Using deep learning methods to process text data from thyroid ultrasound

Abstract

The purpose of this article is to study various intelligent approaches to processing Russian-language textual medical information obtained as a result of ultrasound of the thyroid gland, solving problems of classifying diseases according to the EU-TIRADS system and generating a doctor’s conclusion based on the description of the disease. As part of the research, a machine learning pipeline was developed, including the stages of data preprocessing and model training. Transformer and hybrid architectures have been used to design deep learning models. The paper proposes methods for preprocessing unstructured medical descriptions to adapt them to the required format of the tasks being solved. The results obtained during the study showed that when solving a classification problem, achieving stable and high results using neural network architectures is possible only with careful selection of hyperparameters and taking into account their mutual influence. When solving the problem of generating ultrasound doctor's reports, transformer architectures and large language models show good results on large volumes of data. The proposed solution within the framework of the “Intelligent Ultrasound Physician Assistant” software package will automate the doctor’s work and improve the quality of diagnosis.