[New Friends] To know more serialized articles, please follow the blue text below the title “Zhenli Pai”, and check the historical messages.
[Old Friends] Click the top right corner “…”, to forward or share this article with your friends.
Background (Introduction)
With the rapid development of computer networks, the electronicization of information has become an inevitable trend of the era. Text, as the most important and concentrated carrier of information, plays a particularly important role in the process of electronicization. Optical Character Recognition (OCR) technology is a crucial part of the electronicization of text, changing the traditional concept of inputting paper-based materials.
OCR (Optical Character Recognition) refers to the process of electronic devices (such as scanners or digital cameras) checking printed characters on paper, determining their shapes by detecting dark and light patterns, and then translating those shapes into computer-readable text through character recognition methods: scanning text materials, analyzing image files, and obtaining text and layout information. How to debug or use auxiliary information to improve recognition accuracy is the most important issue in OCR. The main indicators to measure the performance of an OCR system include: rejection rate, error rate, recognition speed, user interface friendliness, product stability, usability, and feasibility.
Based on the principles of OCR technology: the process of OCR text recognition is to convert optical image information into text information that can be recognized and processed by computers. The following is the recognition flowchart:
Through OCR technology, users can convert image information obtained from books, manuscripts, forms, and other printed materials through optical input methods like cameras and scanners into text information that can be recognized and processed by computers. Therefore, compared with traditional manual input methods, OCR technology greatly improves the efficiency of data storage, retrieval, and processing.
Currently, OCR technology has matured in the recognition of printed standard fonts and is widely used in various industries for organizing reports and paper documents in banks, governments, and tax offices. However, the application of OCR in the mobile healthcare field is still in the exploratory stage. The Dr.2 team started the OCR project in January this year, and since the initial testing went live in April, they have been continuously improving the system. Last night, a friend in the versatile WeChat group forwarded a message stating that the “Xing Shulin” company once again claimed to have released the country’s first OCR system that can automatically recognize text from medical record photos, just as they previously claimed to have released the mobile healthcare voice recognition system for the first time, even though it merely embedded the “Yunzhisheng” system. Since they dared to speak out, it is always good for the industry; Dr.2 must fairly evaluate the two main mobile healthcare apps currently claiming to have this functionality, and provide an evaluation report based on OCR’s rejection rate, error rate, recognition speed, user interface friendliness, product stability, usability, etc.
Materials and Methods:
Testing Platform:
Two phones (iPhone 4, Redmi 1S; we intentionally chose lower-end phones for broader applicability), Zhenli Pai V1.8.5, Medical Record Folder V3.0.5
Testing Method:
Use the OCR recognition functions of Zhenli Pai and Medical Record Folder to recognize the same images (including: standard medical records, Song font, forms, English, numbers, handwritten cases).
Testing Steps:
1. Install Zhenli Pai and Medical Record Folder on both phones and log in with different accounts.
2. Use the camera to take photos of various recognition materials (to avoid errors from using both systems to take pictures).
3. Create medical records in both systems and import the pre-prepared images from the album into the OCR recognition function.
4. Compare the recognition speed, accuracy, and stability of the two OCR recognition systems.
Evaluation Results
1. OCR (Image Recognition) Interface:
Zhenli Pai: Directly enter the image recognition to take a photo or select from the album, upload, and it will automatically recognize, allowing for classified search of each file for easier management.
Medical Record Folder: Requires creating a medical record, selecting to take a photo or from the album on the create case page, then clicking the “OCR Recognition” above the image, followed by a 24-hour wait for the recognition results.
Zhenli Pai
Medical Record Folder
2. Recognition Speed:
Zhenli Pai: After uploading an image, it is recognized in real-time, with a typical single image recognition speed of about 2 minutes (including upload, recognition, and result return time). We conducted a speed test using a stopwatch for single image recognition speed. This included English, numbers, and Song font characters. Under clear image conditions, the recognition accuracy is above 90%. If the recognition result is unsatisfactory, manual recognition and re-recognition can also be chosen.
This image shows our testers selecting multiple images from the album for recognition at once. The time displayed in the transfer list was: 9:02, and the time displayed in the recognition list after showing the “recognized” status was 9:03. Considering the time difference for selecting images and taking screenshots, the overall recognition time was around 2 minutes, which is consistent with our timing results for single image recognition.
Medical Record Folder: The recognition result was uploaded at 9:55 PM the previous night, and by 9:59 AM the next day, 12 hours later, it was still in recognition, leading Dr.2 to speculate that the Medical Record Folder’s so-called OCR recognition is purely manual recognition. This raises questions about how much efficiency such manual recognition can achieve, even if it can reach nearly 100% accuracy.
Testing Summary (Discussion)
1. It is evident from various tests that the image recognition of the Medical Record Folder is entirely manual recognition, and the results returned the next day did not utilize any automated OCR technology. Therefore, Dr.2 was excited for nothing, thinking that a technical problem that had troubled us for more than half a year had finally been solved, only to be disappointed!
2. The statistical results indicate that the automatic OCR recognition technology of Zhenli Pai has a high recognition rate for standard printed fonts, with English at 95% and Chinese reaching around 80-90%. However, it cannot recognize handwritten fonts, and forms like lab reports are also unrecognizable (this exceeds our technical capabilities). The recognition results are mainly determined by the following factors: 1) image quality 2) image color 3) shooting angle 4) font.
3. For the accuracy of OCR recognition in Zhenli Pai, we can use the following techniques for printed text: for example, appropriately adjusting brightness and contrast when taking photos; avoiding shadows, and leaving white space around the edges. However, for handwritten text with extremely low recognition rates, the current capabilities of the Dr.2 team cannot solve this, and we do not hold any illusions. The only option is to organize the image information that needs recognition manually and then return the organized document information to the user. Zhenli Pai adopts an approach where, after automatic recognition, if the user is unsatisfied or does not want to manually revise, they can mark it for manual assisted recognition. However, it is possible to conduct secondary organization on documents that have been partially recognized successfully, comparing them with the original images, improving efficiency by about three times and reducing costs, especially with the upcoming influx of high-throughput data. Therefore, we can promise that all previous “angel users” who participated in our 500 VIP closed beta testing for one and a half years will have lifetime free access to the “cloud secretary” manual organization function, while all new users can use the manual organization of image recognition results for free for more than three months or even longer, with automated image recognition being free for life.
4. The reason Dr.2 has not launched the Zhenli Pai system, which has been developed for two and a half years, is that they feel this app has not passed their standards. Of course, UI and similar aspects are not the main issue since we focus on applications. However, there are still several critical technical challenges to overcome. Since we use high-definition photos in uncompressed formats, the data volume per case is relatively large. Think about it, when do doctors usually upload? It is likely that 8-10 PM is a very concentrated peak time. If your system is indeed used by many doctors, then during a short period, simultaneous uploads of high-throughput data will face a “flood attack” problem, leading to very slow speeds, frequent disconnections, or server crashes, or triggering automatic interceptions. Additionally, there are technical issues such as cloud service load balancing, cost accounting, and automatic task flow matching.
Overall, we have a long way to go. Let’s work hard! Furthermore, after 8 months of effort and an investment of over 200,000, our investigation into the usage and market share of mobile healthcare apps among doctors in China’s major cities will be released soon, so stay tuned!
PS: Please iPhone users download and install the Zhenli Pai application but do not store medical cases; feel free to play with it, as we are preparing to launch it, transitioning from an enterprise version to a version available for download on the app store. The previous version located at www.medicool.cn will no longer be updated and will only be used for internal testing. The Android version does not have this issue and can be downloaded freely from the official website!
(Unauthorized use of this article is prohibited. Please indicate the source if reprinted. Author: Dr.2. If you wish to communicate with Dr.2, please add WeChat ID: 1340603421)
Creating a Professional Platform for Mobile Healthcare, Everyone is Welcome to Critique!
1、Sina Weibo@medicoolMedical Software
2、www.medicool.cn(Official Website)
3、www.mediskin.cn(Dermatology Doctor’s Home)
4、www.medibone.cn(Orthopedic Doctor’s Home)
5、www.mediendo.cn(Endocrinology Doctor’s Home)
