TECHNOLOGIES FOR DEVELOPING DECISION SUPPORT SYSTEMS FOR THE DIAGNOSIS OF BLOOD DISORDERS USING CONVOLUTIONAL NEURAL NETWORKS

• Main Authors: U. V. Maslikova, A. A. Supilnikov
• Format: Article
• Language: Russian
• Published: Private institution educational organization of higher education "Medical University "ReaViz" 2021-02-01
• Series: Вестник медицинского института «Реавиз»: Реабилитация, врач и здоровье
• Subjects: bone marrow microscopy; cad; decision support systems; semantics analysis, machine vision
• Online Access: https://vestnik.reaviz.ru/jour/article/view/76

In this study, we analyzed technologies for obtaining, processing, segmentation, and transmitting of microphotographs for subsequent recognition. We selected the most promising machine learning algorithms optimal for the processing of medical images, investigated the technologies of analyzing medical texts, studied the aspects of using the Watson neural network for analyzing the semantics of medical images, as well as the aspect of using the unified medical language UMLS for the needs of syndromic diagnostics for the evaluation of medical texts from medical histories in natural language. We also developed an interface for receiving,  processing, segmenting, and transmitting microphotographs to artificial neural networks and an interface for  the primary accepting and processing of microphotographs based on the OMERO medical image processing  platform. To send data online, a demo script for jupiter was prepared. An interface for transmitting medical texts  to the medical text semantics recognition system was also developed. The IBM Watson Annotator for Clinical  Data was used to perform preliminary analysis of medical texts. We created a database of medical images of  the bone marrow for neural network training. We made 3,500 color microphotographs (600×400 pixels) of  bone marrow smears at a resolution of ×600 (light microscopy; hematoxylin and eosin staining). We performed  marking of 11 types of bone marrow cells. We created a database of medical texts (167 patients, 40,000 words) to prepare a neural network. The database was stripped of all personal identifiers.