Erin Beveridge

Detection and location of 127 anatomical landmarks in diverse CT datasets

The automatic detection and localization of anatomical landmarks has wide application, including intra and interpatient registration, study location and navigation, and the targeting of specialized algorithms. In this paper, we demonstrate the automatic detection and localization of 127 anatomically defined landmarks distributed throughout the body, excluding arms. Landmarks are defined on the skeleton, vasculature and major organs. Our approach builds on the classification forests method,1 using this classifier with simple image features which can be efficiently computed. For the training and validation of the method we have used 369 CT volumes on which radiographers and anatomists have marked ground truth (GT) - that is the locations of all defined landmarks occurring in that volume. A particular challenge is to deal with the wide diversity of datasets encountered in radiology practice. These include data from all major scanner manufacturers, different extents covering single and multiple body compartments, truncated cardiac acquisitions, with and without contrast. Cases with stents and catheters are also represented. Validation is by a leave-one-out method, which we show can be efficiently implemented in the context of decision forest methods. Mean location accuracy of detected landmarks is 13.45mm overall; execution time averages 7s per volume on a modern server machine. We also present localization ROC analysis to characterize detection accuracy - that is to decide if a landmark is or is not present in a given dataset.

Computed tomography lung iodine contrast mapping by image registration and subtraction

Pulmonary embolism (PE) is a relatively common and potentially life threatening disease, affecting around 600,000 people annually in the United States alone. Prompt treatment using anticoagulants is effective and saves lives, but unnecessary treatment risks life threatening haemorrhage. The specificity of any diagnostic test for PE is therefore as important as its sensitivity. Computed tomography (CT) angiography is routinely used to diagnose PE. However, there are concerns it may over-report the condition. Additional information about the severity of an occlusion can be obtained from an iodine contrast map that represents tissue perfusion. Such maps tend to be derived from dual-energy CT acquisitions. However, they may also be calculated by subtracting pre- and post-contrast CT scans. Indeed, there are technical advantages to such a subtraction approach, including better contrast-to-noise ratio for the same radiation dose, and bone suppression. However, subtraction relies on accurate image registration. This paper presents a framework for the automatic alignment of pre- and post-contrast lung volumes prior to subtraction. The registration accuracy is evaluated for seven subjects for whom pre- and post-contrast helical CT scans were acquired using a Toshiba Aquilion ONE scanner. One hundred corresponding points were annotated on the pre- and post-contrast scans, distributed throughout the lung volume. Surface-to-surface error distances were also calculated from lung segmentations. Prior to registration the mean Euclidean landmark alignment error was 2.57mm (range 1.43–4.34 mm), and following registration the mean error was 0.54mm (range 0.44–0.64 mm). The mean surface error distance was 1.89mm before registration and 0.47mm after registration. There was a commensurate reduction in visual artefacts following registration. In conclusion, a framework for pre- and post-contrast lung registration has been developed that is sufficiently accurate for lung subtraction iodine mapping.

Thrombus Detection in CT Brain Scans using a Convolutional Neural Network.

Automatic detection and measurement of thrombi may expedite clinical workflow in the treatment planning stage. Nevertheless it is a challenging task on non-contrast computed tomography due to the subtlety of the pathological intensity changes, which are further confounded by the appearance of vascular calcification (common in ageing brains). In this paper we propose a 3D Convolutional Neural Network architecture to detect these subtle signs of stroke. The architecture is designed to exploit contralateral features and anatomical atlas information. We use 122 CT volumes equally split into training and testing to validate our method, achieving a ROC AUC of 0.996 and a Precision-Recall AUC of 0.563 in a voxel-level evaluation. The results are not yet at a level for routine clinical use, but they are encouraging.

Context-Aware Convolutional Neural Networks for Stroke Sign Detection in Non-contrast CT Scans

Detection of acute stroke signs in non-contrast CT images is a challenging task. The intensity and texture variations in pathological regions are subtle and can be confounded by normal physiological changes or by old lesions. In this paper we investigate the use of contextual information for stroke sign detection. In particular, the appearance of the contralateral anatomy and the atlas-encoded spatial location are incorporated into a Convolutional Neural Network (CNN) architecture. CNNs are trained separately for the detection of dense vessels and of ischaemia. The network performance is evaluated on 170 datasets by cross-validation. We find that atlas location is important for dense vessel detection, but is less useful for ischaemia, whereas bilateral comparison is crucial for detection of ischaemia.

Automatic multi-parametric MR registration method using mutual information based on adaptive asymmetric k-means binning

Multi-parametric MR image registration combines different imaging sequences to enhance visualisation and analysis. However, alignment of the different acquisitions is challenging, due to contrast-dependent anatomical information and abundant artefacts. For two decades, voxel-based registration has been dominated by methods based on mutual information, calculated from the joint image histogram. In this paper, we propose a modified framework - based on an asymmetric cluster-to-image mutual information metric - that increases registration speed and robustness. A new parameter, the homogeneous dynamic intensity range, is used to determine to which image clustering is applied. The framework also includes a semi-automatic 3D region of interest, multi-resolution wavelet decomposition, and particle swarm optimization. Performance of the framework, and its individual components, were evaluated on two diverse datasets, comprising cardiac and neonatal brain datasets. The results demonstrated the method was more robust and accurate than mutual information alone.

Evaluation of Dense Vessel Detection in NCCT Scans

Automatic detection and measurement of dense vessels may enhance the clinical workflow for treatment triage in acute ischemic stroke. In this paper we use a 3D Convolutional Neural Network, which incorporates anatomical atlas information and bilateral comparison, to detect dense vessels. We use 112 non-contrast computed tomography (NCCT) scans for training of the detector and 58 scans for evaluation of its performance. We compare automatic dense vessel detection to identification of the dense vessels by clinical researchers in NCCT and computed tomography angiography (CTA). The automatic system is able to detect dense vessel in NCCT scans, however it shows lower specificity in relation to CTA than clinical experts.

A Comparison of Texture Features Versus Deep Learning for Image Classification in Interstitial Lung Disease

Interstitial lung disease (ILD) is a multifactorial condition that is difficult to diagnose. High-resolution computed tomography (CT) is commonly the imaging modality of choice, as it enables detection and mapping of distinctive pathological patterns. The distribution of these patterns gives clues as to the correct histological diagnosis. This paper compares two approaches to detecting these complex patterns: “man-made” features, based on classical handcrafted texture descriptors, and “machine-made” features, built with deep learning convolutional neural networks (CNNs). The two paradigms are evaluated on scans from 132 subjects, derived from two public databases of high resolution ILD CT images and associated expert annotations. Five specific tissue patterns are included: healthy, emphysema, fibrosis, ground glass opacity, and micronodules. The subjects are divided into training, validation and test groups. On the validation data the best handcrafted solution achieves a class assignment accuracy of 76.0%, compared with the best deep learning accuracy of 79.0%. For the test group, which was not used during development and only tested once, the handcrafted method achieves 65.5%, compared with the CNN accuracy of 69.9%. The results indicate that deep learning CNNs can outperform traditional texture measures, even on a low-level texture classification task such as this.

Evaluation of an Automatic ASPECT Scoring System for Acute Stroke in Non-Contrast CT

Determining the severity of ischemic stroke in non-contrast CT is a difficult problem due to a low signal to noise ratio. This leads to variable interpretation of ischemic stroke severity. We investigate the level of agreement between four methods including the use of an automated system with the aim of identifying early ischemic changes within the brain. For the evaluation we divide the middle cerebral artery territory of each hemisphere into ten regions defined according to the Alberta Stroke Programme Early CT Score (ASPECTS). The automatic system uses a specialised Convolutional Neural Network (CNN) based regressor to produce voxel-level confidence masks of which voxels are suspected as showing early ischemic change and from this we compute the score. Additionally, we obtain the score from three other methods that involved trained human graders. We compare the level of agreement between these methods at both a patient level and a territory level through Simultaneous Truth and Performance Level Estimation (STAPLE) and Cohen’s kappa coefficient. We analyse possible causes of disagreement between the methods and statistically validate the performance of the CNN model against the performance of clinical staff. We find that the CNN produces scores that correlate the greatest with its training data at the patient level, but the training data could be improved to strengthen the correlation with the professional standard.

Arterial tree tracking from anatomical landmarks in magnetic resonance angiography scans

This paper reports on arterial tree tracking in fourteen Contrast Enhanced MRA volumetric scans, given the positions of a predefined set of vascular landmarks, by using the A* algorithm to find the optimal path for each vessel based on voxel intensity and a learnt vascular probability atlas. The algorithm is intended for use in conjunction with an automatic landmark detection step, to enable fully automatic arterial tree tracking. The scan is filtered to give two further images using the top-hat transform with 4mm and 8mm cubic structuring elements. Vessels are then tracked independently on the scan in which the vessel of interest is best enhanced, as determined from knowledge of typical vessel diameter and surrounding structures. A vascular probability atlas modelling expected vessel location and orientation is constructed by non-rigidly registering the training scans to the test scan using a 3D thin plate spline to match landmark correspondences, and employing kernel density estimation with the ground truth center line points to form a probability density distribution. Threshold estimation by histogram analysis is used to segment background from vessel intensities. The A* algorithm is run using a linear cost function constructed from the threshold and the vascular atlas prior. Tracking results are presented for all major arteries excluding those in the upper limbs. An improvement was observed when tracking was informed by contextual information, with particular benefit for peripheral vessels.