David Rodriguez

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.

Background and Purpose: Intracranial hemorrhage (ICH) is a common life‐threatening condition that must be rapidly diagnosed and treated. However, there is still a lack of consensus regarding treatment, driven to some extent by prognostic uncertainty. While several prediction models for ICH detection have already been published, here we present a deep learning predictive model for ICH prognosis.Methods: We included patients with ICH (n = 262), and we trained a custom model for the classification of patients into poor prognosis and good prognosis, using a hybrid input consisting of brain CT images and other clinical variables. We compared it with two other models, one trained with images only (I‐model) and the other with tabular data only (D‐model).Results: Our hybrid model achieved an area under the receiver operating characteristic curve (AUC) of .924 (95% confidence interval [CI]: .831‐.986), and an accuracy of .861 (95% CI: .760‐.960). The I‐ and D‐models achieved an AUC of .763 (95% CI: .622‐.902) and .746 (95% CI: .598‐.876), respectively.Conclusions: The proposed hybrid model was able to accurately classify patients into good and poor prognosis. To the best of our knowledge, this is the first ICH prognosis prediction deep learning model. We concluded that deep learning can be applied for prognosis prediction in ICH that could have a great impact on clinical decision‐making. Further, hybrid inputs could be a promising technique for deep learning in medical imaging.

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.

In December 2019, a new coronavirus known as 2019-nCoV emerged in Wuhan, China. The virus has spread globally and the infection was declared pandemic in March 2020. Although most cases of coronavirus disease 2019 (COVID-19) are mild, some of them rapidly develop acute respiratory distress syndrome. In the clinical management, chest X-rays (CXR) are essential, but the evaluation of COVID-19 CXR could be a challenge. In this context, we developed COVID-19 TRAINING, a free Web application for training on the evaluation of COVID-19 CXR. The application included 196 CXR belonging to three categories: non-pathological, pathological compatible with COVID-19, and pathological non-compatible with COVID-19. On the training screen, images were shown to the users and they chose a diagnosis among those three possibilities. At any time, users could finish the training session and be evaluated through the estimation of their diagnostic accuracy values: sensitivity, specificity, predictive values, and global accuracy. Images were hand-labeled by four thoracic radiologists. Average values for sensitivity, specificity, and global accuracy were .72, .64, and .68. Users who achieved better sensitivity registered less specificity (p < .0001) and those with higher specificity decreased their sensitivity (p < .0001). Users who sent more answers achieved better accuracy (p = .0002). The application COVID-19 TRAINING provides a revolutionary tool to learn the necessary skills to evaluate COVID-19 on CXR. Diagnosis training applications could provide a new original manner of evaluation for medical professionals based on their diagnostic accuracy values, and an efficient method to collect valuable data for research purposes.

Brain MRI atlases may be used to characterise brain structural changes across the life course. Atlases have important applications in research, e.g., as registration and segmentation targets to underpin image analysis in population imaging studies, and potentially in future in clinical practice, e.g., as templates for identifying brain structural changes out with normal limits, and increasingly for use in surgical planning. However, there are several caveats and limitations which must be considered before successfully applying brain MRI atlases to research and clinical problems. For example, the influential Talairach & Tournoux atlas was derived from a single fixed cadaveric brain from an elderly female with limited clinical information, yet is the basis of many modern atlases and is often used to report locations of functional activation. We systematically review currently available whole brain structural MRI atlases with particular reference to the implications for population imaging through to emerging clinical practice. We found 66 whole brain structural MRI atlases world-wide. The vast majority were based on T1, T2 and/or proton density (PD) structural sequences, had been derived using parametric statistics (inappropriate for brain volume distributions), had limited supporting clinical or cognitive data, and included few younger (>5 years and 60 years) subjects. To successfully characterise brain structural features and their changes across different stages of life, we conclude that whole brain structural MRI atlases should include: more subjects at the upper and lower extremes of age; additional structural sequences, including fluid attenuation inversion recovery (FLAIR) and T2* sequences; a range of appropriate statistics, e.g., rank-based or nonparametric; and detailed cognitive and clinical profiles of the included subjects in order to increase the relevance and utility of these atlases.

A framework for building an infrastructure that semantically integrates, archives, and reuses data for various research purposes in human brain imaging remains critical. In particular, problems of aligning technical, clinical, and professional systems in order to facilitate data sharing are a recurring issue in brain imaging. However, large samples of well-characterized images with detailed metadata are increasingly needed. This paper outlines the experience of the NeuroGrid Stroke Exemplar and further work in the Brain Research Imaging Centre and Stroke Trials Unit in developing an infrastructure that facilitates the linkage, archiving, and reuse of imaging data from stroke patients for large-scale clinical and epidemiological studies. We examined data from 12 past stroke projects carried out over the past two decades in our center and two large trials with 329 centers. We assessed previously published schemas and those developed specifically for large multicentre ischemic and hemorrhagic stroke treatment trials. We then developed our own harmonized and integrated schema and database with a web-based interface system, Longitudinal Online Research and Imaging System (LORIS), aiming to be flexible and adaptable to future trials and observational studies. We then linked image and metadata from 3,079 patients acquired in stroke research in one center in a 14-year period (1996–2010) with prospective central hospital health statistics to obtain long-term follow-up. Our integrated database includes 3,079 subjects and over 550 federated and searchable data items including imaging details, medical history, and examination, stroke, and laboratory details, which map to large multicentre stroke trials with imaging data from over 10,000 patients from 30 countries. The central linkage identified 879 of 3,079 patients had died, 525 had recurrent strokes, and 291 developed dementia during up to a 19-year period (range = 0–19; median = 9.04; IQR = 12.17) of follow-up, demonstrating its utility. The core metadata schema has benefited from extensive development in large clinical trials. Further trials’ data can now be added. It provides an opportunity to crosslink and reuse data for a range of large-scale stroke brain imaging clinical and research purposes including developing data analytics models for research into common brain diseases and their consequences.