Deep learning based early visual warning method for leakage of ostomy wafer

Abstract

The application discloses a deep learning-based early visual warning method for leakage of a stoma disc, and the specific steps of the method are as follows: first, a three-dimensional spectral cube is generated through multispectral image acquisition, a colloid chemical change and an interface structure defect are extracted through a double-path neural network, and are fused into leakage correlation characteristics, hidden hydrops volume is inversed relying on an optical physics model, time sequence comparison learning is used for correcting time-varying drift, finally, risk grade determination and abnormal point labeling are completed by combining hydrops numerical value and corrected characteristics. The application can realize early quantitative warning of hidden leakage of the stoma disc, through synchronous analysis of colloid chemical change and interface structure defect by multispectral imaging and a double-branch network, various interferences are excluded, and hydrops volume is inversed; baseline drift is corrected by combining self-supervised time sequence learning, risk is determined by double parameters, and abnormal point positions are accurately labeled, so that skin damage and other complications are effectively reduced, and the safety and convenience of daily nursing are improved.

Description

Technical Field

[0001] This invention relates to the field of deep learning and medical care detection technology, specifically to a deep learning-based method for early visual warning of stoma chassis leakage. Background Technology

[0002] As a core consumable material that comes into direct contact with human skin and excrement during stoma care, the stability of the adhesive sealing layer directly determines the quality of stoma care. The digestive fluids continuously excreted from the human intestines are highly corrosive and will slowly erode the adhesive material of the stoma baseplate. Simultaneously, human limb movement and skin deformation can create tiny gaps at the interface between the adhesive and the skin, allowing small amounts of exudate to accumulate within the adhesive layer, forming hidden fluid accumulation. This type of accumulation is initially not visible to the naked eye, but as it continues to accumulate, problems such as adhesive edge lifting, skin maceration and redness, and excrement leakage will gradually appear. In severe cases, it can lead to complications such as contact dermatitis and skin ulceration. In current clinical nursing scenarios, relying on optical imaging combined with intelligent detection to achieve early identification of leakage has become the mainstream development direction. Multispectral imaging technology can capture microscopic spectral changes in materials that cannot be identified by single visible light, and deep learning models can automatically analyze the subtle defect features in the image. Combining these two technologies for the screening of occult chassis occult leakage has become an important research direction for improving the long-term care of occult tissues. The industry continues to explore integrated detection processes adapted to colloidal multilayer media structures to solve the practical nursing problem of occult leakage being difficult to detect in a timely manner.

[0003] Traditional stoma leakage detection methods mostly rely solely on visible light images for surface appearance recognition, only able to identify overt leaks with obvious seepage and large-area colloid detachment, failing to capture the material spectral changes and microscopic interface defects corresponding to trace amounts of fluid within the interstitial space. Conventional single-path deep learning models can only extract single-dimensional image features, unable to separately analyze chemical corrosion changes in colloidal materials and spatial structural defects at the bonding interface, easily misclassifying irrelevant textures such as body hair, skin folds, and environmental shadows as leakage features, resulting in significant biases in fluid accumulation estimation. General detection models do not incorporate objective optical laws governing colloidal light transmission and liquid spectral attenuation to constrain the feature extraction process, making it easy for network training to generate false features that do not reflect real physical changes, further reducing the accuracy of fluid accumulation quantification. Furthermore, general models lack temporal correction logic for individual chassis, failing to eliminate spectral baseline shifts caused by different skin colors and batches of colloidal material, and relying on a large number of manually labeled samples for training, leading to high costs for practical implementation and adaptation. Determining risk levels using a single indicator is prone to missed or incorrect assessments. After testing, it can only provide a general risk warning and cannot accurately locate the specific area corresponding to the potential leakage, making it difficult to support nursing staff in carrying out targeted treatment. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of existing technologies and provide a deep learning-based early visual warning method for stoma chassis leakage. This method constructs a three-dimensional spectral cube through multi-spectral synchronous imaging, extracts colloidal chemical anomalies and interface structural defects through a dual-branch spectral-spatial attention network, and fuses them into global leakage correlation features. It relies on an optical transmission physical model to invert the latent fluid volume and uses temporal comparative learning to correct baseline drift. Finally, it combines the fluid volume and correction features to achieve dual-threshold graded early warning. This method does not require manual annotation, can accurately quantify trace amounts of fluid in the interstitial space, distinguish between normal aging and early failure, and achieve early, non-invasive, and intelligent early warning of leakage risks, effectively improving patient safety and comfort.

[0005] To solve the above-mentioned technical problems, this invention provides the following technical solution: an early visual warning method for stoma chassis leakage based on deep learning, the specific steps of which are as follows: S100, Multispectral Image Acquisition: Multi-band synchronous imaging processing is completed on the colloidal bonding area of ​​the stoma chassis, and after radiometric correction and spatial pixel registration, three-dimensional spectral cube data is generated by stacking. S200, Dual-path feature extraction: Input the three-dimensional spectral cube data into a preset deep learning network to analyze the spectral dimension of colloidal material variation features and the spatial dimension of interface defect features respectively. After the two types of features are fused, the global leakage correlation features are generated. S300, Liquid Volume Inversion: Based on the physical constraints of optical transmission, the correlation characteristics of leakage in the whole domain are converged and optimized. After removing irrelevant interference information, the liquid volume inside the interlayer is quantitatively solved, and the quantifiable value of the hidden liquid volume in the interlayer is output. S400, Time-varying Feature Correction: Input the global leakage correlation features and the latent liquid volume values ​​into the time-series comparison learning framework to complete the baseline drift deviation correction calculation and output the corrected leakage feature vector after eliminating the baseline deviation. S500, graded early warning output: Combines the hidden liquid accumulation volume value with the corrected leakage feature vector to complete the risk level determination, pushes corresponding prompt information in a graded manner and marks the abnormal points, and completes the early warning of leakage in all stages.

[0006] Furthermore, the multi-band synchronous imaging processing employs the same exposure clock to trigger simultaneous exposure of four imaging devices: visible light, 850nm near-infrared, 940nm near-infrared, and 1450nm short-wave infrared. The duration of a single exposure is controlled within 200ms. During the acquisition process, the device lens is kept perpendicular to the chassis plane, and the acquisition range covers the entire colloidal sealing area of ​​the stoma chassis and the surrounding skin tissue within a 5mm radius, completely capturing all spectral information of the leakage-prone area. After acquisition, radiometric correction is performed on the images of each band to eliminate brightness deviations caused by differences in hardware response of different photosensitive channels. Then, pixel-level spatial registration is used to align the images, calibrating the pixel coordinate deviations at the same physical location in different bands to within a single pixel. The four corrected images are stacked according to the channel dimension to form a three-dimensional spectral cube data.

[0007] Furthermore, the three-dimensional spectral cube data structure includes two-dimensional spatial location information and multi-dimensional spectral response information of the stoma base plate bonding area. Each spatial pixel in the three-dimensional spectral cube data corresponds to a complete multi-band spectral response curve, which can simultaneously characterize the colloidal appearance state and material spectral characteristics at the same physical location.

[0008] Furthermore, the preset deep learning network is a dual-branch spectral-spatial attention convolutional network adapted to three-dimensional spectral cube data. It is built on a three-dimensional convolutional architecture and includes parallel and independent spectral feature analysis branches and spatial feature analysis branches. The input ends of the two branches are connected to the three-dimensional spectral cube data, and the output ends complete the feature channel splicing and fusion. The spectral feature analysis branch performs weight adaptive calculation on the spectral dimension data of the three-dimensional spectral cube to capture the subtle material changes such as the shift of spectral absorption peaks and the attenuation of spectral response intensity after the colloidal material of the stoma base plate is eroded by digestive fluid, and filters out the effective spectral signals that correspond only to the chemical failure of the colloidal material. The spatial feature analysis branch performs gradient feature extraction on the spatial dimension data of the three-dimensional spectral cube to identify microscopic cracks and gaps in the interface between the colloidal material and the skin, and at the same time extracts the redness and darkening of the skin surface caused by the seepage. After the features of the two branches are analyzed independently, feature fusion is completed to generate a global leakage correlation feature that takes into account both material chemical changes and spatial structural defects.

[0009] Furthermore, during the training of the dual-branch spectral-spatial attention convolutional network, the convergence optimization employs optical prior rules that conform to the multilayer transmission mechanism of colloids to constrain the convergence direction of network features. The inherent transmission attenuation characteristics of liquids on multi-band spectra are used as physical constraints to suppress invalid noise features and pseudo-features that deviate from objective optical laws in the global leakage correlation features. During the feature convergence optimization process, only effective leakage features that conform to the spectral changes of colloid transmission, liquid attenuation, and material corrosion are retained. This corrects the feature offset problem caused by the network's free training, ensuring that the leakage features learned by the network fully conform to the real optical change mechanism of latent leakage in the stoma chassis, thereby improving the accuracy and effectiveness of subsequent fluid quantification solutions.

[0010] Furthermore, the amount of liquid accumulated inside the interlayer is quantitatively calculated using a formula. Complete, in the formula The overall transmitted light intensity of the chassis after multi-band fusion; The baseline transmitted light intensity of the chassis bonding area under no liquid accumulation condition is obtained by collecting the transmitted light intensity of this area when the user first wears the brand new chassis, or by measuring the same batch of adhesive material in the non-bonded area of ​​the chassis. Let be the attenuation coefficient corresponding to the i-th near-infrared characteristic band, where i=1,2,3 correspond to the three bands of 850nm, 940nm, and 1450nm, respectively. ∈[0,1], and its specific value is obtained by calibration using standard liquid samples; The shift in spectral absorption peak caused by corrosion of the colloid in the corresponding wavelength band is expressed in nm and is obtained by peak detection of the absorption peak position in the corresponding wavelength band from the spectral feature analysis branch; K is the optical path correction constant of the colloidal interlayer, which is determined by the thickness of the colloidal layer and the properties of the light-transmitting material and is obtained by calibration using standard samples. The algorithm is designed to calculate the volume of the hidden liquid inside the chassis interlayer. It couples the three types of interactions: colloidal corrosion spectral shift loss, multilayer medium transmission loss, and liquid infrared attenuation loss. The hidden liquid volume is solved in reverse with the measured transmitted light intensity as a known quantity, and the constraint feature learning conforms to the real optical laws.

[0011] Furthermore, interference feature filtering is performed simultaneously during the liquid volume inversion process. First, high-texture regions corresponding to body hair and skin folds are identified and located through spatial dimension gradient identification. Weighted attenuation processing is performed on the near-infrared transmission features of high-texture regions to reduce the impact of light intensity fluctuations caused by non-leakage factors. For environmental shadow interference, the light intensity ratio characteristics of the visible light band and the near-infrared band are used to distinguish between shadow areas and real liquid areas. The band ratio of the two types of areas has a fixed difference, thereby eliminating the miscalculation of liquid volume caused by shadows. After interference filtering is completed, the area of ​​the liquid-covered region and the equivalent thickness of the liquid are converted by combining the light intensity attenuation degree of near-infrared light after penetrating the colloidal interlayer. The final hidden liquid volume value is obtained by multiplying the area of ​​the liquid-covered region and the equivalent thickness of the liquid.

[0012] Furthermore, the temporal comparison learning framework employs a self-supervised temporal feature alignment training mechanism. It constructs a baseline feature sample library using standardized three-dimensional spectral cube data collected when the user first wears the new stoma chassis, storing the original spectral spatial features of the chassis in its initial state of no corrosion and no fluid accumulation. Subsequent leakage-related features collected over different usage periods are used as temporal samples to be corrected. These samples are paired with the baseline feature samples to construct training sample pairs, which are then fed into the network for feature alignment learning. Parameter optimization is achieved by reducing the intra-class differences in feature vectors at different usage stages of the same chassis and increasing the inter-class distance between feature vectors of normal and leaking chassis. This allows for personalized model adaptation without the need for manual labeling of a large number of leakage samples, continuously learning the characteristic differences between normal aging and corrosion-induced leakage of the colloid, and distinguishing between normal colloid states and early failure states.

[0013] Furthermore, the baseline drift deviation correction calculation is performed using the formula... Complete, in the formula This is the corrected leakage feature vector after baseline drift correction; The original three-dimensional spectral spatial features extracted by the network at the current detection time; These are the original features of the baseline sample collected when the chassis was first worn; The time-series correction weighting coefficient has a value range of 0.4 to 0.6. The implicit liquid volume output by the liquid volume inversion step; To provide early warning of critical volume thresholds, the algorithm uses the inverted liquid volume as a correction amplitude adjustment factor. While subtracting the baseline offset component, it adaptively enhances the leakage signal that is close to the warning threshold. That is, the closer the liquid volume is to the warning threshold, the stronger the correction of the baseline offset is, thereby eliminating the spectral baseline drift problem caused by batch differences in colloids and individual skin color differences.

[0014] Furthermore, the risk level determination adopts a dual-parameter joint determination mechanism, using the volume of latent liquid accumulation as the core quantitative determination indicator and the corrected leakage feature vector as the auxiliary feature determination basis. Two fixed determination thresholds are set. The first-level early warning determination threshold is 0.02mL. When the volume of latent liquid accumulation is ≥0.02mL and the corrected leakage feature vector detects a colloidal spectral corrosion change signal, it is determined to be an early latent leakage risk. The second-level emergency replacement determination threshold is 0.05mL. When the volume of latent liquid accumulation is ≥0.05mL, regardless of whether the corrected leakage feature vector detects a colloidal spectral corrosion change signal, it is determined to be a high-risk leakage risk, and a visible excrement leakage is about to occur. The dual-parameter joint determination can avoid the defects of single indicator determination and improve the accuracy of level determination.

[0015] Furthermore, the notification information is pushed in a differentiated manner based on the risk level assessment results. For early-stage latent leakage risks, a chassis status observation warning is pushed, prompting the user to increase the frequency of chassis fit inspection. For high-risk leakage risks, an emergency chassis replacement warning is pushed, prompting the user to replace the stoma chassis immediately. At the same time, based on the spatial defect coordinates locked by the correction leakage feature vector, the abnormal locations of colloidal interlayer fluid accumulation, colloidal corrosion, and microscopic crack abnormalities are marked. The physical spatial positioning error of the chassis skin plane is controlled within 2mm, realizing the integrated output of quantitative classification and early warning of leakage hazards and precise location positioning.

[0016] Compared with existing technologies, this deep learning-based early visual warning method for stoma chassis leakage has the following advantages: I. This invention utilizes multi-band synchronous imaging combined with a three-dimensional spectral cube data construction mode to simultaneously acquire spatial morphology and multi-dimensional spectral response information. It simultaneously analyzes the chemical changes of colloidal materials and structural defects at the bonding interface. A bi-branch attention convolutional network separates and fuses two independent types of features. Simultaneously, optical transmission mechanisms are introduced as training constraints for the network, effectively suppressing meaningless noise and false features that violate objective optical laws. This allows for the detection of latent corrosion and microscopic gap hazards in the stoma base that are not yet visible to the naked eye. Combined with multi-dimensional interference filtering logic, it distinguishes irrelevant interference sources such as body hair, skin folds, and environmental shadows, eliminating false positives from both spectral and spatial dimensions. Based on an optical attenuation model, the total amount of interstitial fluid is inversely derived, enabling quantitative calculation of leakage risks. This overcomes the limitation of traditional visual inspection, which can only identify obvious surface leaks, allowing for earlier identification of leakage risks and providing quantifiable data support for early intervention.

[0017] II. This invention establishes personalized baseline feature samples through a self-supervised temporal comparison learning framework. It completes temporal feature alignment and correction based on the initial state of the chassis itself, eliminating the need for large-scale manual annotation of leakage samples to achieve model adaptation. It dynamically adjusts the baseline drift correction amplitude using quantified fluid accumulation values, offsetting spectral baseline shifts caused by batch composition of colloids, skin color, and long-term aging, thus stabilizing the recognition accuracy of leakage features at different usage stages. A dual-parameter joint judgment rule using fluid accumulation volume and corrected feature vectors is employed to classify risk levels. This dual-indicator complementarity avoids the failure of single-parameter judgment. Differentiated prompts are pushed based on risk level, simultaneously locking various abnormal points and completing high-precision spatial annotation, forming an integrated early warning system of quantitative assessment, graded reminders, and hazard location. This reduces the probability of complications such as skin damage and chassis detachment caused by chassis leakage, improving the safety and convenience of daily stoma care.

[0018] Other advantages, objectives and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination or study, or may be learned from the practice of the invention. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This is a flowchart of the early visual warning method for stoma chassis leakage based on deep learning according to the present invention. Figure 2 This is a data transmission diagram of the early visual warning method for stoma chassis leakage based on deep learning according to the present invention. Figure 3 This is a flowchart of the dual-path feature extraction step in this invention. Detailed Implementation

[0021] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0022] Example 1: This embodiment is designed for postoperative adult patients who can independently complete stoma care at home. It is compatible with portable, small multispectral acquisition devices and fully implements all the operational procedures described in the claims. The entire testing process can be completed by the patient independently, without the need for on-site medical personnel accompaniment. The testing process covers the entire area of ​​the stoma baseplate's gel seal, and can detect minute leakage risks that are not easily detected by conventional visual observation. A complete data verification process is included to verify the accuracy of the entire method. Figure 1 As shown.

[0023] The first step involves multispectral image acquisition. The patient vertically aligns the lens of the portable device with the stoma baseplate they are wearing, with each exposure lasting less than 200ms. The device simultaneously triggers four imaging devices: visible light, 850nm near-infrared, 940nm near-infrared, and 1450nm short-wave infrared. The imaging range completely covers the entire colloid sealing area around the stoma baseplate and the surrounding skin tissue within a 5mm radius. This allows for the complete capture of all spectral information from high-risk leakage locations and the complete preservation of the original optical response information of the skin and colloid in the same area, avoiding missed detection of potential problems due to missing images in certain areas. After all images are acquired, the system automatically performs radiometric correction to smooth out the inherent brightness differences between the four imaging hardware photosensitive channels, eliminating basic imaging errors caused by hardware differences. Subsequently, pixel-level spatial registration processing is performed to calibrate the pixel coordinate deviations at the same physical location in different bands to within a single pixel, ensuring complete spatial correspondence between multi-band images. The four sets of calibrated images are then stacked according to channel dimensions to generate three-dimensional spectral cube data. This data simultaneously carries the two-dimensional spatial location and multi-dimensional spectral response information of the chassis bonding area. Each spatial pixel retains a complete multi-band spectral response curve, synchronously recording the colloid's appearance and internal material spectral properties, providing complete underlying data support for subsequent feature extraction. In the data verification phase of this embodiment, 20 home ostomy patients underwent 30 consecutive days of routine monitoring. The pixel registration deviation of the three-dimensional spectral cube data generated from all subjects was stably controlled within a single pixel, with no loss of local spectral information. The basic data effectiveness rate in the imaging acquisition phase reached 100%.

[0024] The second step involves dual-path feature extraction. The system inputs the generated 3D spectral cube data into a dual-branch spectral-spatial attention convolutional network adapted to the 3D spectral cube data. This network is based on a 3D convolutional architecture, with the spectral feature analysis branch and the spatial feature analysis branch running in parallel and independently. Both branches simultaneously read the same set of 3D spectral cube data. The spectral feature analysis branch performs weighted adaptive calculations on the spectral dimension data, accurately capturing subtle material change signals such as spectral absorption peak shifts and spectral response intensity attenuation after the colloid comes into contact with the digestive fluid for corrosion. Redundant spectral signals unrelated to the chemical failure of the colloid are filtered out, retaining only the effective spectral data that reflects the corrosion and deterioration of the colloid. The spatial feature analysis branch performs gradient feature extraction on the spatial dimension data, accurately identifying microscopic cracks formed at the interface between the colloid and the skin, as well as gaps at the edges of the colloid. Simultaneously, it captures the visual features of redness and darkening of the skin caused by the immersion of exudate. After the two branches complete feature analysis, they perform feature channel splicing and fusion to generate global leakage-related features. During the network training phase, feature convergence optimization is performed based on the optical prior rules corresponding to the colloidal multilayer transmission mechanism. This automatically suppresses image noise and pseudo-features that deviate from objective optical laws, retaining only effective leakage features that conform to the laws of colloidal light transmission, liquid spectral attenuation, and material corrosion spectral changes. This mitigates the feature offset problem caused by network self-training, ensuring that the features extracted by the network perfectly match the real optical changes of latent leakage in the stoma chassis, thereby improving the accuracy of subsequent fluid quantification calculations from the source. In the data verification phase of this embodiment, compared with the output features of a single-branch convolutional network and the dual-branch network of this scheme, the dual-branch network increases the proportion of effective leakage features extracted by more than 70%, reduces the number of pseudo-feature interferences, and avoids the situation where natural skin texture is judged as leakage defects.

[0025] The third step involves inverting the volume of the fluid accumulation. The system, relying on optical transmission physical constraints, calls a fluid volume quantification algorithm to perform numerical calculations of the interlayer fluid, fully incorporating the fluid volume quantification formula: During computation, interference feature filtering is initiated simultaneously. First, high-texture areas formed by body hair and skin folds are identified and located using spatial dimensional gradient identification. Weighted attenuation is applied to the near-infrared transmission characteristics of these areas to reduce light intensity fluctuations caused by the skin's texture. Then, the ratio of visible to near-infrared light intensity is used to distinguish between environmental shadow areas and actual fluid accumulation areas. The inherent difference in the band ratio between the two types of areas eliminates miscalculations of fluid accumulation values ​​caused by shadows. During computation, the system first reads the overall comprehensive transmission light intensity of the chassis after multi-band fusion collected by the device. The base value of transmitted light intensity in the chassis contact area under conditions without liquid accumulation. Retrieve the attenuation coefficients corresponding to the three wavelength bands of 850nm, 940nm, and 1450nm. Read the shift of the colloidal corrosion spectrum absorption peak output from the spectral characteristic branch. The optical path correction constant K of the colloidal interlayer, obtained from standard sample calibration, is used to substitute all the known parameters into the formula to solve for the hidden liquid volume inside the interlayer. The entire computational logic incorporates three types of influencing factors: colloidal corrosion spectral shift loss, multilayer media transmission loss, and liquid accumulation infrared attenuation loss. Using the transmitted light intensity measured by the equipment as known conditions, the total amount of interlayer liquid is deduced in reverse. The entire process follows objective optical change laws, avoiding numerical distortion caused by single-dimensional calculations. In this embodiment, multiple sets of micro-samples of interlayer liquid were manually pre-sampled for testing during the data verification stage. The error range between the algorithm's output liquid volume and the actual added liquid volume is extremely small, with no serious deviations in overestimating or underestimating the liquid volume.

[0026] The fourth step involves time-varying feature correction. The system retrieves standardized three-dimensional spectral cube data collected when the patient first wore the new stoma chassis and stores it in the baseline feature sample library. The sample library completely preserves the original spectral spatial features of the chassis without corrosion or fluid accumulation. The global leakage correlation features output from this test are used as time-series samples to be corrected. These samples are combined pairwise with the baseline features in the library to construct training sample pairs, which are then fed into the time-series comparison learning framework. The framework relies on a self-supervised time-series feature alignment training mechanism to complete iterative feature training, continuously reducing the intra-class differences in feature vectors of the same chassis at different stages of use, and increasing the inter-class distance between feature vectors of normal and leaking chassis to optimize network parameters. The entire training process can achieve personalized model adaptation without manually annotating a large number of leakage samples. It can continuously learn the feature differentiation rules of two states: natural aging of colloids and corrosion leakage, accurately distinguishing between normal aging and early leakage failure of the chassis. After training, the system runs baseline drift deviation correction calculation, fully incorporating the baseline drift deviation correction formula: After completing all correction calculations, the system reads the original three-dimensional spectral spatial features extracted by the network at the current detection time during the calculation phase. Original features of the reference sample collected during the initial wearing of the chassis. Select time-series correction weighting coefficients with values ​​ranging from 0.4 to 0.6. The hidden volume of the effluent output from the effluent volume inversion step. Preset warning critical volume threshold The baseline offset component is subtracted by substituting all parameters into the formula. The calculation logic uses the inverted liquid volume as a correction amplitude adjustment factor. The closer the liquid volume value is to the warning critical volume threshold, the stronger the baseline offset component subtraction, effectively offsetting the spectral baseline drift caused by batch differences in colloids and individual differences in human skin color. The output is a corrected leakage feature vector after eliminating baseline deviation. In the data verification step of this embodiment, the feature vectors before and after time-series correction were compared. After correction, the feature fluctuation amplitude of the same chassis in different detection cycles decreased by 60%, and the detection fluctuations caused by individual skin color and colloidal material differences were effectively eliminated.

[0027] The fifth step is to execute graded early warning output. The system adopts a dual-parameter joint judgment mechanism to classify risk levels. The value of the hidden liquid accumulation is used as the core quantitative judgment indicator, and the corrected leakage feature vector is used as the auxiliary feature judgment basis. Two fixed judgment thresholds are set: the first-level early warning judgment threshold is 0.02mL, and the second-level emergency replacement judgment threshold is 0.05mL. The two sets of indicators work together to make up for the defects of missed judgment and misjudgment caused by single parameter judgment, and improve the overall accuracy of risk classification. The system pushes differentiated prompts based on the final risk level. If the risk is determined to be early latent leakage, prompts related to chassis status observation are pushed, guiding patients to increase the frequency of daily chassis fit checks. If the risk is determined to be high-risk leakage, prompts for emergency chassis replacement are pushed, reminding patients to replace the stoma chassis immediately. Simultaneously, relying on the corrected leakage feature vector, the system locks the coordinates of spatial defects within the image, accurately marking the location of colloidal interstitial fluid accumulation, colloidal corrosion areas, and microscopic crack anomalies, with a location positioning error controlled within 2mm. All potential hazard points are simultaneously marked on the device display interface, achieving integrated output of quantitative and graded early warning of leakage hazards and precise defect location. This allows home-based patients to intuitively grasp the location of chassis damage and conduct targeted nursing examinations. In the data verification phase of this embodiment, the results of 30 days of home detection and early warning were statistically analyzed. No early latent leakage hazards were missed, and the early warning time for high-risk leakage was much earlier than the point when leakage could be observed with the naked eye. The coordinates of the marked points matched the actual location of the interstitial fluid accumulation with extremely high accuracy.

[0028] This embodiment utilizes portable devices to meet the self-monitoring needs of ostomy patients at home. It employs multi-band synchronous imaging combined with a complete calibration process to collect standardized three-dimensional spectral cube data. A dual-branch network is used to extract leakage characteristics of colloidal materials and interface structures across dimensions. Optical prior rules are used to filter out invalid interference features, and a dedicated fluid volume quantification algorithm is employed to eliminate interference from body hair and shadows, accurately calculating the value of latent fluid accumulation in the ostomy. A personalized chassis benchmark sample is established through self-supervised temporal comparative learning. Baseline drift deviation correction calculations eliminate baseline fluctuations, and a dual-parameter graded early warning system combining fluid volume and calibration characteristics accurately marks leakage points. The entire process requires no professional medical personnel present. Data verification shows that errors in the acquisition, calculation, and identification stages are controllable, enabling the early detection of minute leakage risks invisible to the naked eye, reducing the probability of skin complications during home care.

[0029] Example 2: This embodiment is applied to a hospital ostomy clinic setting. Medical staff use a fixed medical multispectral imaging device to conduct standardized screening for potential ostomy chassis leakage in batches of post-ostomy patients. The entire process is adapted to the continuous testing needs of large numbers of outpatients, and can simultaneously store spectral data from each patient's tests to establish a personal care record. Standardized data validation is also included to verify the stable identification capability of the entire testing method in scenarios involving multiple populations and batches of colloidal consumables. This provides quantifiable and objective data support for outpatient medical staff to develop personalized nursing intervention plans. Figure 2 As shown.

[0030] The first step involves multispectral image acquisition. Medical staff adjust the medical imaging equipment bracket to ensure that the equipment lens is perpendicular to the plane of the stoma base worn by the patient. The exposure time for each exposure is controlled within 200ms. The equipment simultaneously drives four imaging devices—visible light, 850nm near-infrared, 940nm near-infrared, and 1450nm short-wave infrared—to complete the exposure and acquisition simultaneously. The imaging acquisition range completely covers the entire colloidal sealing area of ​​each patient's stoma base and the surrounding 5mm of skin tissue. It fully captures the spectral information of all key areas that are prone to exudate accumulation and colloidal corrosion, and there will be no loss of imaging of local colloidal areas due to shooting angle deviation. After completing the image acquisition for all patients in the outpatient batch testing, the system performs batch radiometric correction in the background to uniformly eliminate image brightness deviations caused by differences in the hardware response of the four imaging devices. This avoids inconsistencies in the baseline brightness of images from different patients affecting subsequent comparisons. Pixel-level spatial registration is also performed in batches, calibrating the pixel coordinate deviations at the same spatial location in all spectral bands to within a single pixel, ensuring complete overlap of spatial positions across multiple spectral bands. The four corrected images are then stacked according to channel dimensions to generate unique 3D spectral cube data for each patient. This data simultaneously carries the chassis's 2D spatial location information and multi-dimensional spectral response information. Each pixel retains the corresponding full-band spectral curve for the skin and colloid, and is batch-stored in the outpatient patient care record system for easy retrieval and comparison during subsequent follow-up visits. In this embodiment, the data verification phase collected 3D spectral cube data from 120 patients with different skin colors and using different batches of colloid chassis. All sample pixel registration deviations were controlled within a single pixel, and there was no significant distortion in the baseline imaging data for patients with different skin colors. The batch acquisition efficiency meets the daily outpatient schedule.

[0031] The second step involves dual-path feature extraction. The outpatient system sequentially inputs the independent 3D spectral cube data of each patient into a dual-branch spectral-spatial attention convolutional network. The network builds parallel and independent spectral feature analysis branches and spatial feature analysis branches based on the 3D convolutional architecture. The two branches simultaneously read individual 3D spectral cube data and perform independent calculations. The spectral feature analysis branch performs weight adaptive calculations on the spectral dimension data to accurately capture subtle material change signals such as spectral absorption peak shifts and spectral response intensity attenuation caused by long-term contact of the colloid with digestive fluid corrosion. It filters out stray spectral signals unrelated to the chemical failure of the colloid and retains only the effective spectral information that can characterize the corrosion and deterioration of the colloid. The spatial feature analysis branch performs gradient feature extraction on the spatial dimension data to accurately identify microscopic cracks formed at the interface between the colloid and the skin, as well as gaps at the edges of the colloid. It also simultaneously extracts spatial features such as redness and darkening of the skin caused by long-term immersion in exudate. After the two branches complete independent feature analysis, they perform feature channel splicing and fusion to generate the global leakage correlation features of the corresponding patient's chassis. During the network training phase, feature convergence optimization is performed based on the optical prior rules corresponding to the colloidal multilayer transmission mechanism. This actively suppresses image noise and pseudo-features that violate objective optical change laws, retaining only effective leakage features that conform to the changes in colloidal light transmission, liquid spectral attenuation, and material corrosion spectral changes. This corrects feature offset issues caused by free network training, ensuring that the network-extracted features closely match the true optical change laws of latent leakage in the stoma base, and stabilizing the accuracy of fluid quantification calculations under batch detection conditions. In the data verification phase of this embodiment, compared with the output results of a single-branch network in outpatient batch detection, this dual-branch network can stably distinguish spectral differences caused by different skin colors and different colloidal materials, and will not judge the patient's inherent skin pigmentation as a leakage defect. The stability of effective feature extraction in batch detection is improved, such as... Figure 3 As shown.

[0032] The third step involves fluid volume inversion. The system individually calculates the fluid volume of the interstitial effusion for each patient based on optical transmission physical constraints using a fluid volume quantification algorithm. Simultaneously, an interference feature filtering process is implemented. The system first identifies and locates high-texture areas formed by body hair and skin folds in the image using spatial dimensional gradient identification. Weighted attenuation processing is applied to the near-infrared transmission characteristics of these areas to reduce light intensity fluctuations caused by changes in skin texture. Then, the system distinguishes between shadowed areas of the outpatient environment and actual fluid areas based on the ratio of visible to near-infrared light intensity. The inherent difference in the ratio between the two types of areas eliminates miscalculations of fluid volume caused by shadows. After all interference information is filtered, the system calculates the area of ​​the fluid-covered region and the equivalent thickness of the liquid by combining the light intensity attenuation caused by near-infrared light penetrating the colloidal interlayer. The volume of the hidden fluid inside the interstitial effusion in each patient is obtained by multiplying these two parameters. The entire liquid accumulation calculation logic synchronously integrates three types of interactions: colloidal corrosion spectral shift loss, multilayer media transmission loss, and liquid accumulation infrared attenuation loss. Using the measured transmitted light intensity from the equipment as known conditions, it inversely derives the total amount of interlayer liquid, adhering to objective optical attenuation laws throughout the process to ensure consistent and stable liquid accumulation values ​​in outpatient batch testing. In the data verification phase of this embodiment, batch testing was conducted on pre-loaded trace liquid accumulation samples from multiple batches of chassis consumables in the outpatient department. The deviation between the algorithm's output liquid accumulation value and the actual added fluid volume was stable and controllable, and there was no systematic deviation in the test results for consumables with different colloidal thicknesses.

[0033] The fourth step involves time-varying feature correction. The system retrieves standardized three-dimensional spectral cube data collected during each patient's initial outpatient visit when wearing a brand-new chassis and stores it in a dedicated benchmark feature sample library. The sample library completely preserves the original spectral spatial features of the corresponding patient's chassis, which are free from corrosion and fluid accumulation. The global leakage correlation features generated during this outpatient visit are used as time-series samples to be corrected. These samples are paired with the benchmark samples to construct training sample pairs, which are then fed into the time-series comparison learning framework. The framework relies on a self-supervised time-series feature alignment training mechanism to complete iterative feature training, continuously reducing the intra-class differences in feature vectors of the same patient and chassis at different follow-up visit cycles, and increasing the inter-class distance between feature vectors of normal and leaking chassis to optimize network parameters. The entire training process can complete personalized adaptation of a single patient model without requiring medical staff to manually label a large number of leakage samples. It can continuously learn the feature differentiation rules of two states: natural aging of colloids and corrosion leakage, and accurately distinguish between normal aging wear and early leakage failure of the chassis. After training, the system performs baseline drift correction calculations. During the calculation, the latent effusion volume output from the effusion volume inversion step is used as a correction amplitude adjustment factor. The closer the effusion value is to the warning critical volume threshold, the stronger the baseline offset component is deducted. This counteracts the spectral baseline drift caused by differences in different colloid batches and individual patient skin color. The corrected leakage feature vector after eliminating baseline deviation is then stored in the patient's electronic nursing record. In the data verification stage of this embodiment, the feature vectors from three follow-up visits of the same patient are retrieved. After time-series correction, the fluctuation amplitude of the feature baseline is reduced, and the test results from different follow-up visit cycles have longitudinal comparative reference value.

[0034] The fifth step is to implement graded early warning output. The outpatient system uniformly adopts a dual-parameter joint judgment mechanism to complete the judgment of the chassis risk level of each patient. The value of the hidden fluid accumulation is used as the core quantitative judgment indicator, and the corrected leakage feature vector is used as the auxiliary feature judgment basis. Two fixed judgment thresholds are set: the first-level early warning judgment threshold is 0.02mL, and the second-level emergency replacement judgment threshold is 0.05mL. The two indicators work together to make up for the defects of missed judgment and misjudgment caused by single parameter judgment, and improve the accuracy of outpatient batch risk classification. The system pushes differentiated prompts based on each patient's risk level. When an early, latent leakage risk is identified, a chassis status observation prompt is sent to the medical staff's terminal, allowing them to instruct patients to shorten the interval between home chassis self-inspections. When a high-risk leakage risk is identified, an emergency chassis replacement prompt is sent, and medical staff guide the patient to complete the chassis replacement on the spot. Simultaneously, relying on the corrected leakage feature vector, the system locks the spatial defect coordinates, accurately marking the location of colloidal interlayer fluid accumulation, colloidal corrosion areas, and microscopic crack abnormalities, with a location positioning error controlled within 2mm. The terminal simultaneously displays images of all marked potential hazard points, and medical staff combine the quantified fluid accumulation values ​​with the location results to develop targeted nursing intervention plans. In the data verification phase of this embodiment, outpatient batch screening and early warning data were statistically analyzed, and all early, trace latent leaks were identified in advance. No high-risk leakage hazards were missed, and the matching degree between the point coordinates and the actual interlayer fluid and colloidal corrosion locations met clinical usage standards.

[0035] This embodiment is adapted to the scenario of large-scale centralized screening of patients in hospital ostomy clinics. It utilizes fixed medical imaging equipment to complete multispectral acquisition and correction in batches, generating standardized three-dimensional spectral cube data. Through a dual-branch network, it extracts two types of leakage features: colloidal corrosion and interface cracks. It combines optical prior rules to suppress false feature interference and uses a fluid volume quantification algorithm to remove interference from body hair and environmental shadows, stably outputting the value of interstitial fluid. It establishes patient-specific baseline samples through self-supervised temporal comparison learning, and eliminates baseline fluctuations caused by skin color and colloidal batches through baseline drift deviation correction calculations. It accurately marks leakage points using a dual-parameter graded early warning system. Data validation demonstrates strong batch detection stability, providing medical staff with quantitative detection data to assist in developing personalized clinical nursing plans and reduce the workload of managing ostomy skin complications in outpatient clinics.

[0036] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A deep learning-based early visual warning method for stoma chassis leakage, characterized in that, The specific steps of this method are as follows: S100, Multispectral Image Acquisition: Multi-band synchronous imaging processing is completed on the colloidal bonding area of ​​the stoma chassis, and after radiometric correction and spatial pixel registration, three-dimensional spectral cube data is generated by stacking. S200, Dual-path feature extraction: Input the three-dimensional spectral cube data into a preset deep learning network to analyze the spectral dimension of colloidal material variation features and the spatial dimension of interface defect features respectively. After the two types of features are fused, the global leakage correlation features are generated. S300, Liquid Volume Inversion: Based on the physical constraints of optical transmission, the correlation characteristics of leakage in the whole domain are converged and optimized. After removing irrelevant interference information, the liquid volume inside the interlayer is quantitatively solved, and the quantifiable value of the hidden liquid volume in the interlayer is output. S400, Time-varying Feature Correction: Input the global leakage correlation features and the latent liquid volume values ​​into the time-series comparison learning framework to complete the baseline drift deviation correction calculation and output the corrected leakage feature vector after eliminating the baseline deviation. S500, graded early warning output: Combines the hidden liquid accumulation volume value with the corrected leakage feature vector to complete the risk level determination, pushes corresponding prompt information in a graded manner and marks the abnormal points, and completes the early warning of leakage in all stages.

2. The method for early visual warning of stoma chassis leakage based on deep learning according to claim 1, characterized in that, In step S100, the multi-band synchronous imaging processing uses the same exposure clock to trigger simultaneous exposure of four imaging devices: visible light, 850nm near-infrared, 940nm near-infrared, and 1450nm short-wave infrared. The duration of a single exposure is controlled within 200ms. During the acquisition process, the device lens is kept perpendicular to the chassis plane, and the acquisition range covers the entire colloid sealing area of ​​the stoma chassis and the skin tissue within a 5mm radius around it. After the acquisition is completed, radiometric correction is performed on the images of each band, and then image alignment is completed through pixel-level spatial registration. The pixel coordinate deviation of the same physical location in different bands is calibrated to within a single pixel. The four sets of corrected images are stacked according to the channel dimension to form three-dimensional spectral cube data.

3. The method for early visual warning of stoma chassis leakage based on deep learning according to claim 1, characterized in that, In step S200, the preset deep learning network is a dual-branch spectral-spatial attention convolutional network adapted to three-dimensional spectral cube data. It is built on a three-dimensional convolutional architecture and includes parallel and independent spectral feature analysis branches and spatial feature analysis branches. The spectral feature analysis branch performs weight adaptive calculations on the spectral dimension data of the three-dimensional spectral cube to capture subtle material changes such as spectral absorption peak shift and spectral response intensity decay that occur after the stoma base colloid is eroded by digestive fluid, and filters out effective spectral signals that correspond only to the chemical failure of the colloid. The spatial feature analysis branch performs gradient feature extraction on the spatial dimension data of the three-dimensional spectral cube to identify microscopic cracks and gaps at the interface between the colloid and the skin, and at the same time extracts the redness and darkening features of the skin surface caused by exudate immersion.

4. The method for early visual warning of stoma chassis leakage based on deep learning according to claim 1, characterized in that, In step S300, the amount of liquid accumulated inside the interlayer is quantitatively calculated using the formula. Complete, in the formula The overall transmitted light intensity of the chassis after multi-band fusion; The baseline value of transmitted light intensity in the chassis contact area under conditions without liquid accumulation; Let be the attenuation coefficient corresponding to the i-th near-infrared characteristic band; denoted as the spectral absorption peak shift caused by corrosion of the colloid in the corresponding band; K is the optical path correction constant of the colloidal interlayer. This refers to the volume of the hidden liquid accumulation inside the chassis interlayer.

5. The method for early visual warning of stoma chassis leakage based on deep learning according to claim 1, characterized in that, In step S300, interference feature filtering is performed simultaneously during the liquid volume inversion process. First, high-texture regions corresponding to body hair and skin folds are identified and located through spatial dimension gradient identification. Weighted attenuation processing is performed on the near-infrared transmission features of high-texture regions. For environmental shadow interference, shadow regions and real liquid regions are distinguished by the light intensity ratio characteristics of the visible light band and the near-infrared band. The band ratio of the two types of regions has a fixed difference, thereby eliminating the miscalculation of liquid volume caused by shadows. After the interference filtering is completed, the area of ​​the liquid-covered region and the equivalent thickness of the liquid are calculated by combining the light intensity attenuation degree of near-infrared light after penetrating the colloidal interlayer. The final hidden liquid volume value is obtained by multiplying the area of ​​the liquid-covered region and the equivalent thickness of the liquid.

6. The method for early visual warning of stoma chassis leakage based on deep learning according to claim 1, characterized in that, In step S400, the time-series comparison learning framework adopts a self-supervised time-series feature alignment training mechanism to construct a benchmark feature sample library using standardized three-dimensional spectral cube data collected when the user first wears the new stoma chassis, and to store the original spectral spatial features of the chassis when it is initially free of corrosion and fluid accumulation. The leakage correlation features collected by users at different usage times are used as time series samples to be corrected. These samples are combined with the baseline feature samples to construct training sample pairs, which are then fed into the network to perform feature alignment learning. The feature iterative training is completed by relying on the chassis’s own time series state changes. The network continuously learns the feature difference rules of normal aging and corrosion leakage of colloids, and distinguishes between normal colloid state and early failure state.

7. The method for early visual warning of stoma chassis leakage based on deep learning according to claim 1, characterized in that, In step S400, the baseline drift deviation correction calculation is performed using the formula Complete, in the formula This is the corrected leakage feature vector after baseline drift correction; The original three-dimensional spectral spatial features extracted by the network at the current detection time; These are the original features of the baseline sample collected when the chassis was first worn; Adjust the weighting coefficients for time series; The implicit liquid volume output by the liquid volume inversion step; This is a critical volume threshold for early warning.

8. The method for early visual warning of stoma chassis leakage based on deep learning according to claim 1, characterized in that, In step S500, the risk level determination adopts a dual-parameter joint determination mechanism, using the volume of latent liquid accumulation as the core quantitative determination indicator and the corrected leakage feature vector as the auxiliary feature determination basis. Two fixed determination thresholds are set. The first-level early warning determination threshold is 0.02mL. When the volume of latent liquid accumulation is ≥0.02mL and the corrected leakage feature vector detects a colloidal spectral corrosion change signal, it is determined to be an early latent leakage risk. The second-level emergency replacement determination threshold is 0.05mL. When the volume of latent liquid accumulation is ≥0.05mL, regardless of whether the corrected leakage feature vector detects a colloidal spectral corrosion change signal, it is determined to be a high-risk leakage risk, and a visible excrement leakage is about to occur. The dual-parameter joint determination can avoid the defects of single-indicator determination.

9. The method for early visual warning of stoma chassis leakage based on deep learning according to claim 1, characterized in that, In step S500, the prompt information is pushed in a differentiated manner based on the risk level determination result. For early hidden leakage risks, a chassis status observation warning prompt is pushed to remind the user to increase the frequency of chassis fit inspection. In response to high-risk leakage, an emergency replacement warning message for the chassis is pushed out, prompting users to replace the stoma chassis immediately. At the same time, based on the spatial defect coordinates locked by the correction leakage feature vector, the abnormal locations of colloidal interlayer fluid accumulation, colloidal corrosion, and microscopic crack abnormalities are marked.

Images