Search

The cornerstone of stroke care is expedient management that varies depending on the time since stroke onset. Consequently, clinical decision making is centered on accurate knowledge of timing and often requires a radiologist to interpret Computed Tomography (CT) of the brain to confirm the occurrence and age of an event. These tasks are particularly challenging due to the subtle expression of acute ischemic lesions and the dynamic nature of their appearance. Automation efforts have not yet applied deep learning to estimate lesion age and treated these two tasks independently, so, have overlooked their inherent complementary relationship. To leverage this, we propose a novel end-to-end multi-task transformer-based network optimized for concurrent segmentation and age estimation of cerebral ischemic lesions. By utilizing gated positional self-attention and CT-specific data augmentation, the proposed method can capture long-range spatial dependencies while maintaining its ability to be trained from scratch under low-data regimes commonly found in medical imaging. Furthermore, to better combine multiple predictions, we incorporate uncertainty by utilizing quantile loss to facilitate estimating a probability density function of lesion age. The effectiveness of our model is then extensively evaluated on a clinical dataset consisting of 776 CT images from two medical centers. Experimental results demonstrate that our method obtains promising performance, with an area under the curve (AUC) of 0.933 for classifying lesion ages

Background It remains uncertain whether the drip‐and‐ship (DS) strategy (transport to the nearest primary stroke centers) or the mothership (MS) model (direct transportation to the comprehensive stroke center) is the organizational stroke model associated with the best functional outcome in acute stroke patients with large vessel occlusion. In this study, we compared the periprocedural complications and outcomes at 90 days of acute stroke patients with large vessel occlusion of the anterior circulation directly admitted to our comprehensive stroke center compared to those referred from primary stroke centers treated with mechanical thrombectomy. Methods This is a single‐center prospective observational study where patients with DS and MS were compared regarding the disability at 90 days, as assessed by applying the modified Rankin scale, the rate of successful reperfusion, and the rate of immediate complications postprocedure. Results Of 579 patients, 216 (37.30%) were MS, and 363 (62.7%) were DS. There was no difference regarding the modified Rankin scale at 3 months between the MS (36.4%) and DS (39.3%) groups of patients treated with mechanical thrombectomy ( P =0.362). There was no significant difference regarding the National Institutes of Health Stroke Scale at 24 hours, mortality at 90 days, and rate of successful recanalization postprocedure between the 2 groups. DS was associated with an increased risk of hemorrhagic transformation and symptomatic intracranial hemorrhage (odds ratio, 5.414 [95% CI, 1.572–18.644]; P =0.007). Conclusion Our single‐center study showed no difference in terms of functional independence between the DS and MS organizational paradigm. DS was associated with an increased risk of hemorrhagic transformation and symptomatic intracranial hemorrhage.

SummaryThe role of the thalamus in high-level cognition—attention, working memory (WM), rule-based learning, and decision making—remains poorly understood, especially in comparison to that of cortical frontoparietal networks [1–3]. Studies of visual thalamus have revealed important roles for pulvinar and lateral geniculate nucleus in visuospatial perception and attention [4–10] and for mediodorsal thalamus in oculomotor control [11]. Ventrolateral thalamus contains subdivisions devoted to action control as part of a circuit involving the basal ganglia [12, 13] and motor, premotor, and prefrontal cortices [14], whereas anterior thalamus forms a memory network in connection with the hippocampus [15]. This connectivity profile suggests that ventrolateral and anterior thalamus may represent a nexus between mnemonic and control functions, such as action or attentional selection. Here, we characterize the role of thalamus in the interplay between memory and visual attention. We show that ventrolateral lesions impair the influence of WM representations on attentional deployment. A subsequent fMRI study in healthy volunteers demonstrates involvement of ventrolateral and, notably, anterior thalamus in biasing attention through WM contents. To further characterize the memory types used by the thalamus to bias attention, we performed a second fMRI study that involved learning of stimulus-stimulus associations and their retrieval from long-term memory to optimize attention in search. Responses in ventrolateral and anterior thalamic nuclei tracked learning of the predictiveness of these abstract associations and their use in directing attention. These findings demonstrate a key role for human thalamus in higher-level cognition, notably, in mnemonic biasing of attention.

Lesion-symptom mapping studies are a critical component of addressing the relationship between brain and behaviour. Recent developments have yielded significant improvements in the imaging and detection of lesion profiles, but the quantification of motor outcomes is still largely performed by subjective and low-resolution standard clinical rating scales. This mismatch means than lesion-symptom mapping studies are limited in scope by scores which lack the necessary accuracy to fully quantify the subcomponents of motor function. The first study conducted aimed to develop a new automated system of motor function which addressed the limitations inherent in the clinical rating scales. A wearable fusion system was designed that included the attachment of inertial sensors to record the kinematics of upper extremity. This was combined with the novel application of mechanomyographic sensors in this field, to enable the quantification of hand/wrist function. Novel outputs were developed for this system which aimed to combine the validity of the clinical rating scales with the high accuracy of measurements possible with a wearable sensor system. This was achieved by the development of a sophisticated classification model which was trained on series of kinematic and myographic measures to classify the clinical rating scale. These classified scores were combined with a series of fine-grained clinical features derived from higher-order sensor metrics. The developed automated system graded the upper-extremity tasks of the Fugl-Meyer Assessment with a mean accuracy of 75\% for gross motor tasks and 66\% for the wrist/hand tasks. This accuracy increased to 85\% and 74\% when distinguishing between healthy and impaired function for each of these tasks. Several clinical features were computed to describe the subcomponents of upper extremity motor function. This fine-grained clinical feature set offers a novel means to complement the low resolution but well-validated standardised clinical rating scales. A second study was performed to utilise the fine-grained clinical feature set calculated in the previous study in a large-scale region-of-interest lesion-symptom mapping study. Statistically significant regions of motor dysfunction were found in the corticospinal tract and the internal capsule, which are consistent with other motor-based lesion-symptom mapping studies. In addition, the cortico-ponto-cerebellar tract was found to be statistically significant when testing with a clinical feature of hand/wrist motor function. This is a novel finding, potentially due to prior studies being limited to quantifying this subcomponent of motor function using standard clinical rating scales. These results indicate the validity and potential of the clinical feature set to provide a more detailed picture of motor dysfunction in lesion-symptom mapping studies. Open Access

Machine learning is playing a pivotal role in medical image analysis. Many algorithms based on machine learning have been applied in medical imaging to solve classification, detection, and segmentation problems. Particularly, with the wide application of deep learning approaches, the performance of medical image analysis has been significantly improved. In this thesis, we investigate machine learning methods for two key challenges in medical image analysis: The first one is segmentation of medical images. The second one is learning with weak supervision in the context of medical imaging. The first main contribution of the thesis is a series of novel approaches for image segmentation. First, we propose a framework based on multi-scale image patches and random forests to segment small vessel disease (SVD) lesions on computed tomography (CT) images. This framework is validated in terms of spatial similarity, estimated lesion volumes, visual score ratings and was compared with human experts. The results showed that the proposed framework performs as well as human experts. Second, we propose a generic convolutional neural network (CNN) architecture called the DRINet for medical image segmentation. The DRINet approach is robust in three different types of segmentation tasks, which are multi-class cerebrospinal fluid (CSF) segmentation on brain CT images, multi-organ segmentation on abdomen CT images, and multi-class tumour segmentation on brain magnetic resonance (MR) images. Finally, we propose a CNN-based framework to segment acute ischemic lesions on diffusion weighted (DW)-MR images, where the lesions are highly variable in terms of position, shape, and size. Promising results were achieved on a large clinical dataset. The second main contribution of the thesis is two novel strategies for learning with weak supervision. First, we propose a novel strategy called context restoration to make use of the images without annotations. The context restoration strategy is a proxy learning process based on the CNN, which extracts semantic features from images without using annotations. It was validated on classification, localization, and segmentation problems and was superior to existing strategies. Second, we propose a patch-based framework using multi-instance learning to distinguish normal and abnormal SVD on CT images, where there are only coarse-grained labels available. Our framework was observed to work better than classic methods and clinical practice. Open Access

Background: Motor impairment and attention deficits are common in stroke. Little is known about how a patient’s attentional capacity influences their motor function, motor-learning and recovery. This relationship may present a target for rehabilitation. This thesis aimed to: 1) survey the prevalence of attention deficits following stroke; 2) investigate the relationship between attention deficits and motor performance; 3) assess how the attention and motor profile of patients related to lesion location and disruption of functional brain networks; and 4) to develop and test a practical tool that allows for measurement and rehabilitation of attention-motor deficits in combination. Methods: Study 1: Anatomically-unselected stroke patients performed the Attention Network Task, a sensitive measure of attention, with performance related to lesion anatomy. Study 2: Stroke patients and controls were tested on a novel visuomotor tracking task, with variable distractors, using a commercially available hand-grip controller. Relationships between motor-tracking performance and distractibility were determined, as were the dependency of these behavioural measures on lesion location and functional network integrity. Study 3: A separate group of subjects performed a visumotor tracking task while functional MRI was obtained. Performance and motor-learning was related to changes in resting-state networks before and after the task. Study 4: A novel portable hand-grip and variant of the visuomotor tracking task were designed and developed for bedside assessment and rehabilitation. The novel system was tested on hemiparetic patients, and its accessibility was compared with existing mobile gaming technologies. Results: Study 1: A majority of stroke patients showed attention deficits, especially attention-control deficits; even though a far smaller proportion showed attentional-neglect on standard bedside tests. Attention-control impairments were seen equally with lesions to subcortical, premotor and prefrontal cortices. Study 2: Motor performance was closely related to attention-control performance. This was dependent upon lesion location and interference with both attention-control and motor network connectivity. Study 3: The visuomotor task influenced changes in connectivity of visuo-spatial, sensorimotor and cerebellar resting-state networks. These differed between patient and controls, and related to motor-learning. Study 4: A significantly greater proportion of hemiparetic patients – particularly those with a severe motor deficit - could engage with our novel attention-motor trainer than existing technologies. Conclusions: This work provides evidence that attention deficits frequently accompany stroke and have a significant effect on a patient’s motor ability and recovery potential. Variability in patients’ motor function can be accounted for by lesions that damage both corticospinal and attention-control systems. A novel portable electronic device, designed as part of the PhD, allows for both testing and training of motor stroke patients, for both their motor and related attention deficits. Open Access

The present thesis investigates the neural mechanisms supporting working memory (WM) guidance of visual attention, focusing on the role of the thalamus. Chapter 1 is a review of the relevant literature and sets-up the specific research aims. Chapters 2 and 3 explore the role of the thalamus on guidance of attention by WM contents. Stroke patients with focal-brain lesions performed a WM-guided search task. In valid conditions, the colour of the search target was pre-cued by the WM cue while on neutral conditions there was no cue prior to search. In invalid conditions, the WM cue specified the colour of a search distracter and the target appeared elsewhere. First, it was hypothesized that lesions to the thalamus could lead to deficits in attentional control (e.g. failing to separate irrelevant memory contents with relevant target information and leading to increased capture from WM-like distracters during invalid search conditions). An alternative hypothesis was that the thalamus may support the capture of attention by WM contents, hence thalamic patients would display little bias of attention from the WM contents, despite those contents are being maintained in memory. It was found that patients with focal-thalamic lesions especially in the ventrolateral nucleus, showed no search benefit from the valid cues on search as opposed to a control group of patients with lesions outside the thalamus and non-stroke patients. In the invalid condition, thalamic patients showed no capture by the irrelevant search item that matched the WM cue, whereas a group of healthy age-match controls exhibited the normal effect of capture by irrelevant contents held in WM. These observations suggest that lesions to the ventrolateral nucleus of the thalamus impair the capture of attention from WM contents. In Chapter 4, I aimed to establish the role of cortical structures that are known to be structurally connected with the ventrolateral nucleus of the thalamus (i.e. superior frontal gyrus) in WM guidance of attention. To do this, I investigated the effects of transcranial direct current stimulation (tDCS) of the dorsal frontal cortex in WM guidance of attention under distinct WM loads. I found that despite the effect of WM guidance of attention decreasing as WM load increased, frontal-tDCS modulated WM guidance in these conditions. We suggest that the dorsal frontal cortex forms part of a network alongside the thalamus in supporting WM guidance of attention. Finally, I conducted a functional Magnetic Resonance Imaging (fMRI) experiment (Chapter 5) with healthy volunteers to test the hypothesis that the thalamus plays a role in WM guidance when learning of abstract cue-target feature associations needs to take place for guidance of behaviour to emerge. I used four Japanese ideograms as WM cues, each associated with the colour surrounding the sought after target in the subsequent search display (valid trials). In the neutral condition, four different Japanese ideograms were presented that did not predict the colour of the target. Hence, for WM to guide attention the association between the abstract cue and the colour that surrounded the search target needed to be learned. I found that responses in the thalamus and the frontoparietal cortex displayed sensitivity to the predictiveness of the ideogram cues as WM guidance of attention emerged during learning. The findings reported in this thesis demonstrate the pivotal role of the thalamus in WM guidance of attention. Open Access