A retrospective study of deep learning generalization across two centers and multiple models of X-ray devices using COVID-19 chest-X rays.

Generalization of deep learning (DL) algorithms is critical for the secure implementation of computer-aided diagnosis systems in clinical practice. However, broad generalization remains to be a challenge in machine learning. This research aims to identify and study potential factors that can affect the internal validation and generalization of DL networks, namely the institution where the images come from, the image processing applied by the X-ray device, and the type of response function of the X-ray device. For these purposes, a pre-trained convolutional neural network (CNN) (VGG16) was trained three times for classifying COVID-19 and control chest radiographs with the same hyperparameters, but using different combinations of data acquired in two institutions by three different X-ray device manufacturers. Regarding internal validation, the addition of images from an external institution to the training set did not modify the algorithm's internal performance, however, the inclusion of images acquired by a device from a different manufacturer decreased the performance up to 8% (p < 0.05). In contrast, generalization across institutions and X-ray devices with the same type of response function was achieved. Nonetheless, generalization was not observed across devices with different types of response function. This factor was the key impediment to achieving broad generalization in our research, followed by the device's image-processing and the inter-institutional differences, which both reduced generalization performance to 18.9% (p < 0.05), and 9.8% (p < 0.05), respectively. Finally, clustering analysis with features extracted by the CNN was performed, revealing a substantial dependence of feature values extracted by the pre-trained CNN on the X-ray device which acquired the images.

A primer on deep learning and convolutional neural networks for clinicians.

Deep learning is nowadays at the forefront of artificial intelligence. More precisely, the use of convolutional neural networks has drastically improved the learning capabilities of computer vision applications, being able to directly consider raw data without any prior feature extraction. Advanced methods in the machine learning field, such as adaptive momentum algorithms or dropout regularization, have dramatically improved the convolutional neural networks predicting ability, outperforming that of conventional fully connected neural networks. This work summarizes, in an intended didactic way, the main aspects of these cutting-edge techniques from a medical imaging perspective.

Use of brain MRI atlases to determine boundaries of age-related pathology: the importance of statistical method.

INTRODUCTION: Neurodegenerative disease diagnoses may be supported by the comparison of an individual patient's brain magnetic resonance image (MRI) with a voxel-based atlas of normal brain MRI. Most current brain MRI atlases are of young to middle-aged adults and parametric, e.g., mean ± standard deviation (SD); these atlases require data to be Gaussian. Brain MRI data, e.g., grey matter (GM) proportion images, from normal older subjects are apparently not Gaussian. We created a nonparametric and a parametric atlas of the normal limits of GM proportions in older subjects and compared their classifications of GM proportions in Alzheimer's disease (AD) patients. METHODS: Using publicly available brain MRI from 138 normal subjects and 138 subjects diagnosed with AD (all 55-90 years), we created: a mean ± SD atlas to estimate parametrically the percentile ranks and limits of normal ageing GM; and, separately, a nonparametric, rank order-based GM atlas from the same normal ageing subjects. GM images from AD patients were then classified with respect to each atlas to determine the effect statistical distributions had on classifications of proportions of GM in AD patients. RESULTS: The parametric atlas often defined the lower normal limit of the proportion of GM to be negative (which does not make sense physiologically as the lowest possible proportion is zero). Because of this, for approximately half of the AD subjects, 25-45% of voxels were classified as normal when compared to the parametric atlas; but were classified as abnormal when compared to the nonparametric atlas. These voxels were mainly concentrated in the frontal and occipital lobes. DISCUSSION: To our knowledge, we have presented the first nonparametric brain MRI atlas. In conditions where there is increasing variability in brain structure, such as in old age, nonparametric brain MRI atlases may represent the limits of normal brain structure more accurately than parametric approaches. Therefore, we conclude that the statistical method used for construction of brain MRI atlases should be selected taking into account the population and aim under study. Parametric methods are generally robust for defining central tendencies, e.g., means, of brain structure. Nonparametric methods are advisable when studying the limits of brain structure in ageing and neurodegenerative disease.

Variance in brain volume with advancing age: implications for defining the limits of normality.

BACKGROUND: Statistical models of normal ageing brain tissue volumes may support earlier diagnosis of increasingly common, yet still fatal, neurodegenerative diseases. For example, the statistically defined distribution of normal ageing brain tissue volumes may be used as a reference to assess patient volumes. To date, such models were often derived from mean values which were assumed to represent the distributions and boundaries, i.e. percentile ranks, of brain tissue volume. Since it was previously unknown, the objective of the present study was to determine if this assumption was robust, i.e. whether regression models derived from mean values accurately represented the distributions and boundaries of brain tissue volume at older ages. MATERIALS AND METHODS: We acquired T1-w magnetic resonance (MR) brain images of 227 normal and 219 Alzheimer's disease (AD) subjects (aged 55-89 years) from publicly available databanks. Using nonlinear regression within both samples, we compared mean and percentile rank estimates of whole brain tissue volume by age. RESULTS: In both the normal and AD sample, mean regression estimates of brain tissue volume often did not accurately represent percentile rank estimates (errors=-74% to 75%). In the normal sample, mean estimates generally underestimated differences in brain volume at percentile ranks below the mean. Conversely, in the AD sample, mean estimates generally underestimated differences in brain volume at percentile ranks above the mean. Differences between ages at the 5(th) percentile rank of normal subjects were ~39% greater than mean differences in the AD subjects. CONCLUSIONS: While more data are required to make true population inferences, our results indicate that mean regression estimates may not accurately represent the distributions of ageing brain tissue volumes. This suggests that percentile rank estimates will be required to robustly define the limits of brain tissue volume in normal ageing and neurodegenerative disease.

Parallel perfusion imaging processing using GPGPU.

BACKGROUND AND PURPOSE: The objective of brain perfusion quantification is to generate parametric maps of relevant hemodynamic quantities such as cerebral blood flow (CBF), cerebral blood volume (CBV) and mean transit time (MTT) that can be used in diagnosis of acute stroke. These calculations involve deconvolution operations that can be very computationally expensive when using local Arterial Input Functions (AIF). As time is vitally important in the case of acute stroke, reducing the analysis time will reduce the number of brain cells damaged and increase the potential for recovery. METHODS: GPUs originated as graphics generation dedicated co-processors, but modern GPUs have evolved to become a more general processor capable of executing scientific computations. It provides a highly parallel computing environment due to its large number of computing cores and constitutes an affordable high performance computing method. In this paper, we will present the implementation of a deconvolution algorithm for brain perfusion quantification on GPGPU (General Purpose Graphics Processor Units) using the CUDA programming model. We present the serial and parallel implementations of such algorithms and the evaluation of the performance gains using GPUs. RESULTS: Our method has gained a 5.56 and 3.75 speedup for CT and MR images respectively. CONCLUSIONS: It seems that using GPGPU is a desirable approach in perfusion imaging analysis, which does not harm the quality of cerebral hemodynamic maps but delivers results faster than the traditional computation.

An open source toolkit for medical imaging de-identification.

OBJECTIVE: Medical imaging acquired for clinical purposes can have several legitimate secondary uses in research projects and teaching libraries. No commonly accepted solution for anonymising these images exists because the amount of personal data that should be preserved varies case by case. Our objective is to provide a flexible mechanism for anonymising Digital Imaging and Communications in Medicine (DICOM) data that meets the requirements for deployment in multicentre trials. METHODS: We reviewed our current de-identification practices and defined the relevant use cases to extract the requirements for the de-identification process. We then used these requirements in the design and implementation of the toolkit. Finally, we tested the toolkit taking as a reference those requirements, including a multicentre deployment. RESULTS: The toolkit successfully anonymised DICOM data from various sources. Furthermore, it was shown that it could forward anonymous data to remote destinations, remove burned-in annotations, and add tracking information to the header. The toolkit also implements the DICOM standard confidentiality mechanism. CONCLUSION: A DICOM de-identification toolkit that facilitates the enforcement of privacy policies was developed. It is highly extensible, provides the necessary flexibility to account for different de-identification requirements and has a low adoption barrier for new users.