CN-122024991-A - Postoperative skin targeted detection nursing method based on image photo data analysis

CN122024991ACN 122024991 ACN122024991 ACN 122024991ACN-122024991-A

Abstract

The invention provides a postoperative skin targeted detection nursing method based on image photo data analysis, which comprises the steps of constructing a standardized anatomical region semantic map database with regional healing benchmarks, complication risks and key characterization indexes, combining deep learning to realize spatial positioning and anatomical partition automatic identification of postoperative skin images, extracting multi-dimensional skin state apparent features, carrying out individual restoration dynamic modeling through a sparse sampling LSTM network, fusing regional priori knowledge and individual hidden state features, realizing restoration trend prediction through a attention mechanism, self-adaptively adjusting population and individual information weight according to confidence, continuously calibrating an individual model through closed loop feedback and incremental learning, and finally outputting an interpretable nursing suggestion combined with patient constraint conditions.

Inventors

SUN DEJIN
LUO DAN
YANG JINGHUI
XU TANGYI

Assignees

深圳市哈喽云医科技有限公司

Dates

Publication Date: 20260512
Application Date: 20260205

Claims (10)

1. The postoperative skin pertinence detection nursing method based on image photo data analysis is characterized by comprising the following steps of: S1, constructing a standardized anatomical region semantic map database based on physiological characteristics and healing rules of a main operation region of a human body; s2, segmenting the postoperative skin image of the newly-entered group of patients, identifying the spatial position of the wound, and generating a region attribution result based on the matching of the corresponding anatomical partition labels of the standardized anatomical region semantic map database; S3, collecting a skin state image sequence of a patient at a limited time point, extracting multi-dimensional apparent feature vectors from each frame of image, inputting the multi-dimensional apparent feature vectors into a sparse sampling LSTM network, and generating an individualized recovery dynamic hidden state; s4, invoking historical group recovery pattern tracks matched with the current anatomical partition from the anatomical region semantic map database, calculating similarity weight distribution between the personalized recovery dynamic hidden state and the group reference tracks, and generating an initial recovery trend prediction curve; And S5, calculating confidence scores of the initial recovery trend prediction curves based on the number of the current observation time points and the characteristic stability indexes, adjusting contribution weights of the group reference tracks and the personalized recovery dynamic hidden states in the semantic graphs of the standardized anatomical regions according to comparison results of the confidence scores and preset thresholds, and outputting optimized recovery trend prediction results.
2. The method for post-operative skin pertinence detection and care based on image photo data analysis according to claim 1, wherein the step S5 further comprises: s6, after each nursing intervention is executed, collecting skin state change data of a dry prognosis, calculating a deviation value between an actual response track and a predicted trend, triggering an incremental learning mechanism when the deviation value continuously exceeds a set tolerance range, and finely adjusting the personalized recovery dynamic hidden state parameter by using latest observation data; s7, based on the calibrated recovery trend prediction result, searching a staged nursing target set matched with the current prediction stage in a nursing knowledge base, and generating an interpretable nursing scheme suggestion sequence by combining patient allergy history, medication contraindications and lifestyle constraint conditions; And S8, feeding back the interpretable nursing scheme suggestion sequence to a clinical execution end, recording the actual execution condition and the subsequent skin state response, and updating the intervention effect mapping relation in the nursing knowledge base by utilizing closed-loop feedback data for optimizing the accuracy and suitability of future nursing scheme generation.
3. The post-operative skin targeted detection care method based on image photo data analysis according to claim 1, wherein each node in the standardized anatomical region semantic map database represents an anatomical partition, the side relationship represents an adjacent or functionally associated region, and the attribute field stores medical prior parameters, wherein each anatomical partition is labeled with a typical healing rhythm reference curve, a common complication risk type and a skin state key characterization index.
4. The method for detecting and caring skin after surgery based on image photo data analysis according to claim 1, wherein the step S3 specifically comprises: Acquiring a multi-time point postoperative skin image sequence of a new patient, determining a spatial ROI (region of interest) of an anatomical partition where a wound is located based on an image segmentation result, performing brightness normalization and color temperature compensation treatment on each frame of image, and outputting a time sequence image set with enhanced illumination consistency; Performing multi-dimensional apparent feature extraction on the time sequence image set with enhanced illumination consistency to generate a multi-dimensional apparent feature vector sequence; Taking the multi-dimensional apparent feature vector sequence as input, loading a pre-trained sparse sampling LSTM network model, optimizing and supporting non-equidistant time point input by the sparse sampling LSTM network model, performing time sequence dependent modeling on the irregularly sampled feature sequence based on a gating mechanism, updating a hidden state frame by frame, and outputting a hidden state vector sequence; performing maximum pooling operation on the hidden state vector sequence, extracting the most representative global hidden state characteristics, and generating an individualized recovery dynamic hidden state representation; And calculating an initial confidence weight coefficient represented by the personalized recovery dynamic hidden state, and generating a characteristic stability index based on the number of the currently acquired time points and the Euclidean distance change rate between the continuous two-step hidden states.
5. The post-operative skin pertinence detection and care method based on image photo data analysis according to claim 4, wherein the multi-dimensional apparent feature vector sequence comprises an erythema index, a edema degree, a texture disturbance degree and a wound edge definition, wherein the erythema index is calculated by adopting a pixel analysis method based on HSV and Lab color space fusion and is used as an inflammatory response intensity representation, the swelling degree is measured by a local contrast adaptive threshold method in combination with the texture disturbance degree, the texture disturbance degree is evaluated by Gabor filter bank response entropy, and the wound edge definition is quantified by a Canny edge detection and contour smoothness ratio.
6. The post-operative skin pertinence detection and care method based on image photo data analysis of claim 4, wherein the sparse sampling LSTM network model structure comprises three gating units, namely an input gate, a forgetting gate and an output gate.
7. The method for detecting and caring skin after surgery based on image photo data analysis according to claim 1, wherein the step S4 specifically comprises: Based on the regional attribution result output in the step S2, searching a historical group recovery mode track matched with the anatomical partition where the current wound is positioned from a standardized anatomical region semantic map database; performing time dimension alignment processing on the individuation recovery dynamic hidden state representation output in the step S3, mapping the individuation recovery dynamic hidden state representation to time resolution consistent with the historical group recovery mode track, and extracting multi-dimensional apparent characteristic hidden vectors of the individuation recovery dynamic hidden state representation at each key time point to obtain an individuation recovery state sequence after time sequence alignment; Converting the historical group recovery mode track into a key vector sequence and a value vector sequence to form a priori memory library calculation; based on the personalized recovery state sequence as query input, performing multi-head self-attention calculation with the group prior memory bank, determining the dynamic association strength between the current recovery path of the individual and the group typical mode by calculating a cosine similarity matrix, and generating normalized attention weight distribution; And carrying out weighted summation operation on the value vectors in the historical group recovery mode track according to the normalized attention weight distribution, and fusing individual recovery dynamics and group priori knowledge to generate an initial recovery trend prediction curve.
8. The post-operative skin pertinence detection and care method based on image photo data analysis according to claim 7, wherein the key vector sequence is composed of a reference healing feature vector corresponding to each time step in the historical group recovery pattern track, and the value vector sequence comprises a complication risk level and skin state key representation index matched with the time step key vector.
9. The method for detecting and caring skin after surgery based on image photo data analysis according to claim 1, wherein the step S5 specifically comprises: Acquiring the number of the skin state image observation time points of the current patient as an input condition, executing grading quantization processing based on a preset time window threshold range, and generating a discretized observation density grade label; Calculating a variation coefficient of the multi-dimensional apparent feature vector sequence extracted in the step S3 to generate a feature stability index, and carrying out trend smoothing on the feature stability index to obtain a continuous measurement value of stability; Constructing a two-factor confidence evaluation function based on the observation density grade label and the characteristic stability index, and generating a comprehensive confidence score of an initial recovery trend prediction curve in a weighted linear combination mode; Judging whether the comprehensive confidence score is lower than a preset critical threshold value, if so, generating a high priori dependent control signal, and increasing the contribution weight of the group reference track in the semantic map of the anatomical region in dynamic weighted fusion; If the comprehensive confidence score is higher than or continuously rises above a set threshold, a high-data confidence control signal is generated, the adjustment strength of the group prior is gradually reduced, the transition is carried out to take the personalized recovery dynamic hidden state as the leading mode, the proportional switching of the two source information is completed through the self-adaptive gating unit, and the optimized recovery trend prediction result is output.
10. The method of claim 2, wherein the set of staged care targets comprises frequency of cleaning, anti-inflammatory intervention intensity, maintenance level of moisturization, and scar prevention priority.

Description

Postoperative skin targeted detection nursing method based on image photo data analysis Technical Field The invention relates to the technical field of medical image intelligent analysis and nursing decision support, in particular to a postoperative skin targeted detection nursing method based on image photo data analysis. Background With the rapid development of medical informatization and intelligent care technologies, postoperative skin monitoring and care management is gradually evolving towards data driving and personalized recommendation directions. In the current mainstream postoperative skin recovery trend prediction scheme in industry, an image recognition and time sequence modeling method based on deep learning or machine learning is generally adopted, and the trend prediction of recovery process is carried out by collecting postoperative skin photos of a patient, extracting apparent features such as erythema, edema and pigment and combining with neural networks such as LSTM or GRU. For example, in a large medical image algorithm platform and part of intelligent nursing auxiliary systems, dynamic analysis and risk prompt of skin healing based on continuous observation data are partially realized. The existing literature and application technology also comprises a recommendation system based on a standardized nursing flow, and nursing measure templates are automatically searched through parameters such as morphological characteristics, recovery days and the like, so that auxiliary decision support is provided for clinical medical staff; The technology can obtain ideal recovery trend prediction and nursing suggestion effects under the condition of high data density or sufficient sample size. However, in the face of individual differences, multi-region wound distribution, and sparse sampling (such as high frequency inconvenience of follow-up or irregular photographing time) in actual clinical scenarios, the prior art generally lacks efficient modeling of anatomical region healing rules, complications risks, and care cycle priors. Meanwhile, most systems only conduct time sequence inference based on current observation information of a single patient, and the historical data of regional groups, medical ontology knowledge and structural semantic tags cannot be fully fused, so that the generalization capability of the model is limited. Especially under the conditions of small sample individual cases, severe change in postoperative recovery stage, abnormal individual response and the like, the recovery trend prediction accuracy is obviously reduced, and the prospective and individuation level of nursing recommendation is difficult to meet clinical demands. In addition, the traditional algorithm often ignores dynamic feedback and a model online calibration mechanism, can not continuously correct and predict according to the actual skin response after nursing intervention, and easily solves the problems of scheme matching lag, risk early warning dullness and the like. Disclosure of Invention The invention aims to solve the technical problems and provides a postoperative skin targeted detection nursing method based on image photo data analysis. The technical scheme of the invention is realized in such a way that the postoperative skin pertinence detection nursing method based on image photo data analysis comprises the following steps: S1, constructing a standardized anatomical region semantic map database comprising parts such as head, face, chest, abdomen, limbs and the like based on physiological characteristics and healing rules of a main operation region of a human body, wherein a typical healing rhythm reference curve, a common complication risk type and a skin state key characterization index are marked for each anatomical partition by a map to be used as prior knowledge representation of postoperative recovery trend prediction; S2, segmenting the postoperative skin image of the newly-entered group of patients, identifying the spatial position of the wound, and generating an area attribution result with semantic identification based on the anatomical area semantic map database matched with the corresponding anatomical partition label for guiding the prior information call in the subsequent time sequence modeling; S3, collecting a skin state image sequence of a patient at a limited time point, extracting multi-dimensional apparent feature vectors including erythema index, edema degree, texture disturbance degree and edge definition for each frame of image, inputting the feature sequences into a sparse sampling LSTM network, and generating an individualized recovery dynamic hidden state representation; S4, calling a historical group recovery mode track matched with the current anatomical partition from the anatomical region semantic map database to serve as a reference prior signal, calculating similarity weight distribution between the personalized recovery dynamic hidden state and the group reference