region can substantially affect the quantification, and is potentially confounded by the presence of pathology or other limitations such as partial volume effect. In this paper we outline the desirable list of criteria for selecting a normalization region of interest, and utilize this approach for quantitative susceptibility mapping (QSM) MRI in a study of neurodegeneration. The proposed criteria includes (i) association between reference region and demographics such as age, (ii) diagnostic group separation effect in the reference region, (iii) correlation between reference and target regions, (iv) local variance in the reference region, and (v) reduced cross-sectional variance within the diagnostic groups using the reference region. The intensity normalization was then evaluated using 119 subjects with normal cognition, mild cognitive impairment and Alzheimer's disease. For QSM applications in ageing we found that normalizing by the white matter regions not only satisfies the criteria but it also provides the best separation between clinical groups in the brain nuclei regions.

Spatio-temporal atlas is a useful tool in imaging studies of neurodevelopment, which characterizes the growth of brain, and allows the monitoring of its development. The imaging of preterm and term born infants provides opportunities to develop a series of spatio-temporal atlases that track the changes during the particular period of neurodevelopment between. The aim of this paper is to develop a spatio-temporal atlas of diffusion MRI for neonatal brains between 32 to 42 weeks postmenstrual age (PMA). We subdivided the cohort consisting of preterm- and term-born infants according to their PMA at the MRI scan based on a kernel ridge regression, and generated the atlases based on Fibre Orientation Distribution (FOD) reconstruction of the diffusion data.

In vivo tau imaging with PET is a promising new modality that offers a unique insight into Alzheimer's disease (AD) pathology. 18F-AV1451 is a tau tracer currently being evaluated in both AIBL and ADNI. While MR-based quantification remains the gold standard, there is great interest in PET-only quantification techniques for use in patients who cannot undergo MRI. In this study, 3 PET-only methods (single atlas, adaptive atlas, and PCA-based atlas) are evaluated and compared to an MR-based quantification in both AIBL (94 subjects) and ADNI (87 subjects). Results show that all quantifications performed using PET-only normalization approaches did equally well, with an average quantification error around 2% in both cohorts.

A popular method for segmentation of magnetic resonance images (MRI) of the brain is to use a mixture model of tissue intensities with an underlying Markov Random Field (MRF) to incorporate spatial dependence between neighbouring voxels. Most current available mixture-MRF-based implementations require the user to fix the values of the MRF parameters. There is no clear method of choosing these values. In this paper we propose the use of maximum pseudolikelihood (MPL) estimation of the MRF parameters, which has not previously been used in the context of MRI segmentation, and compare this to an existing least-squares (LS) estimator. We compare the performance of both estimators on real brain MRI, and also to fixing the MRF parameters. We found that the MPL estimator was better able to recover expert manual segmentations than the LS estimator, as measured by Dice coefficient. Likewise, estimation by either method was superior to fixing the MRF parameters.

Thin leaves, fine stems, self-occlusion, non-rigid and slowly changing structures make plants difficult for three-dimensional (3D) scanning and reconstruction - two critical steps in automated visual phenotyping. Many current solutions such as laser scanning, structured light, and multiview stereo can struggle to acquire usable 3D models because of limitations in scanning resolution and calibration accuracy. In response, we have developed a fast, low-cost, 3D scanning platform to image plants on a rotating stage with two tilting DSLR cameras centred on the plant. This uses new methods of camera calibration and background removal to achieve high-accuracy 3D reconstruction. We assessed the system's accuracy using a 3D visual hull reconstruction algorithm applied on 2 plastic models of dicotyledonous plants, 2 sorghum plants and 2 wheat plants across different sets of tilt angles. Scan times ranged from 3 minutes (to capture 72 images using 2 tilt angles), to 30 minutes (to capture 360 images using 10 tilt angles). The leaf lengths, widths, areas and perimeters of the plastic models were measured manually and compared to measurements from the scanning system: results were within 3-4\% of each other. The 3D reconstructions obtained with the scanning system show excellent geometric agreement with all six plant specimens, even plants with thin leaves and fine stems.

Traditional shape learning of medical image data has been implemented via Principal Component Analysis (PCA). These PCA based statistical shape models batch process all shapes at once to generate a fixed model of shape variation as principal components, which may require significant computation resources for large number of shapes. This paper applies incremental PCA (IPCA) on a dataset of 728 surfaces (derived from magnetic resonance imaging examinations displaying the articulating bones of the knee joint) that can efficiently adapt to changes in training sets. After comparing the compactness and the accuracy of shape reconstruction of both batch PCA and IPCA models, our results show that IPCA produces a model comparable to batch PCA in terms of compactness and applicability to shape reconstruction, while requiring considerably shorter processing time and computer memory for computation.

Segmentation and quantitative tissue evaluation of the posterior cruciate ligament (PCL) from MRI will facilitate analyses into the morphological and biochemical changes associated with various knee injuries and conditions such as osteoarthritis(OA). In this paper, we validate a multi-scale patch-based method for automated segmentation of the PCL from multiecho spin-echo T2-map MRI of the knee acquired from 26 asymptomatic volunteers. Volume, length and T2-relaxation properties of the PCL were then estimated and validation was performed against manual segmentations of the T2-map images. We apply the method to an MR dataset of 88 patients from the osteoarthritis initiative to investigate differences in T2-properties of the PCL between knees at different stages of OA. A mean Dice's similarity coefficient of 74.4±4.2% was obtained for the PCL segmentation. Moderate and strong correlations were noted between automated and manual volume, length and median T2-values (rV=0.67, rL=0.88, rT2=0.78). Wilcoxon rank-sum tests showed no significant differences in length and median T2-values of the PCL between patients at variable stages of knee OA.

The analysis of the structural brain networks have recently gathered extensive interest due to its crucial role in unveiling the fundamental principles of the brain. The uniformity of structural networks inferred from diffusion tensor imaging across different individuals is however, unknown. This paper presents a method to infer group-wise consistent structural clusters from the connectome Laplacian graph. The spectral clustering of the cortical networks from diffusion tensor imaging was applied on 146 healthy subjects using 3 random groups, and on groups based on gender and age, to determine the optimal number of clusters. The results show six consistent sub-networks of structural connections that was validated using known cluster validity indices, showing highly reproducible clusters for random groups and groups based on gender; while, cluster differences were observed between younger and older groups in areas related to memory.

This paper presents a validation study on the application of a novel interslice interpolation technique for musculoskeletal structure segmentation of articulated joints and muscles on human magnetic resonance imaging data. The interpolation technique is based on morphological shape-based interpolation combined with intensity based voxel classification. Shape-based interpolation in the absence of the original intensity image has been investigated intensively. However, in some applications of medical image analysis, the intensity image of the slice to be interpolated is available. For example, when manual segmentation is conducted on selected slices, the segmentation on those unselected slices can be obtained by interpolation. We proposed a two- step interpolation method to utilize both the shape information in the manual segmentation and local intensity information in the image. The method was tested on segmentations of knee, hip and shoulder joint bones and hamstring muscles. The results were compared with two existing interpolation methods. Based on the calculated Dice similarity coefficient and normalized error rate, the proposed method outperformed the other two methods.

The accurate segmentation of the bone from Magnetic Resonance (MR) images of the hip is important for clinical studies and drug trials into conditions like Osteoarthritis. This paper presents an automatic segmentation scheme that utilises a deformable model robust to different field of views by training shape priors from partial and full bone surfaces. The deformable model with these priors were used to segment the hip joint within 16 unilateral 3T MR images having different field of views, so that parts of the model outside the image could be ignored fully without affecting the accuracy of the segmentation within the image. Mean and median Dice's Similarity Coefficients of 0.91 & 0.92 for the femur and 0.86 & 0.88 for one half of the pelvis were obtained using a leave-one-out approach.

The accurate segmentation of the bone and articular cartilages from magnetic resonance (MR) images of the hip is important for clinical studies and drug trials into conditions like osteoarthritis. In current studies, segmentations are obtained using time-consuming manual or semi-automatic algorithms which have high inter- and intra-observer variabilities. This paper presents an important step towards obtaining automatic and accurate segmentations of the hip cartilages, namely an approach to automatically segment the bones. The segmentation is performed using three-dimensional active shape models, which are initialized using an affine registration to an atlas. The accuracy and robustness of the approach was experimentally validated using an MR database of we VIBE, we DESS and MEDIC MR images. The (left, right) femoral and ace tabular bone segmentation had a median Dice similarity coefficient of (0.921, 0.926) and (0.830, 0.813).

This paper presents a novel approach to automatically segment the prostate (including seminal vesicles) using a surface that is actively deformed via shape and gray level models. The surface deformation process utilises the results of a multi-atlas registration approach, where training images are matched to the case image via non-rigid registration. Normalised mutual information is then used to measure the similarity between each image in the training set and the case image. The set of training images with a similarity greater than a threshold is then used to build the initialisation and the gray level model of the segmentation process. This case specific gray level model is used to deform the initial surface to more closely match the prostate boundary via normalised cross-correlation based template matching of gray level profiles. Mean and median Dice's Similarity Coefficients of 0.849 and 0.855, as well as a mean surface error of 2.11 mm, were achieved when segmenting 3T Magnetic Resonance clinical scans of fifty patients.

Recent developments in high resolution MRI scanning of the human spine are providing increasing opportunities for the development of accurate automated approaches for pathoanatomical assessment of intervertebral discs and vertebrae. We are developing a fully automated 3D segmentation approach for MRI scans of the human spine based on statistical shape analysis and template matching of grey level intensity profiles. The algorithm reported in the present study was validated on a dataset of high resolution volumetric scans of lower thoracic and lumbar spine obtained on a 3T scanner using the relatively new 3D SPACE (T2-weighted) pulse sequence, and on a dataset of axial T1-weighted scans of lumbar spine obtained on a 1.5T system. A 3D spine curve is initially extracted and used to position the statistical shape models for final segmentation. Initial validating experiments show promising results on both MRI datasets.

The human cortex is a folded ribbon of neurons with a high inter-individual variability. It is a challenging structure to study especially when measuring small changes resulting from normal aging and neurodegenerative disorders such as Alzheimer's Disease (AD). Recent studies have proposed surface based approaches for statistical population comparison of cortical changes since such approaches better cope with the surfacic nature of the cortex. In this paper we present a new multi-scale EM-ICP registration that is embedded into a surface-based approach. We compare this new registration algorithm with the shape context in the context of statistical population analysis. When comparing the cortical thickness between healthy elderly subjects to Alzheimer's disease patients, the new pipeline reduces the intra class variability while increasing the statistical power of the T-tests between both groups.

One of the main challenges in high-throughput plant data acquisition is the robust and automated analysis of the data. This includes a high-resolution 3D plant model reconstruction and an automated 3D segmentation. In this paper we present our top-down partitioning pipeline used to automatically segment high-resolution plant meshes. The proposed method produces a smart partition of the initial mesh that allows to identify the main stem, branches, and leaves of the plant. Extracted regions are then processed through the next stage of the automated analysis, which retrieves accurate plant information such as stem length, leaf width, length or area. Results involved applying our top-down approach on a prototype population of 6 cotton-plant meshes studied at 3 or 4 time points. Using our partitioning pipeline, we obtained accurate meshes segmentations for 20 plants out of the initial 22. Results validate the feasibility of an automated analysis of plant data. Future work will involve extending our approach to multiple plant varieties and using an atlas-based iterative feedback scheme to improve the 3D plant reconstruction.

Automatic segmentation of white matter hyperintensities (WMH) from T2-Weighted and FLAIR MRI is a common task that needs to be performed in the analysis of many different diseases. A method to segment the WMH is proposed whereby a local intensity model (LIM) of normal tissue is generated. WMH are detected as outliers from this model. The LIM enables an accurate modeling of intensity variations thus reducing false positives. Moreover only scans with normal tissues are required to create the model. Twelve normal scans were used to generate the LIM and validation was conducted on a set of 46 scans. Similarity indices between the proposed approach and manual segmentations were 0.59±0.15, 0.65±0.08 and 0.77±0.08 for subjects with small, moderate and large volume of lesions respectively. The proposed approach performed better than support vector machines on the same dataset and compared favorably to approaches in literature.

This study presents a novel method for the automatic segmentation of the quadratus lumborum (QL) muscle from axial magnetic resonance (MR) images using a hybrid scheme incorporating the use of non-rigid registration with probabilistic atlases (PAs) and geodesic active contours (GACs). The scheme was evaluated on an MR database of 7mm axial images of the lumbar spine from 20 subjects (fast bowlers and athletic controls). This scheme involved several steps, including (i) image pre-processing, (ii) generation of PAs for the QL, psoas (PS) and erector spinae+multifidus (ES+MT) muscles and (iii) segmentation, using 3D GACs initialized and constrained by the propagation of the PAs using non-rigid registration. Pre-processing of the images involved bias field correction based on local entropy minimization with a bicubic spline model and a reverse diffusion interpolation algorithm to increase the slice resolution to 0.98 x 0.98 x 1.75mm. The processed images were then registered (affine and non-rigid) and used to generate an average atlas. The PAs for the QL, PS and ES+MT were then generated by propagation of manual segmentations. These atlases were further analysed with specialised filtering to constrain the QL segmentation from adjacent non-muscle tissues (kidney, fat). This information was then used in 3D GACs to obtain the final segmentation of the QL. The automatic segmentation results were compared with the manual segmentations using the Dice similarity metric (DSC), with a median DSC for the right and left QL muscles of 0.78 (mean = 0.77, sd=0.07) and 0.75 (mean =0.74, sd=0.07), respectively.

Accurate cortical thickness estimation is important for the study of many neurodegenerative diseases. Amongst the approaches previously proposed in the literature, mesh based techniques typically lack computational efficiency, whereas voxel based techniques tend to be faster but are less accurate. The aim of this work is to propose a novel voxel based method using the Laplacian definition of thickness, being both accurate and computationally efficient. A subvoxel estimate of the location of the boundary is obtained through ray-casting using both partial volume estimation and the direction of the streamlines. This estimate is then used to initialise the boundary conditions when computing the length of the streamlines. The approach was validated on synthetic phantoms and real data, showing an improved accuracy and reproducibility.

Statistical shape models are widely used as a compact way of representing shape variation. Fitting a shape model to unseen data enables characterizing the data in terms of the model parameters. In this paper a Gauss-Newton optimization scheme is proposed to estimate shape model parameters of 3D surfaces using distance maps, which enables the estimation of model parameters without the requirement of point correspondence. For applications with acquisition limitations such as speed and cost, this formulation enables the fitting of a statistical shape model to arbitrarily sampled data. The method is applied to a database of 3D surfaces from a section of the porcine pelvic bone extracted from 33 CT scans. A leave-one-out validation shows that the parameters of the first 3 modes of the shape model can be predicted with a mean difference within [-0.01,0.02] from the true mean, with a standard deviation less than 0.34.

This paper presents an automated segmentation approach for MR images of the knee bones. The bones are the first stage of a segmentation system for the knee, primarily aimed at the automated segmentation of the cartilages. The segmentation is performed using 3D active shape models (ASM), which are initialized using an affine registration to an atlas. The 3D ASMs of the bones are created automatically using a point distribution model optimization scheme. The accuracy and robustness of the segmentation approach was experimentally validated using an MR database of fat suppressed spoiled gradient recall images.

Deformable models are a highly accurate and flexible approach to segmenting structures in medical images. The primary drawback of deformable models is that they are sensitive to initialisation, with accurate and robust results often requiring initialisation close to the true object in the image. Automatically obtaining a good initialisation is problematic for many structures in the body. The cartilages of the knee are a thin elastic material that cover the ends of the bone, absorbing shock and allowing smooth movement. The degeneration of these cartilages characterize the progression of osteoarthritis. The state of the art in the segmentation of the cartilage are 2D semi-automated algorithms. These algorithms require significant time and supervison by a clinical expert, so the development of an automatic segmentation algorithm for the cartilages is an important clinical goal. In this paper we present an approach towards this goal that allows us to automatically providing a good initialisation for deformable models of the patella cartilage, by utilising the strong spatial relationship of the cartilage to the underlying bone.