Selected Topic: Artificial Intelligence
Research and industry groups working with artificial neural networks can use this work as example to encourage the use of cloud-based environments as development and production tools with medical applications. Furthermore, this work is part of the ATMOSPHERE project, and will be employed to validate its use.
Convolutional Neural Networks (CNNs) have already been established as an efficient tool for a number of image processing tasks, such as, image classification, object identification, and face recognition. Through the use of CNNs, in this work, multiple CNN models will be proposed and evaluated according to their appropriateness to the interpretation of echocardiograms. The focus of the interpretation is on a specific heart condition, known as the rheumatic heart disease (RHD), which is a leading cause of heart failure in unprivileged populations.
The RHD is a heart condition caused by abnormal immune response to streptococcal infection, which is a bacteria normally associated with poor sanitation and hygiene conditions. The disease mainly attacks the mitral and aortic valves, resulting in permanent disability or death on the worst case, if not properly treated. Clinical diagnosis of RHD is dependent on auscultation and detection of valvular murmur, which is only possible when valve damage is more significant. Echocardiographic exams are an important tool for treating RHD, because it enables early detection of the disease and adequate treatment.
Furthermore, RHD is the highest worldwide cause of morbidity and mortality among heart valve diseases. One of the reasons for such is that effective detection of early RHD requires echocardiographic imaging, which can be inaccessible for certain regions (low-income and/or rural areas). As a solution, the UFMG PROVAR group proposed the use of hand-held echocardiographic devices by non-experts as a way to improve the accessibility of RHD diagnoses. Nevertheless, this solution requires a large amount of human resources. These non-experts also need to undergo a training process that enables them to recognize RHD features in echocardiograms.
In this work, we propose to improve the PROVAR approach leveraging the artificial neural network to improve the accuracy of RHD identification, while also reducing the human cost, on top of a cloud-based environment to ensure availability and data security. The use of the ATMOSPHERE environment will be essential for such endeavor, providing the trustworthiness and availability requirements for this medical application. Also, it will facilitate the development and improvement of the neural network itself, easing the optimization process for the network hyper-parameters.