Framework to Support Radiologist Personnel in the Diagnosis of Diseases in Medical Images Using Deep Learning and Personalized DICOM Tags

Abstract

Technological innovations in the healthcare field have allowed medical images to be widely used in the diagnostic care of patients since medical personnel can analyze different body organs to identify any disease through these images. The analysis of these images is entirely within the domain of the specialist, who, based on his/her experience, interprets them and discloses the results to the patient. This paper presents the architecture of a framework that seeks to support the decision-making of medical personnel regarding the diagnosis of diseases. The framework integrates custom tags in the metadata of Digital Imaging and Communications in Medicine(DICOM) files. The tags contain the classification results of supervised learning models. Different convolutional neural network (CNN) architectures trained on medical images were developed using transfer learning and existing pre-trained CNNs to evaluate the framework’s performance. A web viewer was also developed to show medical personnel the custom tags. Due to the characteristics of the framework, its use could be extended to patients so that they could obtain a preliminary diagnosis and go to the doctor as soon as possible, which could be crucial.