Increasing the accuracy of handwriting text recognition in medical prescriptions with generative artificial intelligence

Abstract

The object of the research is a system for recognizing handwritten text in medical prescriptions. The peculiarities of handwriting, the variety of calligraphy styles, as well as the specificity of medical prescriptions, create many problems and challenges for recognition algorithms, causing errors and reducing recognition accuracy. The work presents a new system with additional components of post-processing the recognition results to increase the accuracy of the final results. An algorithm for combining words into lines and blocks is proposed, which makes it possible to group words while preserving contextual connections between them. Also, a generative neural network with a large language model is used to analyze the recognition result and correct possible errors. The results of the testing show an improvement in recognition accuracy by 0.13 %. Successful cases of generative artificial intelligence usage are analyzed, as well as examples of the results deterioration, that are related to grammatical errors in the initial input data. The obtained results show the use of generative artificial intelligence as an additional step for processing the recognition results really can improve the accuracy of text recognition systems. The results of the study can be used for further experiments to improve recognition results in other tasks related to text recognition and in related fields.