Ian Poole

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.

Objective evaluation of the correction by non-rigid registration of abdominal organ motion in low-dose 4D dynamic contrast-enhanced CT

We objectively evaluate a straightforward registration method for correcting respiration-induced movement of abdominal organs in CT perfusion studies by measuring the distributions of alignment errors between corresponding landmark pairs. We introduce the concept and describe the advantages of using the surface-normal component of distance between pairs of corresponding landmarks selected so that their surface normal is in one of the three coordinate axis directions, and show that such landmarks can be precisely placed with respect to the surface normal. Using a large population of landmark pairs on a substantial quantity of 4D dynamic contrast-enhanced CT volume data, we quantify the average alignment errors of abdominal organs that remain uncorrected by registration.

An Evaluation of Supervised, Novelty-Based and Hybrid Approaches to Fall Detection Using Silmee Accelerometer Data

Elderly people often experience a fear of falling. A reliable fall detector could increase their confidence in receiving prompt help after a fall, thus reducing their mental distress. A wearable sensors such as Toshibas Silmee device can gather accelerometer data, which can be used to detect falls. We collected data from 20 volunteers wearing Silmee during simulated falls and activities of daily living (ADL). This gave 168 fall and 375 ADL recordings. We used these recordings in three experiments conducted to compare the performance of machine learning techniques for the detection of falls from accelerometer data. These experiments evaluate supervised methods, novelty based fall detection techniques, and finally our proposed hybrid techniques which use supervised methods for feature learning, but can be applied in the context of novelty detection. We found that the best performing supervised method was the Convolutional Neural Network (CNN) and the best performing unsupervised method was the one-class Nearest Neighbour Classifier. The best performing hybrid approach resulted from a combination of the CNN and the one-class Support Vector Machine. It draws on the strengths of the CNN (appropriate feature learning) and may offer more accurate real world fall identification.

Stein's paradox and improved quadratic discrimination of real and simulated data by covariance weighting

The cross-validation error rates of quadratic discriminant classifiers were substantially reduced by weighting the off-diagonal elements of the covariance matrices by a constant less than one. The improvement was found both when real chromosome data was classified, and also when simulated multivariate normal random vectors were classified. It is shown empirically that the optimum value of the weighting constant depended on the size of the training set, and that its relative benefit was greatest when the training set was small. The optimum weight appeared to be largely independent of the number of features. The relationship of this heuristic to Steins paradox (1956) was explored, and it is shown that near-optimal values of weight could be predicted directly, thereby avoiding the need for an expensive empirical search.

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.

Optimal atlas selection using image similarities in a trained regression model to predict performance

An atlas in the context of atlas-based segmentation refers to a pre-selected image with labelled anatomical regions of interest. Atlas-based segmentation is the propagation of these labels to a novel image after both images have been registered. The goal of an atlas is to be representative of an anatomical category, but in practice there exists variability in human anatomy. One solution to maintain consistent segmentation accuracies is to use multiple atlases, with a system for selecting the most appropriate atlas at the time of segmentation. This paper describes a method for selecting an atlas using a linear regression model to predict the segmentation accuracy based on image similarity measures. It goes further to present an offline method for automatically selecting a set of atlases, representative of the training set to be used during segmentation; all of this illustrated by segmentation of the heart and kidneys in 3D CT images.

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.

Fast, Simple, Accurate Multi-atlas Segmentation of the Brain

We are concerned with the segmentation of structures within the brain particularly the gyri of the cerebral cortex, but also subcortical structures from volumetric T1-weighted MRI images. A fully automatic multi-atlas registration based segmentation approach is used to label novel data. We use a standard affine registration method combined with a small deformation (non-diffeomorphic), non-linear registration method which optimises mutual information, with a cascading set of regularisation parameters. We consistently segment 138 structures in the brain, 98 in the cortex and 40 in the sub-cortex. An overall Dice score of 0.704 and a mean surface distance of 1.106 mm is achieved in leave-one-out cross validation using all available atlases. The algorithm has been evaluated on a number of different cohorts which includes patients of different ages and scanner manufacturers, and has been shown to be robust. It is shown to have comparable accuracy to other state of the art methods using the MICCAI 2012 multi-atlas challenge benchmark, but the runtime is substantially less.

Experience of developing a cervical cytology scanning system using Gofer and Haskell

We present as a case study, the development of an application in automated image microscopy, using the functional programming languages Gofer and Haskell. Gofer is used as a formal specification language and subsequently for animation. I/O is sequenced by an I/O monad similar to that proposed for Haskell 1.3. The final implementation is in Haskell 1.2, although pragmatically, a C-coded image processing library is exploited.

Using the Functional Programming Language Haskell to Specify Image Analysis Systems

This paper demonstrates the feasibility (and discusses the advantages) of using a functional programming language as a vehicle for both formal specification and final implementation of diagnostic imaging systems. We show how the syntax of a particular language, Haskell, can be used to record implicit specifications, general purpose theorems and proofs. An example of the derivation of a constructive definition from an imperative one, notated in Haskell is given. Some practical issues concerning the implementation of imaging systems in a functional language are discussed briefly.