Medical Prescription Recognition Using Heuristic Clustering and Similarity Search

The necessity to convert printed documents to facilitate the storage and retrieval of information is growing, particularly in the medical and healthcare industries. In our last work, we presented a method to extract prescriptions from images using CRAFT and TESSERACT so that patients could quickly save and check up on their pharmaceutical use information. However, the slow processing speed and the limited number of medication names lead to it being impractical. Based on this model structure, a new system is introduced, using bounding box clustering heuristics to detect the featured text areas, before employing VietOCR tool to identify the texts in prescription images. Simultaneously, a fast and accurate technique for extracting prescriptions is developed, utilizing word embedding and the vector search algorithm. The experiment results reveal that the proposed model significantly reduces the error of the retrieved data on the two standard measures, WER and CER, prominently with CER lowered to 26.95. Furthermore, the execution time decreases from 17.81 s to an average of 3.64 s, demonstrating the great effectiveness of our effort to improve the prior system.