Developing an app to interpret chest X-rays to support the diagnosis of respiratory pathology with artificial intelligence

• Autor(a) principal: Elkins, Andrew
• Data de Publicação: 2020
• Outros Autores: Freitas, Felipe F., Sanz, Verónica
• Tipo de documento: Artigo
• Idioma: eng
• Título da fonte: Repositório Científico de Acesso Aberto de Portugal (Repositórios Cientìficos)
• Texto Completo: http://hdl.handle.net/10773/29880
• topic: Chest X-ray; Pulmonary disease; Deep learning; Computer vision; Diagnosis app

Background: Medical images, including results from X-rays, are an integral part of medical diagnosis. Their interpretation requires an experienced radiologist. One of the main problems in developing countries is access to timely medical diagnosis. Lack of investment in health care infrastructure, geographical isolation and shortage of trained specialists are common obstacles to providing adequate health care in many areas of the world. In this work we show how to build and deploy a Deep Learning computer vision application for the classification of 14 common thorax disease using X-rays images. Methods: We make use of the FAST.AI and pytorch framework to create and train the DenseNet-121 model to classify the X-ray images from the ChestX-ray14 data set which contains 112,120 frontal-view X-ray images of 30,805 unique patients. After training and validate our model we create a web-app using Heroku, this web-app can be accessed by any mobile device with internet connection. Results: We obtained 70% for detecting pneumothorax for the one-vs-all task. Meanwhile, for the multilabel-multiclass task we are able to achieve state-of-the-art accuracy with fewer epochs, reducing drastically the training time of the model. We also demonstrate the feature localization of our model by using the Grad-CAM methodologies, feature which can be useful for early diagnostic of dangerous illnesses. Conclusions: In this work we present our study of the use of machine learning techniques to identify diseases using X-ray information. We have used the new framework of Fast.AI, and imported the resulting model to an app which can be tested by any user. The app has an intuitive interface where the user can upload an image and obtain a likelihood for the given image be classified as one of the 14 labeled diseases. This classification could assist diagnosis by medical providers and broaden access to medical services to remote areas.