Study design
The design of this study is shown in Fig.1, including general information, pretreatments, knowledge mining, and results. After the participant IRT data is preprocessed, the image can be segmented into ROI sets. After the temperature characteristics in the ROI are granulated, the knowledge graph can be generated through APOS. After the knowledge mining of the control group and the MS group, the skin temperature distribution of the two groups can be obtained, and finally the mining results are verified by statistical methods. This research passed the review of the Ethics Committee of Dongzhimen Hospital in Beijing University of Chinese Medicine (No. DZMEC-KY-2020-102), and conducted in accordance with the Declaration of Helsinki.
The design of this study.
Participants in this study came from the Physical Examination Center of the International Department of Dongzhimen Hospital, Beijing University of Chinese Medicine, and all participants signed the informed consent form. Participants did not receive any financial incentive during the trial, but all examinations were free of charge.
The diagnostic of the participant should comply with MS diagnostic standards (Detailed diagnostic criteria are written in the supplementary materials section1);
The age of the participant was greater than 18years old;
The participant agreed to participate in this study.
Women during pregnancy or lactation;
Those with a history of cardiovascular and cerebrovascular diseases, severe liver and kidney insufficiency, blood diseases, severe trauma or major surgery;
Diseases with metabolic abnormalities as the main manifestation, such as hyperthyroidism, hypothyroidism, acute thyroiditis, hypovolemia, etc.;
Participants have obvious allergic rash on the skin;
Those who have a history of mental illness or who do not agree to participate in this research.
Due to the high sensitivity of IRT to temperature, it is prone to interference during data acquisition. This study refers to the consensus statement on human skin temperature measurement proposed in the literature14.
The individual data of all participants were recorded in the data table;
Participants had no alcohol, smoking, caffeine, large meals, ointments, cosmetics, and showers in the four hours before the images were collected;
Participants did not exercise vigorously within two hours before the image was collected;
The ambient temperature was 251, and the relative humidity of the air was 4050%;
There was no interfering heat source in the IRT field of view, and no obvious air flow;
The equipment is HIR-2000A cabin type IRT produced by Beijing Yuetian Optoelectronics Technology Co., Ltd. The thermal sensitivity was 0.05 at 30, the spectral range was 7.513m, the image pixel size was 256336, the spatial resolution was 0.9mrad, and the acquisition frequency was 9Hz.
The IRT collection environment is a thermostatic chamber, and all participants are in the same space to ensure that the environmental baseline is consistent;
Power on the equipment half an hour before data collection to keep the equipment in a steady state;
The vertical distance between the participant and the lens is 2.5m;
The center line of the lens field of view is perpendicular to the human body through the liftable pan/tilt;
The emissivity of the detector was set to 0.98 (default);
The time for all participants to collect images is 9:0011:00 am;
All participants maintained the same standing posture, with their arms drooping naturally, palms forward, arms and legs slightly extended;
Before collecting images, participants remove their clothes and accessories, and rest for 15min in the thermostatic chamber;
The IRT data processing software is independently developed by our team. The data save format is a *.dat file, which contains all the temperature values of any space coordinates in the field of view, and is filled in a 16-bit integer data format.
The preprocessing include the background filtering method and the ROI segmentation method based on anchor point positioning. The purpose of background filtering is to filter out all the temperature information except the human body information in the image, so as to facilitate the automatic calculation of boundary coordinates during the ROI segmentation process. The purpose of ROI segmentation is to obtain the sub-domain information of the human skin temperature distribution, which is convenient to find the different regions of different groups of people through the method of knowledge mining.
IRT is a pseudo-color image, and the color channel of the image can be changed by adjusting the temperature forming width (tfw), temperature forming starting point (tfs), and palette15. Therefore, the conventional background filtering method based on color channels is not suitable for processing IRT.
This paper proposes a bimodal background filtering method based on the temperature field distribution, which can filter IRT background succinctly and quickly, and retain the effective data of human skin IRT. The flow chart of the algorithm is shown in Fig.2, and the specific operation flow is shown as follows:
Extracting the IRT temperature field data to be processed as input;
Calculating the distribution frequency of each temperature value in the field and drawing a curve;
Performing gaussian smooth filtering on the curve, and seting the operator size to 5;
Detecting and locating the first two peaks of the processed curve;
Calculating the median between the two peaks as the threshold to segment the image background.
Flow chart of bimodal background filtering method.
The purpose of this study is to explore the rule of skin temperature distribution on different parts, but there are individual differences in the participants on height and weight, and it is difficult to perform rapid segmentation through image processing. A semi-automatic region segmentation method based on anchor marking was proposed in this study. This process required the operator to manually mark the anchor points, and then generate ROI through mathematical calculation and image processing to complete the image segmentation.
Determining the measurement targets are the premise of segmenting the ROI of the people IRT. In this study, the human image was divided into 18 ROI of symmetry according to the midline, and corresponding to 12 anchor points, as shown in Fig.3. Each anchor point corresponds to the body surface marking part of the participant as shown in Table 1. According to the information of 12 anchor points, the vertices of each ROI can be quickly calculated. The comparison relationship between ROI labels and anchor points is shown in Table 2.
The region segmentation target.
After background segmentation, the background had been set to 0C, so the boundaries of each ROI could be found through row and column traversal according to the anchor points. The circumscribed quadrilateral area of the target region could be extracted, and the background information contained in the quadrilateral could be excluded through later calculation. The key to the extraction of the quadrilateral region is the acquisition of the coordinates of its four vertices. Different regions have different calculation methods. The ROI vertex coordinates are mainly calculated based on the same-size scale.The specific calculation method can be found in the supplementary materials section2.
APOS is proposed on the basis of order theory, granular computing, and formal concept analysis. Its main purpose is to visually analyze and mine the knowledge structure between attributes and objects in data tables16. The generation principle of APOS as follows:
The attribute nodes with more coverage objects are higher in the hierarchy, and the attribute nodes with fewer coverage objects are lower in the hierarchy;
The same or similar objects are close to each other, and different objects are far away from each other;
The higher-level attribute nodes are more universal and the lower-level attribute nodes are more specific;
Each branch is a collection of attributes of an object.
This article does not give a detailed introduction to the specific generation method. The supplementary material provides a simple explanation of the method through a simple example (section3). For the specific generation code, please refer to the literature17,18.
The formal background is a binarized NS data table composed of attributes and objects. Each object corresponds to N attributes, and each attribute corresponds to S objects. The formal background also is a binary data table, and the generation of APOS is based on the table. Therefore, how to transform the image features of the IRT into the formal background, will be the key point to this research.
This paper proposed a regional feature granulation method based on K-Means.The flow chart is shown in Fig.4. The core idea of this method was to cluster the ROI data into three types: high temperature, medium temperature, and low temperature, and then distinguish the high and low temperature according to the clustering results and the temperature features of all ROI. A formal background with two attributes of high temperature and low temperature should be generated in each ROI.
Flow chart of data granulation.
K-Means is a commonly used unsupervised clustering algorithm, which has obvious advantages, but also has disadvantages, such as: sensitivity to noise and discrete points; the hyperparameter K must be determined in advance; if the center point is initialized randomly, each clustering result is inconsistent19. Therefore, the design goal of this algorithm is to reduce the impact of the above disadvantages of the clustering results as much as possible.
Different tfw and tfs can show the same data as different images. The difference in the discrete points of the image is particularly obvious, which has a great impact on the clustering of K-Means. Therefore, the drawing rule should be standardized, and the parameters of all ROI of each participant must be consistent. To solve this problem, this paper proposed an adaptive image adjustment method, which was based on temperature data. The tfw usually be set to 10 in medical IRT research15, and in this study, tfw was initialized to 10, and tfs was set to the difference between the maximum temperature and tfw. At the same time, the grayscale of the standardized image could further reduce the impact of the palette on the data and reduce the computational complexity.
The goal of the algorithm is to extract the high-temperature attribute and the low-temperature attribute in the ROI. As some of the extracted ROI contain background data, but some do not. The background has an influence on the number of classifications, when there was background data in the ROI, the super parameter K was set to 4, otherwise the K was set to 3, which could solve the problem of preset K in the K-Means algorithm.
The random center point method is often used in the initialization of K-Means, but this method may result in inconsistent clustering results each time, and the robustness is poor. In order to solve this problem, this study can initialize the center point according to the distribution features of IRT temperature data to ensure the consistency of the clustering results of all ROI in the same image. When K was 4, the four center points were set to 0, non-zero minimum value, non-zero average value, and non-zero maximum value, and when K was 3, the three center points were set to the minimum, average and maximum value. This method directly enhances the robustness of the algorithm.
After K-Means clustering, each ROI can be divided into four clusters: high temperature, medium temperature, low temperature and background. When the temperature of the high-temperature cluster in the ROI was higher than the average of the high-temperature clusters in all regions, it means that the high-temperature attribute in the region was dominant. Similarly, whether the low-temperature attribute was dominant could also be judged. This is the process of converting the image data into qualitative data. The specific discriminant arithmetic model is shown in formula (1):
$$ left{ begin{gathered} ROI{ - }L = 1;{text{else }};{0},;min (T_{roi} ) le overline{{T_{min} }} hfill \ ROI{ - }H = 1;{text{else}};{ 0},;max (T_{roi} ) ge overline{{T_{max} }} hfill \ end{gathered} right. $$
(1)
where ROI-L represents the low-temperature attribute of the ROI, ROI-H represents the high-temperature attributeof the ROI region, Troi represents the average temperature set of the non-zero clusters obtained by K-Means of the target region, (overline{{T_{min} }}) represents the average temperature of the smallest non-zero cluster in all ROI of the whole body, and (overline{{T_{max} }}) represents the average temperature of the largest cluster in all ROIs of the whole body.
The human IRT segmentation method described in section2.3 was used to divide the image into 18 ROI. Combined with the data granulation method, 18 ROI could be granulated into 218 attribute sets, where each ROI corresponded to two attributes of ROI-L and ROI-H. The two formal backgrounds of this study would be constructed corresponding to the control group and the MS group, with the same number of data objects and the same scale of knowledge graph, and the qualitative description of the IRT image can be realized, and a knowledge graph can be constructed to facilitate subsequent knowledge mining.
Continuous variables are expressed as the mean (standard deviation), and categorical variables are expressed as numbers and percentages.The continuous variables of baseline characteristics were evaluated with ANOVA, and the categorical variables were evaluated with 2 test. For the characteristics of the target area obtained through knowledge mining were evaluated with independent-samples T test. Analyses were performed with SPSS (IBM SPSS Statistics, New York) Version 26 with P<0.05 considered significant.
Read this article:
An exploration of new methods for metabolic syndrome examination by infrared thermography and knowledge mining | Scientific Reports - Nature.com
Read More..