An algorithm for identifying and extracting nameplate information of medical devices

Abstract

This invention relates to an algorithm for identifying and extracting nameplate information from medical devices. Its key feature is the cross-matching of information from a standard medical device library with text recognition results to ultimately determine the basic information of the medical device. This application extracts device information such as name, model, brand, registration certificate number, serial number, and medical device category by recognizing the nameplate image of the medical device. If some information is missing or incorrect, it can be corrected using the standard library. For Chinese nameplates of medical devices, this method achieves 96% accuracy in identifying device information, including device name, model, brand, registration certificate number, and medical device category; for English nameplates, the accuracy reaches 80%. Another advantage of this method is that it can return all device information even if the nameplate contains only partial information.

Description

Technical Field

[0001] This invention relates to the field of image recognition and processing, and in particular to an algorithm for recognizing and extracting information from medical device nameplates. Background Technology

[0002] Currently, there are relatively mature image text recognition technologies. Image text recognition technology has achieved high accuracy in recognizing ordinary text, but when it comes to extracting key information from text, its application areas are standardized texts, such as ID cards, driver's licenses, vehicle registration certificates, license plates, business licenses, and machine-printed invoices. Standardized texts are well-structured and use standardized terms, making information extraction very easy. Commercial algorithm platforms like Alibaba Cloud, iFlytek, and Baidu AI have mature products for standardized texts. However, there are currently no mature products for non-standardized texts. Medical device nameplates are an example of non-standardized text. Due to the variety of different devices and manufacturers, there are hundreds of thousands of different styles of medical device nameplates, making information extraction very difficult. Summary of the Invention

[0003] In view of the above technical problems, the present invention provides an algorithm for identifying and extracting medical device nameplate information. The present invention is a method for determining the basic information of medical devices by cross-matching information from a standard library of medical devices with text recognition results.

[0004] This invention is based on a medical device standard library, which is a database obtained by our company through official data and on-site investigations, and then manually processed. This database covers accurate information such as the name, model, brand, registration certificate number, and medical classification of most medical devices.

[0005] An algorithm for identifying and extracting information from medical device nameplates, the specific identification steps of which are as follows:

[0006] Step 1: Obtain a digital image of the medical device nameplate;

[0007] Step two: Use OCR technology to recognize the text in the above image. Because the text on the nameplate is not continuous, and is not on the same line or has large blank spaces in the middle, it will be divided into different "text fragments". Each text fragment contains the text string and the region coordinates of the text fragment.

[0008] Step 3: After segmenting each text fragment obtained in Step 2, match it with the medical device name keyword database. If the matching rate is high, the text fragment is confirmed as a device name. The segmentation uses the commonly used jieba segmentation tool to divide a text fragment into a group of words. For example, "multi-parameter patient monitor" will be segmented into "multi-parameter," "patient," and "monitor." The device name keyword database is a selection of representative device name keywords based on previously collected segmented device names. The database contains words strongly related to medical devices, such as "Doppler," "ultrasound," "monitor," "X-ray," "disinfection," "MRI," and "dialysis." The words obtained from the segmented text fragments are matched against the medical device name database. Each successful match increments the matching score of the text fragment by 1. Finally, the name matching score Sname of the text fragment and the list of matched medical device name keywords are obtained. The keyword list corresponding to the highest Sname is then matched against the device name database to obtain the device name Equ_name with the highest similarity.

[0009] Step four: Search all text fragments for the words "model," "specification," "model," and "type." If found, the text fragment containing consecutive English letters and numbers, along with the text to its right, represents the candidate device model text fragments. If no device model text fragments are extracted in step four, extract all consecutive strings consisting of non-Chinese characters (numbers, English letters, hyphens, spaces, etc.) from all text fragments as model candidates. Match the obtained model candidate strings with models in the standard device library. The standard device library contains tens of thousands of models; to improve the matching speed, we introduce a prefix tree to aid the query. Each node in the prefix tree is a two-character key, and its corresponding value is all device models containing those two characters. We set the moving window of the candidate string to 2. Starting from the first character, the two characters of the moving window are used as keys to search the prefix tree for the corresponding device model list. If a match is found and the matching rate is higher than a threshold, the device model is added to the candidate model list. Then, the moving window is moved one position to the right, and the above process is repeated until the last position of the candidate string is reached. Finally, a candidate model list is obtained.

[0010] Step 5: Extract the text fragments that conform to the national medical device registration certificate number specifications from the text obtained in Step 2, and use them as the registration certificate number. The national medical device registration certificate number has specific specifications. According to the national specifications, use regular expression matching to try to match the text fragments on the nameplate. If a match is found, it can be used as the registration certificate number of the device. Match the device registration certificate number with the registration certificate number database, and add the registration certificate numbers with a matching rate higher than the threshold to the candidate registration certificate number list.

[0011] Step 6: Take the candidate lists of device names, device models, and registration numbers obtained in Steps 3, 4, and 5, and form a device list L_name (ID: Score_name_id) based on device name, a device list L_model (ID: Score_model_id) based on device model, and a device list L_reg (ID: Score_reg_id) based on device registration number. Each list contains the device ID and matching score in the standard library. Take the intersection of the devices contained in the three lists to obtain the device list [Equ1, Equ2, Equ3, ..., Equ_N]. For each Equ_i, sum the matching scores of the three lists to obtain Score_equ_i = SUM(Score_name_i, Score_model_i, Score_reg_i).

[0012] The device ID corresponding to the largest Score_equ_i is the most likely matching device ID. Based on the device's basic information, the device's basic information is retrieved from the standard library and returned as the final result.

[0013] The basic device information returned in step six includes the device's name, model, brand, registration certificate number, and device category.

[0014] The beneficial effects of this invention are as follows: This application extracts equipment information such as the name, model, brand, registration certificate number, serial number, and medical equipment classification of medical equipment by recognizing the nameplate image of the medical equipment. If some information is missing or incorrect, it can be corrected through a standard library.

[0015] For Chinese nameplates of medical devices, this method identifies device information, including device name, model, brand, registration certificate number, and medical device classification, with an accuracy of 96%. For English nameplates of medical devices, the accuracy reaches 80%. Another advantage of this method is that it can return all device information even if the nameplate contains only partial information. Typically, as long as the registration certificate number or name + model can be identified on the nameplate, this method can return complete device information. Attached Figure Description

[0016] Figure 1 This is a flowchart of the present invention. Detailed Implementation

[0017] Example 1

[0018] The present invention will be further described in conjunction with the following embodiments.

[0019] The medical equipment standard library upon which this method for matching equipment information is based is a database compiled and manually verified by our company, derived from the national medical equipment database and data collected from on-site investigations at hundreds of hospitals. Each record in this database corresponds to a medical device, and each record contains the following information: [Equipment ID, Equipment Name, Equipment Model, Equipment Brand, Equipment Registration Certificate Number, Equipment Medical Classification].

[0020] The specific identification method is carried out in the following steps:

[0021] Step 1: Use a mobile phone, camera, or other means to take a digital image of the medical equipment nameplate;

[0022] Step two: Use Alibaba's OCR technology to recognize the text in the image. Because the text on the nameplate is divided by line breaks and whitespace, OCR technology can recognize different text segments based on these line breaks and whitespace. Each text segment contains the string and the coordinates of its four corners.

[0023] Step 3: Each text fragment obtained in Step 2 is segmented and matched against a medical device name keyword database. If the matching rate is high, the text fragment is confirmed as a device name. The segmentation uses the commonly used jieba segmentation tool to divide a text fragment into a group of words. For example, "multi-parameter patient monitor" will be segmented into "multi-parameter," "patient," and "monitor." The device name keyword database is a selection of representative device name keywords based on previously collected segmented device names. The database contains words strongly related to medical devices, such as "Doppler," "ultrasound," "monitor," "X-ray," "disinfection," "MRI," and "dialysis." The words obtained from the segmented text fragments are matched against the medical device name database. Those with a matching rate higher than a threshold are selected to form a candidate list of device names. For example, device names matching "patient monitor" include ["clinical monitor," "multi-parameter monitor," "electrocardiogram monitor," etc.].

[0024] Step 4: Search all text fragments for words such as "Model", "Specification", "model", "type", etc. If any are present, the continuous string of English and numbers in that text fragment and the text to its right is the alternative text fragment for the device model. If no text fragment for the device model is extracted in Step 4, then extract all continuous strings composed of continuous non-Chinese characters (numbers, English letters, dashes, spaces, etc.) from all text fragments as model alternatives. Match the obtained alternative model strings with the models in the standard device library. There are tens of thousands of models in the standard device library. To improve the matching speed, we introduce a prefix tree to assist in the query. Each node in the prefix tree is a key composed of two characters, and its corresponding value is all device models containing these two characters. We set the moving window of the alternative string to 2. Starting from the first character, use the two characters in the moving window as the key to search for the corresponding list of device models in the prefix tree. Check if there is a match. If the matching rate is higher than the threshold, add that device model to the candidate model list. Then move the moving window one position backward and repeat the above process until reaching the last character of the alternative string. Finally, obtain a candidate model list.

[0025] For example, if an alternative model fragment is "Model EVL000M", use "mo", "od", "de", "el", "le", "ev", "vL", "L0", "00", "0m" in the window as keys in sequence to query in the prefix tree. The device model with a high matching rate found is "EV1000M" (here, due to OCR error, the number 1 was recognized as L, but the device model in the standard library can correct it to the correct model).

[0026] Step 5: Extract the text fragments that conform to the national medical device registration certificate number specification from the text obtained in Step 2 as the registration certificate number; the extraction of the registration certificate number can be carried out according to the naming rules of the national medical device registration certificate number. The naming rules of the national medical device registration certificate number are: ×1 Medical Device Registration ×2 ××××3 ×4 ××5 ××××6. ×1 is the abbreviation of the location of the registration approval department; ×2 is the registration form; ××××3 is the year of first registration; ×4 is the product management category; ××5 is the product classification code; ××××6 is the first registration serial number. Then, the registration certificate number can be extracted through regular matching methods.

[0027] For example, the registration certificate number extracted from the nameplate can be: "State Food and Drug Administration (Import) No. 2013 8211526".

[0028] Step 6: Take the candidate lists of device names, device models, and registration numbers obtained in Steps 3, 4, and 5, and form a device list L_name (ID: Score_name_id) based on device name, a device list L_model (ID: Score_model_id) based on device model, and a device list L_reg (ID: Score_reg_id) based on device registration number. Each list contains the device ID and matching score in the standard library. Take the intersection of the devices contained in the three lists to obtain the device list [Equ1, Equ2, Equ3, ..., Equ_N]. For each Equ_i, sum the matching scores of the three lists to obtain Score_equ_i = SUM(Score_name_i, Score_model_i, Score_reg_i).

[0029] The device ID corresponding to the largest Score_equ_i is the most likely matching device ID. Based on the device's basic information, the device's basic information is retrieved from the standard library and returned as the final result.

[0030] Step six, by matching devices across three dimensions and selecting the most probable direction based on probability, can correct errors caused by OCR recognition mistakes. It can also repair missing information from nameplates. For example, if two devices have very similar models—one a multi-parameter monitor (CL1000) and the other a Doppler ultrasound diagnostic instrument (CLL000)—and the OCR identifies the device as the multi-parameter monitor (CLL000), the model matching score is similar, but the name matching score is significantly higher than that of the multi-parameter monitor. Therefore, it can be assumed that the device is the multi-parameter monitor (CL1000), thus correcting the OCR recognition error.

[0031] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. The various components mentioned in this invention are common technologies in the existing field. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of this invention is defined by the appended claims and their equivalents.

Claims

1. An algorithm for identifying and extracting nameplate information from medical devices, characterized in that... The specific identification steps are as follows: Step 1: Obtain a digital image of the medical device nameplate; Step 2, Text Recognition The text in the above image was identified using OCR technology; Step 3, Device Name Recognition Each text fragment obtained in step two is segmented and matched with the medical device name keyword database. If the matching rate is high, the text fragment is confirmed as a candidate device name. Based on the cross-matching of information from the medical equipment standard library and text recognition results, the basic information of the medical equipment is finally determined. The medical equipment name keyword library is a representative keyword library of equipment names selected after the previous collection of equipment names is segmented. The library contains words that are strongly related to medical equipment, such as "Doppler", "ultrasound", "monitor", "X-ray", "disinfection", "magnetic resonance", and "dialysis". Step 4: Device Model Identification Extract the model number from all text fragments to obtain a list of candidate models; The specific method for obtaining the candidate model list is as follows: First, extract the model number by checking if the words "model", "specification", "model", or "type" are present. If they are, the text fragment containing consecutive English letters and numbers in the text to its right is the candidate text fragment for the device model. If no text fragment containing the device model is extracted in step four, then all consecutive strings consisting of continuous non-Chinese characters are extracted from all text fragments as model candidates. The obtained model candidate strings are matched with models in the standard device library. The standard device library contains tens of thousands of models. To improve the matching speed, a prefix tree is introduced to assist in the query. Each node in the prefix tree is a key consisting of two characters, and its corresponding value is all device models containing these two characters. The moving window of the candidate string is 2. Starting from the first character, the two characters of the moving window are used as the key to search for the corresponding device model list in the prefix tree. If there is a match, if the matching rate is higher than the threshold, the device model is added to the candidate model list. Then the moving window is moved one position to the right, and the above process is repeated until the last position of the candidate string is reached, finally obtaining a candidate model list. Step 5: Extract the text fragments that conform to the national medical device registration certificate number specifications from the text obtained in Step 2, and use them as the registration certificate number. Match the device registration certificate number with the registration certificate number database, and include the registration certificate numbers with a matching rate higher than the threshold in the candidate registration certificate number list. Step six: Take the candidate lists of device names, device models, and registration numbers obtained in steps three, four, and five, and form a device list L_name based on device name, a device list L_model based on device model, and a device list L_reg based on device registration number, respectively. Each list contains the device's ID and matching rate score in the standard library. After comparison and matching, the final result is obtained and then returned.

2. The medical device nameplate information recognition and extraction algorithm according to claim 1, characterized in that... The word segmentation uses the commonly used jieba word segmentation tool to divide a text fragment into a group of words.

3. The medical device nameplate information recognition and extraction algorithm according to claim 1, characterized in that... The words obtained from the word segmentation of the text fragment are matched against the medical device name database. Each successful match will increase the matching score of the text fragment by 1. Finally, the name matching score Sname of the text fragment and the list of matched medical device name keywords are obtained. The keyword list corresponding to the highest Sname is selected, and then the keyword list is matched against the device name database to obtain the device name Equ_name with the highest similarity.

4. The medical device nameplate information recognition and extraction algorithm according to claim 3, characterized in that... The non-Chinese characters refer to characters containing numbers, English letters, hyphens, and spaces.

5. The medical device nameplate information recognition and extraction algorithm according to claim 1, characterized in that... The national medical device registration certificate number is standardized. According to national standards, regular expression matching is used to try to match the text fragments on the nameplate. If a match is found, it can be used as the registration certificate number for the device.

6. The medical device nameplate information recognition and extraction algorithm according to claim 1, characterized in that... The comparison and matching method is as follows: Each of the three lists—device list based on device name, device list based on device model, and device list based on device registration number—contains the device's ID and matching score in the standard library. The intersection of the devices in the three lists yields the device list [Equ1, Equ2, Equ3, ..., Equ_N]. For each Equ_i, its matching score in the three lists is summed to obtain Score_equ_i = SUM(Score_name_i, Score_model_i, Score_reg_i). The device ID corresponding to the largest Score_equ_i is the most likely matching device ID. Based on the device's basic information, the device's basic information is retrieved from the standard library and returned as the final result.

7. The medical device nameplate information recognition and extraction algorithm according to claim 1, characterized in that... The basic device information returned in step six includes the device's name, model, brand, registration certificate number, and device category.

Images