Text Detection and Recognition for Images of Medical Laboratory Reports With a Deep Learning Approach

Wenyuan Xue,Qingyong Li,Qiyuan Xue

doi:10.1109/access.2019.2961964

Wenyuan Xue, Qingyong Li + Show 1 more

Open Access

https://doi.org/10.1109/access.2019.2961964

Copy DOI

Abstract

The adoption of electronic health records (EHRs) is an important step in the development of modern medicine. However, complete health records are not often available during treatment because of the functional problem of the EHR system or information barriers. This paper presents a deep-learning-based approach for textual information extraction from images of medical laboratory reports, which may help physicians solve the data-sharing problem. The approach consists of two modules: text detection and recognition. In text detection, a patch-based training strategy is applied, which can achieve the recall of 99.5% in the experiments. For text recognition, a concatenation structure is designed to combine the features from both shallow and deep layers in neural networks. The experimental results demonstrate that the text recognizer in our approach can improve the accuracy of multi-lingual text recognition. The approach will be beneficial for integrating historical health records and engaging patients in their own health care.

Highlights

The medical laboratory report is one kind of important clinical data, which helps health care professionals with patient assessment, diagnosis, and long-term monitoring
The results demonstrate that the proposed approach can effectively detect and recognize texts from medical laboratory reports
EXPERIMENTS AND RESULTS The proposed approach is evaluated for both text detection and recognition

Summary

INTRODUCTION

The medical laboratory report is one kind of important clinical data, which helps health care professionals with patient assessment, diagnosis, and long-term monitoring. A deep learning approach is presented to detect and recognize texts from a laboratory report image In this approach, a patch-based strategy and a concatenation structure are proposed to handle the problems mentioned above. Recent studies [23], [48], [49] show that the features from shallow layers are important in image classification, object detection, and semantic segmentation Inspired by these works, we take the strategy that combines the features from both deep and shallow layers to solve the local similarity problem between multi-lingual characters. This objective function can directly calculate the cost value for one pair of prediction and ground truth so that the whole network can be trained by an end-to-end way

EXPERIMENTS AND RESULTS

DATASET AND METRICS

CONCLUSION