An icg contrast multimodal fusion postoperative upper limb lymphedema management system

Abstract

The application discloses an ICG contrast multi-modal fusion postoperative upper limb lymphedema management system, which comprises an ICG dynamic lymphangiography acquisition module, an upper limb circumference volume measurement module, a traditional Chinese medicine multi-modal data acquisition module, a western medicine clinical data acquisition module, an edema property stratification judgment module, a Chinese and western medicine time sequence data fusion module and an edema management decision module, extracts time sequence parameters such as lymphatic vessel filling speed, imaging delay time and reflux duration through an infrared fluorescence imager after injection of a contrast agent; combines upper limb circumference volume calculation with traditional Chinese medicine tongue appearance, pulse appearance and time sequence score data of syndrome elements, realizes deep fusion of Chinese and western medicine data by adopting an attention mechanism embedded with a time sequence correlation matrix; and outputs an individualized treatment scheme and prognosis risk evaluation of the combination of Chinese and western medicine based on liquid / solid edema stratification judgment through a reinforcement learning model.

Description

Technical Field

[0001] This invention relates to the field of medical information processing technology, specifically to a postoperative upper limb lymphedema management system based on ICG angiography multimodal fusion. Background Technology

[0002] Breast cancer-related lymphedema (BCRL) is a core complication of breast cancer treatment, with an incidence rate of approximately 20% to 50% after surgery, affecting over 50 million patients worldwide. Once BCRL develops, it is difficult to cure, causing not only upper limb swelling and limited function but also pain and infection, severely impacting patients' postoperative quality of life. Therefore, early and accurate diagnosis of BCRL, individualized treatment planning, and scientific prognostic assessment are crucial for improving patient outcomes.

[0003] Currently, indocyanine green (ICG) lymphangiography is an important technique for assessing the dynamic function of the lymphatic system, enabling non-invasive visualization of lymphatic vessel course and reflux. Meanwhile, the temporal evolution of TCM syndrome differentiation (such as qi stagnation, blood stasis, dampness, qi deficiency, spleen deficiency, and kidney deficiency) has been proven to be closely related to the progression of BCRL, serving as an important basis for guiding treatments such as herbal medicine, acupuncture, and cupping. However, in clinical practice, ICG lymphatic function temporal parameters (such as lymphatic vessel filling velocity, imaging delay time, and reflux duration) and TCM syndrome temporal scoring data belong to different assessment systems, lacking a physiological and pathological correlation model pre-trained on a large amount of clinical data. Existing technologies, even when introducing artificial intelligence for multimodal data fusion, only static feature splicing or simple weighting of each modality's data is performed, resulting in the fused feature vector failing to reflect clinically validated coupling patterns such as "exacerbation of dampness → delayed lymphatic reflux" and "progression of qi deficiency → decreased filling velocity."

[0004] Due to the lack of the aforementioned temporal correlation and fusion capabilities, existing systems struggle to achieve dynamic and precise stratification of the pathological nature of BCRL edema. Clinically, patients with the same volume of swelling may exhibit drastically different pathological components: "liquid" edema, primarily composed of lymph and inflammation, responds well to conservative treatments (such as pressure therapy and traditional Chinese medicine for promoting diuresis); while "solid" edema, mainly characterized by fat deposition and fibrosis, requires surgical intervention. Current technology cannot dynamically determine the nature of edema through combined analysis of ICG dynamic parameters and temporal changes in TCM syndrome elements, leading to ambiguity in the indications for conservative treatment and surgery, a lack of precise basis for treatment selection, and patients often missing the optimal intervention window, exacerbating disease progression and increasing the medical burden.

[0005] Therefore, there is an urgent need for a system that can deeply correlate and integrate the dynamic temporal parameters of ICG lymphatic drainage with the temporal scores of TCM syndrome elements. By introducing prior knowledge of the correlation between TCM and Western medicine temporal parameters based on clinical diagnosis and treatment data, the fusion model can capture the synergistic evolution of "syndrome element-function" with a focus, thereby achieving dynamic and accurate stratification of the edema properties of BCRL and providing a reliable basis for individualized treatment decisions. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides a postoperative upper limb lymphedema management system based on ICG angiography multimodal fusion, which solves the problems mentioned in the background technology.

[0007] To achieve the above objectives, the present invention provides the following technical solution: a postoperative upper limb lymphedema management system based on ICG angiography multimodal fusion, comprising,

[0008] The ICG dynamic lymphangiography acquisition module is used to acquire dynamic image sequences at three time points: 15 minutes, 30 minutes and 60 minutes after the injection of indocyanine green contrast agent using an infrared fluorescence imager, and to extract the timing parameters of lymphatic drainage function, including lymphatic filling speed, imaging delay time and reflux duration.

[0009] The upper limb circumference volume measurement module is used to collect segmental circumference measurement data along the affected and healthy upper limbs at preset intervals, and calculate the edema volume of each segment of the upper limb and the total volume swelling rate based on the frustum volume formula.

[0010] The TCM multimodal data acquisition module is used to collect quantitative data including the color values ​​of the three primary colors of tongue appearance (red, green, and blue), pulse frequency and amplitude, and generate TCM syndrome differentiation labels and time-series score data of six syndrome elements: Qi stagnation, blood stasis, dampness, Qi deficiency, spleen deficiency and kidney deficiency.

[0011] The Western medicine clinical data acquisition module is used to collect basic patient information, surgical and treatment-related information, auxiliary treatment information, and time-series scoring data of the Patient Reported Outcome Scale (PTS). Among them, the time-series scoring data of the PTS includes the Breast Cancer-Related Lymphedema Symptom Index Scale, the Arm-Shoulder-Hand Dysfunction Scale, the Upper Limb Lymphedema Quality of Life Scale, the Hospital Anxiety and Depression Scale, and the Herth Hope Scale.

[0012] The edema nature stratification determination module is used to determine whether the main pathological component of edema is liquid edema or solid edema based on the time sequence parameters of lymphatic drainage function output by the ICG dynamic lymphangiography acquisition module, the total volume swelling rate output by the upper limb circumference volume measurement module, and externally input physical examination / MRI / color Doppler ultrasound data.

[0013] The TCM and Western medicine time-series data fusion module is used to fuse the time-series parameters of the lymphatic drainage function, the total volume swelling rate, the time-series score data of the syndrome elements, the time-series score data of the patient reported outcome scale, and the Western medicine clinical data at the time-series feature level to generate a fused time-series feature vector. The time-series feature-level fusion adopts an attention mechanism that embeds a time-series correlation matrix. The time-series correlation matrix is ​​pre-trained based on the time-series correlation of existing diagnosis and treatment data of TCM syndrome element scores and ICG dynamic feature parameters.

[0014] The edema management decision module is used to output lymphedema risk warning information, multidisciplinary collaborative management suggestions combining traditional Chinese and Western medicine, and prognostic risk prediction information based on the fusion time-series feature vector and the stratification results of the edema nature stratification judgment module.

[0015] Furthermore, the ICG dynamic lymphangiography acquisition module includes:

[0016] The contrast agent injection unit is used to inject indocyanine green water-based solution into the lymphatic-rich area of ​​the second finger web or the dorsum of the foot of the affected limb, at a dose of 0.1 to 0.3 ml.

[0017] Near-infrared excitation imaging unit, used to continuously acquire dynamic fluorescence images at a rate of 30 frames per second in the 750 to 1000 nanometer wavelength band;

[0018] The temporal frame extraction unit is used to extract static keyframes at 15 minutes, 30 minutes and 60 minutes after contrast agent injection from the dynamic fluorescence image, which correspond to the initial stage of lymphatic vessel filling, dynamic drainage period and delayed imaging period, respectively.

[0019] The lymphatic function quantification unit is used to calculate the lymphatic vessel continuity index, imaging delay time, and reflux area ratio based on the static keyframe, and output the stage 0 to stage 4 results according to the ICG lymphography staging standard.

[0020] Furthermore, the upper limb circumference volume measuring module is spaced 8 cm apart along the upper limb, and the formula for the volume of the frustum is:

[0021]

[0022] Where V represents the volume of the frustum, in cubic centimeters; h represents the distance between two adjacent measurement points, in centimeters; C1 represents the circumference of the upper limb of the upper measurement point between two adjacent measurement points, in centimeters; C2 represents the circumference of the upper limb of the lower measurement point between two adjacent measurement points, in centimeters; 0.0265 is the reciprocal correction factor of pi, calculated from 1 / (12π), and is dimensionless.

[0023] Furthermore, the edema nature stratification determination module executes the following determination logic:

[0024] Determination of fluid edema: When the patient complains of milder symptoms in the morning and more severe symptoms in the evening, physical examination reveals obvious pitting with a pressure depth greater than or equal to 6 mm, MRI shows subcutaneous soft tissue swelling and widening of the interstitial spaces, and color Doppler ultrasound shows homogeneous subcutaneous tissue echoes, the fluid edema stratification result is output.

[0025] Determination of solid edema: When physical examination reveals non-pitting edema with a pressure depth of less than 6 mm, or accompanied by acanthosis, verrucous hyperplasia, hyperkeratosis, MRI showing fat deposition and fibrosis with high signal on T1-weighted images and low signal on T2-weighted images, and color Doppler ultrasound showing thickening of the skin and fat layers with enhanced echogenicity, the solid edema stratification result is output.

[0026] Furthermore, the traditional Chinese and Western medicine time-series data fusion module includes:

[0027] The temporal correlation matrix construction unit is used to extract multiple historical diagnosis and treatment data from the encrypted medical record database. Pearson correlation analysis is used to calculate the correlation coefficient between the time-series scores of TCM syndrome elements and the dynamic feature parameters of ICG. The prior probability association rule mining algorithm is used to extract frequent itemsets of the temporal changes of syndrome elements and the temporal changes of ICG. The statistically significant association rules are encoded into a temporal correlation matrix.

[0028] The attention-series fusion unit is used to map the ICG image time-series feature vector, the TCM syndrome element time-series feature vector, the patient-reported outcome scale time-series feature vector, and the Western medicine pathology feature vector to the same hidden layer dimension. It uses a scaled dot product attention mechanism to calculate the time-series attention weights, superimposes the time-series correlation matrix as a bias term, and outputs the fused time-series feature vector.

[0029] Furthermore, the edema management decision module includes:

[0030] The latent period early warning unit is used to output a subclinical BCRL early warning prompt when the patient's upper limb circumference difference is less than 2 cm and the volume difference is less than 10%, but the timing parameters of the lymphatic drainage function are abnormal.

[0031] The risk warning information output unit is used to output the BCRL diagnosis probability and the staging recommendations of the International Society of Lymphoma based on the fused temporal feature vector.

[0032] The management suggestion information generation unit is used to prioritize conservative treatment when the edema nature stratification judgment module outputs liquid edema, and recommend surgical intervention combined with postoperative comprehensive edema reduction treatment when it outputs solid edema. The treatment recommendation unit adopts a multi-objective optimization decision model based on reinforcement learning. The action space includes permutations and combinations of Western medicine physical therapy sub-actions, Western medicine surgical treatment sub-actions, traditional Chinese medicine external treatment sub-actions, and traditional Chinese medicine internal treatment sub-actions. The reward function is a weighted sum of edema regression rate, patient-reported outcome scale quality of life improvement, and treatment safety. The quality of life improvement is quantified by the difference in total score before and after treatment of the upper limb lymphedema quality of life scale or breast cancer treatment function assessment scale. The unit outputs the top three treatment options in descending order of decision confidence and their corresponding evidence-based text.

[0033] The prognostic risk prediction information output unit is used to output the 1-year and 2-year recurrence probability prediction based on time-series follow-up data.

[0034] Furthermore, the management suggestion information generation unit includes:

[0035] The contraindication screening subunit is used to automatically filter action plans containing corresponding Chinese medicine or surgery when there is a history of allergy to Chinese medicine or coagulation dysfunction in the patient's electronic medical record.

[0036] The evidence-based evidence generation subunit is used to call the historical multidisciplinary collaborative case library of integrated traditional Chinese and Western medicine, retrieve the efficacy data of cases similar to the current patient's syndrome element time series score and ICG dynamic feature parameters, and generate the top three treatment plans and corresponding evidence-based evidence texts in descending order of decision confidence. The decision confidence is determined based on the decision quality value output by the reinforcement learning model.

[0037] Furthermore, the prognostic risk prediction information output unit adopts a hybrid architecture of Cox proportional hazards regression and long short-term memory network. The long short-term memory network receives time-series score data of the Breast Cancer Treatment Functional Assessment Scale, Arm, Shoulder and Hand Dysfunction Scale, Upper Limb Lymphedema Quality of Life Scale, Hospital Anxiety and Depression Scale, Herth Hope Scale and Daily Living Ability Index during the patient's follow-up period to capture dynamic trends. The Cox proportional hazards regression fits a risk proportional function with the dynamic trend characteristics and the patient's baseline covariates, and outputs the predicted values ​​of recurrence probability at 1 year and 2 years.

[0038] Furthermore, the system is deployed on a local hospital server. The ICG dynamic lymphangiography acquisition module, the upper limb circumference and volume measurement module, the traditional Chinese medicine multimodal data acquisition module, and the Western medicine clinical data acquisition module are connected to the traditional Chinese and Western medicine time-series data fusion module through a standard data interface. The edema management decision module outputs structured prediction results to the application layer through a representation state transition application programming interface. The application layer includes a data entry module, an image analysis module, a diagnostic suggestion module, a treatment recommendation module, a prognostic assessment module, and a follow-up management module.

[0039] This invention provides a postoperative upper limb lymphedema management system based on ICG angiography multimodal fusion, which has the following beneficial effects:

[0040] This system integrates time-series data from both Traditional Chinese Medicine (TCM) and Western medicine using a TCM-Western medicine time-series data fusion module. This module combines time-series parameters such as lymphatic filling velocity, imaging delay time, and reflux duration extracted by the ICG dynamic lymphangiography acquisition module with time-series scores of six syndrome elements (Qi stagnation, blood stasis, dampness, Qi deficiency, spleen deficiency, and kidney deficiency) generated by the TCM multimodal data acquisition module, along with total edema rate output by the upper limb circumference and volume measurement module and time-series scores of the Patient Reported Outcome Scale (PRI) collected by the Western medicine clinical data acquisition module. The fusion employs an attention mechanism embedding a time-series correlation matrix, which is pre-trained based on the time-series correlation between TCM syndrome element scores and ICG dynamic feature parameters from existing diagnostic and treatment data. Therefore, the system can identify and quantify the synergistic evolution between dynamic changes in TCM syndrome elements and dynamic changes in lymphatic function. This allows the generated fused time-series feature vector to directly reflect the clinically validated physiological and pathological coupling relationship, providing a clinically sound foundation for subsequent edema characterization.

[0041] Based on the fused temporal feature vector, this system incorporates a stratified edema determination module. This module, based on ICG lymphatic drainage function temporal parameters, total volume swelling rate, and externally input physical examination, MRI, and color Doppler ultrasound data, executes a clear logic for determining liquid and solid edema. Liquid edema is determined by criteria including a complaint of worsening in the evening and improving morning symptoms, significant pitting edema with greater depth of pressure, MRI showing subcutaneous soft tissue swelling and widened interstitial spaces, and color Doppler ultrasound showing homogeneous subcutaneous tissue echogenicity. Solid edema is determined by criteria including non-pitting edema with less depth of pressure, accompanied by acanthosis, verrucous hyperplasia, or hyperkeratosis, MRI showing fat deposition and fibrosis, and color Doppler ultrasound showing thickening and increased echogenicity of the skin and fat layers. Compared to existing technologies that rely solely on volume differences or qualitative observations from a single image for empirical judgment, this system can dynamically output precise stratification results, indicating whether edema is primarily composed of liquid components such as lymph and inflammation, or solid components such as fat deposition and fibrosis. This allows the determination of the pathological nature of edema to move from vague experience to quantitative precision.

[0042] The edema management decision-making module of this system outputs lymphedema risk warning information, multidisciplinary collaborative management suggestions combining traditional Chinese and Western medicine, and prognostic risk prediction information based on the stratification results of the fusion time-series feature vector and edema nature stratification module. Specifically, when the stratification result is fluid-type edema, the system prioritizes conservative treatment; when the stratification result is solid-type edema, the system recommends surgical intervention combined with postoperative comprehensive swelling reduction treatment. Thus, this system provides clinicians with clear treatment pathway guidance: for patients with fluid-type edema mainly characterized by lymphatic fluid and inflammation, timely conservative interventions such as pressure therapy and traditional Chinese medicine for diuresis are implemented to avoid unnecessary surgical trauma; for patients with solid-type edema mainly characterized by fat deposition and fibrosis, early assessment of surgical indications is conducted to avoid delaying the optimal surgical time due to prolonged ineffective conservative treatment. This closed-loop decision-making mechanism effectively solves the problem of ambiguity between conservative treatment and surgical indications in existing technologies.

[0043] This system extracts sensitive temporal parameters such as lymphatic vessel filling speed, imaging delay time, and reflux duration through the ICG dynamic lymphangiography acquisition module. When the patient's upper limb circumference and volume difference have not yet reached the clinical diagnostic criteria, but the temporal parameters of lymphatic drainage function are already abnormal, the edema management decision module can output a pre-existing warning for subclinical lymphedema. This allows the system to identify early lymphatic function impairment before the patient develops visible swelling, shifting the intervention time from the symptomatic stage to the latent stage, significantly improving patient prognosis.

[0044] The management suggestion information generation unit of this system employs a multi-objective optimization decision-making model based on reinforcement learning. The action space encompasses various combinations of Western medicine physical therapy, Western medicine surgical treatment, and traditional Chinese medicine external and internal therapies. The reward function comprehensively considers edema reduction rate, improvement in quality of life, and treatment safety. The system outputs the top three treatment options, sorted in descending order of decision confidence. It also retrieves case efficacy data from a historical multidisciplinary collaborative case database of integrated traditional Chinese and Western medicine that are similar to the current patient's syndrome element time-series scores and ICG dynamic feature parameters, generating corresponding evidence-based text. This mechanism allows clinicians not only to receive recommended treatment plans but also to understand the efficacy basis of those plans, significantly enhancing the system's clinical acceptability and decision-making support value.

[0045] The prognostic risk prediction information output unit of this system employs a hybrid architecture of Cox proportional hazards regression and long short-term memory (LSTM) networks. The LSTM network receives time-series score data from multiple scales during the patient's follow-up period to capture dynamic trends. Cox proportional hazards regression fits a hazard proportion function based on the dynamic trend characteristics and the patient's baseline covariates, outputting predicted recurrence probabilities for one and two years. Compared to existing technologies that rely on static prognostic judgments based on a single assessment, this system can dynamically update prediction results using time-series follow-up data, providing a scientific basis for clinical adjustments to treatment intensity and follow-up frequency. Attached Figure Description

[0046] Figure 1 This is a block diagram of a postoperative upper limb lymphedema management system based on ICG angiography multimodal fusion according to the present invention;

[0047] Figure 2 This is a schematic diagram of the upper limb circumference measurement interval in a postoperative upper limb lymphedema management system based on ICG angiography multimodal fusion according to the present invention. Detailed Implementation

[0048] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0049] Example 1

[0050] Please see Figure 1 This invention provides a postoperative upper limb lymphedema management system based on ICG contrast imaging multimodal fusion, comprising an ICG dynamic lymphangiography acquisition module, used to acquire dynamic image sequences at three time points: 15 minutes, 30 minutes, and 60 minutes after injection of indocyanine green contrast agent using an infrared fluorescence imager, and extracting time-series parameters of lymphatic drainage function, including lymphatic filling speed, imaging delay time, and reflux duration;

[0051] The upper limb circumference volume measurement module is used to collect segmental circumference measurement data along the affected and healthy upper limbs at preset intervals, and calculate the edema volume of each segment of the upper limb and the total volume swelling rate based on the frustum volume formula.

[0052] The TCM multimodal data acquisition module is used to collect quantitative data including the color values ​​of the three primary colors of tongue appearance (red, green, and blue), pulse frequency and amplitude, and generate TCM syndrome differentiation labels and time-series score data of six syndrome elements: Qi stagnation, blood stasis, dampness, Qi deficiency, spleen deficiency and kidney deficiency.

[0053] The Western medicine clinical data acquisition module is used to collect basic patient information, surgical and treatment-related information, auxiliary treatment information, and time-series scoring data of the Patient Reported Outcome Scale (PTS). Among them, the time-series scoring data of the PTS includes the Breast Cancer-Related Lymphedema Symptom Index Scale, the Arm-Shoulder-Hand Dysfunction Scale, the Upper Limb Lymphedema Quality of Life Scale, the Hospital Anxiety and Depression Scale, and the Herth Hope Scale.

[0054] The edema nature stratification determination module is used to determine whether the main pathological component of edema is liquid edema or solid edema based on the time sequence parameters of lymphatic drainage function output by the ICG dynamic lymphangiography acquisition module, the total volume swelling rate output by the upper limb circumference volume measurement module, and externally input physical examination / MRI / color Doppler ultrasound data.

[0055] The TCM and Western medicine time-series data fusion module is used to fuse the time-series parameters of the lymphatic drainage function, the total volume swelling rate, the time-series score data of the syndrome elements, the time-series score data of the patient reported outcome scale, and the Western medicine clinical data at the time-series feature level to generate a fused time-series feature vector. The time-series feature-level fusion adopts an attention mechanism that embeds a time-series correlation matrix. The time-series correlation matrix is ​​pre-trained based on the time-series correlation of existing diagnosis and treatment data of TCM syndrome element scores and ICG dynamic feature parameters.

[0056] The edema management decision module is used to output lymphedema risk warning information, multidisciplinary collaborative management suggestions combining traditional Chinese and Western medicine, and prognostic risk prediction information based on the fusion time-series feature vector and the stratification results of the edema nature stratification judgment module.

[0057] In this embodiment, the ICG lymphangiography multimodal fusion-based intelligent diagnosis and treatment system for lymphedema is deployed on a local hospital server. The recommended hardware configuration is an Intel Core i7 processor and 16 gigabytes of memory, while a lightweight version for primary care is compatible with an Intel Core i5 processor and 8 gigabytes of memory. The system connects to an infrared fluorescence imager via a standard DICOM interface, and to a tongue image acquisition device and pulse imager via a USB interface. It also captures Western medicine clinical data through the hospital information system interface. This system is specifically designed for BCRL diagnosis and treatment scenarios. Its core input is the coupling of dynamic ICG lymphangiography image sequences and upper limb circumference and volume measurement data. For the first time, it introduces TCM tongue image RGB values, pulse wave signals, and syndrome element time-series scores as independent data dimensions, forming a data acquisition architecture not disclosed in existing technologies.

[0058] The system adopts a six-layer architecture: an ICG dynamic lymphangiography acquisition module, an upper limb circumference and volume measurement module, a traditional Chinese medicine multimodal data acquisition module, a Western medicine clinical data acquisition module, an edema nature stratification determination module, a traditional Chinese and Western medicine time-series data fusion module, and an edema management decision module. The first four modules constitute the data acquisition layer, which writes multi-dimensional raw data into an encrypted database after cleaning, normalization, and anonymization through a standard interface. The edema nature stratification determination module performs liquid / solid edema state determination based on multi-physical quantity coupling. The traditional Chinese and Western medicine time-series data fusion module reads standardized time-series data from the encrypted database and generates a fused time-series feature vector through an attention mechanism embedding a time-series correlation matrix. The edema management decision module is deployed on the TensorFlow Serving inference service and includes a latent period warning unit, a risk warning information output unit, a management suggestion information generation unit, and a prognostic risk prediction information output unit. It receives the fused time-series feature vector through a RESTful API and returns structured diagnosis and treatment decision results. The application layer is developed based on the Vue.js front-end framework and the Flask back-end framework, and manages patient diagnosis and treatment process data through a MySQL relational database.

[0059] Example 2

[0060] The ICG dynamic lymphangiography acquisition module includes:

[0061] The contrast agent injection unit is used to inject indocyanine green water-based solution into the lymphatic-rich area of ​​the second finger web or the dorsum of the foot of the affected limb, at a dose of 0.1 to 0.3 ml.

[0062] Near-infrared excitation imaging unit, used to continuously acquire dynamic fluorescence images at a rate of 30 frames per second in the 750 to 1000 nanometer wavelength band;

[0063] The temporal frame extraction unit is used to extract static keyframes at 15 minutes, 30 minutes and 60 minutes after contrast agent injection from the dynamic fluorescence image, which correspond to the initial stage of lymphatic vessel filling, dynamic drainage period and delayed imaging period, respectively.

[0064] The lymphatic function quantification unit is used to calculate the lymphatic vessel continuity index, imaging delay time, and reflux area ratio based on the static keyframe, and output the stage 0 to stage 4 results according to the ICG lymphography staging standard.

[0065] Specifically, the ICG dynamic lymphangiography acquisition module in this embodiment is equipped with an infrared fluorescence imager, and the contrast agent used is indocyanine green water-based solution. This module acquires dynamic functional images rather than static structural images, focusing on extracting static keyframe sequences at three time points: 15 minutes, 30 minutes, and 60 minutes after contrast agent injection. These correspond to the initial lymphatic filling phase, the dynamic drainage phase, and the delayed contrast enhancement phase, respectively, and are used to assess the temporal changes in lymphatic drainage function rather than to identify tumor anatomical lesions.

[0066] The contrast agent injection unit performs injections in areas rich in lymphatic vessels, such as the second finger web or the dorsalis pedis of the affected limb, at a dose of 0.1 to 0.3 ml. The near-infrared excitation imaging unit continuously acquires dynamic fluorescence images at a rate of 30 frames per second in the 750 to 1000 nm wavelength band for 60 minutes, totaling 108,000 frames. The temporal frame extraction unit extracts keyframes from the dynamic image stream according to time nodes: the 15-minute keyframe corresponds to frame 27,000, the 30-minute keyframe corresponds to frame 54,000, and the 60-minute keyframe corresponds to frame 108,000.

[0067] The lymphatic function quantization unit calculates the following timing parameters based on three-phase keyframes:

[0068] Lymphatic vessel filling rate:

[0069]

[0070] Where v represents the lymphatic vessel filling rate, measured in millimeters per minute; This indicates the length of lymphatic vessels visualized in a 30-minute keyframe, in millimeters. The value 15 indicates the length of lymphatic vessels visible in a 15-minute keyframe, in millimeters; 15 indicates the time interval between 15 minutes and 30 minutes, in minutes.

[0071] Imaging delay time:

[0072] in, This indicates the image delay time, in seconds. This indicates the time point at which the contrast agent first appears in the lymphatic vessels, expressed in seconds. This indicates the time point at which the contrast agent injection was completed, in seconds.

[0073] Reflux duration:

[0074] in, Indicates the duration of the backflow, in seconds; Indicates the time point at which the backflow phenomenon disappears, in seconds; This indicates the time point at which the backflow phenomenon first occurs, in seconds.

[0075] Lymphatic vessel course continuity index:

[0076] CI stands for Lymphatic vessel continuity index, expressed as a percentage. Indicates the number of consecutive segments of lymphatic vessels; This indicates the total number of segments of the visualized lymphatic vessels.

[0077] Percentage of area in the reflux zone:

[0078] Where RA represents the area percentage of the reflux zone; This represents the pixel area of ​​the reflux region in the image. This represents the pixel area of ​​the affected limb in the image.

[0079] The ICG lymphography staging unit is integrated into the lymphatic function quantification unit, and the staging logic is as follows:

[0080] Installments ICG Reflow Mode ICG distribution characteristics 0th period normal linear Linear imaging of lymphatic vessels; lymphatic fluid flow can be seen in individuals with a slender build. Phase 1 abnormal linearity Abnormal or tortuous linear lymphatic vessels may be visualized. Phase 2 Splash (+) Splash (++) Linear lymphatic vessels are visible; lymph splashing and overflow are visible in one site; lymph splashing and overflow are visible in two or more sites. Phase 3 Stardust (+) Stardust (++) No linear lymphatic vessels were observed; stardust-like visualization was observed in both the upper arm and forearm. Phase 4 Pervasive (+) Pervasive (++) Pervasive (+++) Pervasive (++++) Diffuse reflux in single sites of the upper arm, forearm, and back of the hand; diffuse reflux in two sites of the upper arm, forearm, and back of the hand; diffuse reflux throughout the entire affected limb; diffuse reflux throughout the entire affected limb accompanied by subepidermal lymphatic vessel neovascularization.

[0081] Acquired images are packaged in DICOM standard format with a storage resolution of no less than 1024×1024 pixels and a bit depth of 16 bits. A double-blind quality control process is performed before images are entered into the database: two attending radiologists independently review the image quality, removing images with motion blur, fluorescence saturation, and injection extravasation based on preset artifact exclusion criteria. Only when the two physicians' reviews are consistent and the image usability rating is A can the image sequence proceed to the subsequent algorithmic process.

[0082] Example 3

[0083] like Figure 2 As shown, the upper limb circumference volume measurement module collects segmental circumference data of the affected and unaffected upper limbs at 8-centimeter intervals using an electronic circumference ruler or a quick upper limb lymphedema circumference measurement and assessment ruler. The module has a built-in formula for the volume of a frustum of a cone:

[0084]

[0085] Where V represents the volume of the frustum, in cubic centimeters; h represents the distance between two adjacent measurement points, in centimeters; This indicates the circumference of the upper limb at the upper measuring point among two adjacent measuring points, in centimeters. The value represents the circumference of the upper limb at the lower measuring point between two adjacent measuring points, in centimeters; 0.0265 is the reciprocal correction factor of pi, calculated from 1 / (12π), and is dimensionless.

[0086] The module calculates the volume difference between corresponding segments on the affected and healthy sides, summing them to generate the total volume swelling rate and the unilateral volume reduction rate. The following is an example of upper limb measurement:

[0087] Affected wrist joint segment: = 23.7 cm, = 16.1 cm, h = 8 cm

[0088] =0.0265×8×(23.7²+23.7×16.1+16.1²)

[0089] =0.0265×8×(561.69+381.57+259.21)

[0090] =0.0265 × 8 × 1202.47 = 254.92 cubic centimeters

[0091] healthy wrist joint segment: = 16.1 cm, = 16.1 cm, h = 8 cm

[0092] =0.0265×8×(16.1²+16.1×16.1+16.1²)

[0093] =0.0265×8×(16.1²+16.1×16.1+16.1²)

[0094] =0.0265 × 8 × 777.63 = 164.86 cubic centimeters

[0095] Swelling rate of this segment:

[0096] Where ηᵥ represents the volume swelling rate, in percentage; The volume of a frustum of a cone in a measured segment of the affected upper limb is expressed in cubic centimeters. The volume of the frustum of a cone corresponding to the measurement segment of the unaffected upper limb is expressed in cubic centimeters. This indicates the volume difference between the affected and healthy sides, expressed in cubic centimeters.

[0097]

[0098] The module applies the above calculations to all measurement segments on both the affected and healthy sides, outputting the volume difference and swelling rate segment by segment, and then summing them to generate the total swelling rate. For the upper limb, there are 7 measurement segments from the palm circumference to the shoulder joint. The module calculates the volume of each segment separately and then sums them to generate the total volume of the affected side and the total volume of the healthy side, finally outputting the total swelling rate.

[0099] This module transforms abstract upper limb swelling into segmentally quantifiable volume data, enabling precise location of edema sites, guiding key treatment areas, and providing quantitative indicators of severity for subsequent stratified assessments.

[0100] Example 4

[0101] In this embodiment, the TCM multimodal data acquisition module includes a tongue image acquisition unit, a pulse image acquisition unit, and a syndrome differentiation unit.

[0102] The tongue image acquisition unit is equipped with a standard D65 light source, with a fixed shooting distance of 15 cm, outputting a tongue image with a resolution of 3000×2000 pixels. The backend performs the following image processing steps using the OpenCV image processing library: first, white balance correction is performed to eliminate color cast from the ambient light source; second, tongue segmentation is performed, using the GrabCut algorithm to extract the tongue region from the background; finally, the RGB color values ​​of the tongue body, the thickness of the tongue coating, and the quantitative data of the tongue coating coverage area are extracted. The RGB color values ​​of the tongue body are obtained by calculating the average pixel value of the tongue region, the tongue coating thickness is obtained by analyzing the gray-level gradient distribution of the tongue coating region, and the tongue coating coverage area is obtained by statistically analyzing the proportion of tongue coating pixels to the total tongue area.

[0103] The pulse acquisition unit recorded a 30-second pulse signal at a sampling frequency of 500 Hz, with a total of 15,000 sampling points. After the signal was filtered by a bandpass filter (0.5 to 50 Hz) to remove baseline drift and high-frequency noise, the pulse frequency, main wave amplitude, and dicrotic depth features were extracted using Fast Fourier Transform. The pulse frequency was obtained by calculating the reciprocal of the main wave peak interval time, the main wave amplitude was the difference between the maximum and minimum values ​​of the pulse signal, and the dicrotic depth was the ratio of the secondary peak amplitude to the main peak amplitude.

[0104] The syndrome differentiation unit outputs TCM syndrome differentiation labels based on the "Classification and Determination of Traditional Chinese Medicine Constitution" standards and clinical diagnosis. The specific process is as follows: Three associate chief physicians independently review the patient's tongue image, pulse signal data, and clinical consultation information, and each provides a syndrome differentiation opinion. The system compares the differentiation results of the three physicians for consistency; only when the Kappa consistency coefficient is not less than 0.8 can the patient's syndrome differentiation label be entered into the database. The differentiation labels include Liver Qi Stagnation and Spleen Deficiency Type, Damp-Heat Accumulation Type, Qi Deficiency and Blood Stasis Type, Spleen Deficiency and Dampness Accumulation Type, and Spleen and Kidney Deficiency Type. Simultaneously, the unit generates quantitative scores for six syndrome elements: Qi stagnation, blood stasis, dampness, Qi deficiency, spleen deficiency, and kidney deficiency. Each syndrome element score ranges from 0 to 10 points, assessed by the physicians based on the severity of symptoms.

[0105] The generation method of syndrome element time-series scoring data: At four time points—patient enrollment (baseline), 1 month after treatment, 3 months after treatment, and 6 months after treatment—the same group of TCM physicians conducted syndrome differentiation and scoring, generating a time-series scoring sequence. , , , ], and calculate the rate of change of the evidence element:

[0106]

[0107] Wherein, ΔS represents the rate of change of the evidence element, which is dimensionless; The syndrome score indicates the syndrome factor score 3 months after treatment; This indicates the baseline evidence score.

[0108] Example 5

[0109] The Western medicine clinical data acquisition module is used to collect basic patient information, surgical and treatment-related information, auxiliary treatment information, and time-series scoring data of the Patient Reported Outcome Scale (PTS). Among them, the time-series scoring data of the PTS includes the Breast Cancer-Related Lymphedema Symptom Index Scale, the Arm-Shoulder-Hand Dysfunction Scale, the Upper Limb Lymphedema Quality of Life Scale, the Hospital Anxiety and Depression Scale, and the Herth Hope Scale.

[0110] The patient-reported outcome scale time-series score data were collected at four time points: when the patient was enrolled, 1 month after treatment, 3 months after treatment, and 6 months after treatment. The time-series score sequence was generated and the scale change rate was calculated. The data were then input into the traditional Chinese and Western medicine time-series data fusion module and the BCRL diagnosis and treatment decision module.

[0111] The Western medicine clinical data acquisition module uses the hospital information system interface and electronic data acquisition system to structurally capture patients' comprehensive Western medicine clinical data, including four sub-modules:

[0112] (1) Basic information unit: Collect demographic data (age, gender, height, weight, calculate BMI, dominant hand, occupation and whether upper limb weight is required), basic medical history (diabetes and its course, blood sugar control, hypertension, connective tissue disease, history of shoulder disease) and lifestyle habits (smoking history, frequency of physical exercise).

[0113] (2) Surgery and treatment related information unit: The surgical procedure (breast-conserving surgery, simple mastectomy, modified radical mastectomy, etc.), axillary lymph node treatment (number of sentinel lymph node biopsies, axillary lymph node dissection range grade I / II), incision details (incision type such as radial / transverse, location such as upper quadrant of the breast / lower quadrant / axillary, length), operation time, intraoperative blood loss, whether a drainage tube is placed and its indwelling time are mapped through the surgical operation coding library.

[0114] (3) Adjunctive therapy information unit: The radiotherapy regimen (whether it includes supraclavicular / axillary field, total dose, irradiation range), chemotherapy regimen (whether it includes taxanes / anthracyclines, number of cycles), endocrine therapy (specific drugs such as tamoxifen and aromatase inhibitors) and targeted therapy (drug type) are retrieved through the chemotherapy regimen coding library and radiotherapy structured report.

[0115] (4) Patient-Reported Outcome (PRO) Scale Collection Unit: Deploy an electronic data collection system to input and collect the following scales online and in a time-series manner:

[0116] The Breast Cancer-Associated Lymphedema Symptom Index (BCLE-SEI) contains 24 items, covering symptoms such as limited shoulder / elbow / wrist / finger / arm movement, swelling, heaviness, stiffness, tightness, thickened skin, inflexibility, numbness, tenderness, pain / pain / ache, redness, increased skin temperature, blistering, burning pain, tingling, needle-like sensation, weakness in the affected limb, and fatigue in the affected limb. Each item is scored on a five-point scale from 0 (not present) to 4 (very severe), with a total score of 0–96. Higher scores indicate more severe symptoms.

[0117] The Disability of the Arm, Shoulder and Hand Scale (DASH) contains 30 indicators, divided into limitations in daily living activities (23 items) and upper limb discomfort symptoms and sleep disturbances (7 items), with an additional section (4 items) for musicians / athletes. Higher scores indicate more severe functional impairment.

[0118] The Upper Limb Lymphedema Quality of Life Scale (ULLQoL) includes 14 items across two dimensions: physical and psychological status. The Chinese version has a Cronbach's α coefficient of 0.930 and is used to assess the specific factors influencing the quality of life of BCRL patients.

[0119] The Hospital Anxiety and Depression Scale (HADS) includes an anxiety subscale (HADS-A, 7 items) and a depression subscale (HADS-D, 7 items). Each item is scored on a four-level scale (0-3 points) based on the frequency of symptoms in the past month. It is used for rapid screening of anxiety and depression.

[0120] Herth Hope Scale (HHI): It consists of 12 items in 3 dimensions: positive attitudes, positive actions, and close relationships. It uses a Likert 4-point rating scale (1-4 points) with a total score of 12-48 points. The higher the score, the higher the level of hope.

[0121] The above scales were collected at four time points: patient enrollment (baseline), 1 month, 3 months, and 6 months after treatment, generating a time-series scoring sequence. And calculate the rate of change of the scale:

[0122]

[0123] in, This represents the rate of change in the PRO scale, and is dimensionless. The scale score indicates the results 3 months after treatment. This indicates the baseline scale score.

[0124] Once all scale data has been verified, it is automatically associated with anonymized patient IDs and encrypted and stored on the hospital server, accessible only to authorized researchers. A full backup is performed weekly.

[0125] Example 6

[0126] The edema nature stratification determination module executes the following determination logic:

[0127] Determination of fluid edema: When the patient complains of milder symptoms in the morning and more severe symptoms in the evening, physical examination reveals obvious pitting with a pressure depth greater than or equal to 6 mm, MRI shows subcutaneous soft tissue swelling and widening of the interstitial spaces, and color Doppler ultrasound shows homogeneous subcutaneous tissue echoes, the fluid edema stratification result is output.

[0128] Determination of solid edema: When physical examination reveals non-pitting edema with a pressure depth of less than 6 mm, or accompanied by acanthosis, verrucous hyperplasia, hyperkeratosis, MRI showing fat deposition and fibrosis with high signal on T1-weighted images and low signal on T2-weighted images, and color Doppler ultrasound showing thickening of the skin and fat layers with enhanced echogenicity, the solid edema stratification result is output.

[0129] Specifically, the edema nature stratification determination module performs continuous state determination based on the coupling of multiple physical quantities. The specific implementation of the liquid edema determination logic is as follows:

[0130] The first step is to receive the patient's chief complaint text, extract the keywords "worse in the evening and milder in the morning" through natural language processing, and assign weights to successful matches. = 0.3;

[0131] The second step involves receiving the physical examination data and determining the indentation depth. Specifically, the physician uses a standard pressure compressor to apply 30 Newtons of pressure to the edematous area for 5 seconds, then measures the rebound time after release. If the indentation depth is greater than or equal to 6 mm and the rebound time is less than or equal to 10 seconds, a weight is assigned. = 0.3;

[0132] The third step involves receiving MRI data and extracting the signal intensity of T2-weighted images of subcutaneous soft tissue. If the signal intensity is 20% higher than the corresponding area on the healthy side and widening of the interstitial space is visible, weighting is assigned. = 0.2;

[0133] The fourth step is to receive color Doppler ultrasound data and determine the homogeneity of subcutaneous tissue echoes. If the echoes are homogeneous and there are no fibrotic hyperechoic bands, a weighting is assigned. = 0.2;

[0134] Step 5: Weighted summation determination:

[0135]

[0136] in, The score represents the determination of liquid edema and is dimensionless. Indicates the weight of the main complaint text; Indicates the weight of the physical examination data; Indicates the weight of MRI data; This indicates the weight of the color Doppler ultrasound data.

[0137] like If the value is greater than or equal to 0.8, output the liquid edema determination result.

[0138] Specific implementation of the solid edema determination logic:

[0139] The first step involves assessing the patient's physical condition by applying pressure to a depth of less than 6 mm or a rebound time of more than 10 seconds, and assigning weights accordingly. = 0.25;

[0140] The second step involves assigning weights to cases with acanthosis, verrucous hyperplasia, or hyperkeratosis (identified via dermoscopy images). =0.25;

[0141] The third step, MRI, showed fat deposition (high signal on T1-weighted images) and fibrosis (low signal on T2-weighted images), with weighting assigned. = 0.25;

[0142] Step four, color Doppler ultrasound showed thickening of the skin and fat layers (thickness greater than 50% of the healthy side), with enhanced echogenicity accompanied by fibrotic, cord-like hyperechoic areas, weighted and assigned values. = 0.25;

[0143] Step 5: Weighted summation determination:

[0144]

[0145] in, The score represents the determination of solid edema and is dimensionless. to The meaning is the same as above.

[0146] like If the value is greater than or equal to 0.8, output the solid edema determination result.

[0147] The stratified results are output to the management suggestion information generation unit, serving as a rigid constraint for the selection of treatment plans.

[0148] Example 7

[0149] The TCM-Western medicine time-series data fusion module is used to integrate the time-series features of ICG dynamic lymphangiography, upper limb circumference and volume data, TCM syndrome element time-series score data and Western medicine clinical data. This module uses an attention mechanism that embeds a time-series correlation matrix to perform time-series feature-level fusion.

[0150] The traditional Chinese and Western medicine time-series data fusion module adopts a multi-stream Transformer architecture to achieve deep alignment of heterogeneous data, including:

[0151] (1) Image stream: Using Swin Transformer or ResNet-50 as the backbone network, input the three-phase key frames of ICG dynamic lymphangiography (15 minutes, 30 minutes, 60 minutes), extract the spatiotemporal features such as lymphatic vessel filling length, reflux area, and imaging stage, and output a 256-dimensional ICG image temporal feature vector.

[0152] (2) Structured / Text Flow (Tabular Transformer): The FT-Transformer is used to process the structured data of Western medicine clinical practice (basic information, surgical information, auxiliary treatment information) and PRO scale time series score data. Interactive embeddings are learned for categorical features, and bucketing and linear projection are performed on numerical features to output a 128-dimensional structured time series feature vector.

[0153] (3) TCM syndrome flow: The temporal scores and change rates of the six syndrome elements of Qi stagnation, blood stasis, dampness, Qi deficiency, spleen deficiency and kidney deficiency are encoded into a 128-dimensional TCM syndrome element temporal feature vector.

[0154] (4) Cross-Attention multimodal fusion: The above three feature vectors are mapped to the same hidden layer dimension through a fully connected layer. =512 dimensions, generating a query matrix based on the temporal features of TCM syndrome elements. Key matrix generated from ICG image temporal features Value matrix generated using structured Western medicine and PRO scale data. Where m is the number of time-series nodes (e.g., baseline, January, March, and June, a total of 4 nodes). A scaled dot product attention mechanism is employed.

[0155]

[0156] in, This is the attention score matrix between time-series nodes; The elements of the pre-trained temporal correlation bias matrix are... This represents the cross-modal correlation strength between the change in TCM syndrome element scores at the i-th time-series node and the change in ICG dynamic features at the j-th time-series node; this bias matrix is ​​derived from the global temporal correlation matrix. The dimension is obtained through modal subspace projection and dimensional mapping. Specifically, Pearson correlation analysis is used to calculate the correlation coefficient between the rate of change of TCM syndrome elements and the ICG time-series feature parameters from historical medical data. The Apriori association rule mining algorithm is then used to extract frequent itemsets (minimum support 10%, minimum confidence 60%), and statistically significant association rules are encoded as follows: The corresponding element; For dimensional scaling factor, This mechanism enables the model to automatically learn cross-modal associations such as "when ICG shows prolonged duration of lymphatic reflux, the time-series weight of blood stasis syndrome elements automatically increases," simulating the comprehensive judgment process of clinicians.

[0157] (5) Prior embedding of temporal correlation matrix: global temporal correlation matrix The (512×512 dimensional) constraint is loaded into memory during system initialization as a prior constraint at the hidden layer level. It is then mapped to a bias matrix matching the current batch time series length m through a linear projection layer. This ensures dynamic adaptation under different follow-up node numbers.

[0158] The traditional Chinese and Western medicine time-series data fusion module includes:

[0159] The temporal correlation matrix construction unit is used to extract multiple historical diagnosis and treatment data from the encrypted medical record database. Pearson correlation analysis is used to calculate the correlation coefficient between the time-series scores of TCM syndrome elements and the dynamic feature parameters of ICG. The prior probability association rule mining algorithm is used to extract frequent itemsets of the temporal changes of syndrome elements and the temporal changes of ICG. The statistically significant association rules are encoded into a temporal correlation matrix.

[0160] The attention-series fusion unit is used to map the ICG image time-series feature vector, the TCM syndrome element time-series feature vector, the patient-reported outcome scale time-series feature vector, and the Western medicine pathology feature vector to the same hidden layer dimension. It uses a scaled dot product attention mechanism to calculate the time-series attention weights, superimposes the time-series correlation matrix as a bias term, and outputs the fused time-series feature vector.

[0161] Specific implementation of the temporal correlation matrix construction unit:

[0162] Step 1, Data Preparation: Extract time-series clinical data from 500 patients from the encrypted database, including:

[0163] ICG timing characteristics: 15-minute fill length 30-minute filling length 60-minute filling length , filling speed v, reflux duration T_reflux, imaging delay time T_delay;

[0164] TCM syndrome element time-series scoring: baseline score S_pre, score 1 month after treatment S_1m, score 3 months after treatment S_3m, syndrome element change rate ΔS.

[0165] The second step is Pearson time-series correlation analysis: calculating the Pearson time-series correlation coefficient between the rate of change of each TCM syndrome element and each ICG time-series characteristic parameter.

[0166]

[0167] in, represents the time-series Pearson correlation coefficient, with a value range of [-1, 1], and is dimensionless; n represents the total number of samples, and n is greater than or equal to 500; Let i represent the rate of change of syndrome elements in the i-th patient, where i belongs to [1, n]. Let i represent the ICG time-series characteristic parameters of the i-th patient, where i belongs to [1, n]; The sample mean representing the rate of change of the evidence element, i.e. ; The sample mean of the ICG time series feature parameters is represented by... = ; This represents the deviation of the rate of change of the syndrome elements of the i-th patient from the mean; Σ represents the deviation of the ICG time series characteristics from the mean of the i-th patient; Σ represents the summation operation, i is accumulated from 1 to n; √ represents the square root operation.

[0168] When the absolute value of the correlation coefficient is greater than 0.3 and the p-value is less than 0.05, it is considered that the temporal variation of the evidence element is statistically significantly associated with the temporal characteristics of ICG.

[0169] The third step is to mine Apriori time-series association rules: the direction of change of TCM syndrome elements (increase / decrease / stable) and the direction of change of ICG phase (acceleration / delay / stable) are used as transaction itemsets, and the minimum support is set to 10% and the minimum confidence is set to 60% to mine frequent itemsets.

[0170] For example, if the support for the simultaneous occurrence of "elevated blood stasis syndrome elements" and "prolonged reflux duration" is 15%, and the confidence level for "elevated blood stasis syndrome elements" under the condition of "prolonged reflux duration" is 65%, then the rule is encoded as the temporal correlation matrix element M_blood stasis, reflux = 0.65.

[0171] The fourth step is matrix encoding: statistically significant association rules are encoded into a 512×512-dimensional temporal association matrix M_temporal. The matrix element Mᵢⱼ represents the association strength between the i-th TCM syndrome element temporal feature and the j-th ICG temporal feature, with a value ranging from -1 to 1. The matrix is ​​loaded into memory during system initialization and serves as a bias term for the attention mechanism.

[0172] Specific implementation of the attention-sequential fusion unit:

[0173] The first step, feature mapping: The ICG image temporal feature vector (256 dimensions, containing...) is mapped... , , , v, T_reflux, T_delay and their temporal difference features), TCM syndrome element temporal feature vector (128 dimensions, including ... , , The ΔS and correlation features between syndrome elements) and the Western medicine pathological feature vector (128 dimensions, including pathological molecular subtyping, surgical method, number of lymph node dissections, etc.) are mapped to the same hidden layer dimension of 512 through a fully connected layer to generate the Query matrix Q, Key matrix K and Value matrix V.

[0174] The second step is to calculate the attention score: The temporal attention weights are calculated using the scaled dot product formula.

[0175]

[0176] Where Q represents the Query matrix, which is obtained by mapping the temporal feature vectors of ICG images, and has a dimension of n×d. k K represents the Key matrix, which is obtained by mapping the time-series feature vectors of TCM syndrome elements, and has a dimension of n×d. k V represents the Value matrix, which is obtained by mapping the feature vectors of Western medicine pathology, and has a dimension of n×dᵥ; QKᵀ represents the matrix product of the Query matrix and the transpose of the Key matrix, and has a dimension of n×n; This represents the dimension of the key vector. = 512, dimensionless; This indicates scaling of the dimension to prevent the dot product result from becoming too large and causing the softmax gradient to vanish; M_temporal represents the pre-trained temporal correlation matrix with dimension M. × The matrix element Mᵢⱼ represents the correlation strength between the i-th TCM syndrome element temporal feature and the j-th ICG temporal feature; Softmax represents the normalized exponential function, which converts the attention score into a probability distribution; n represents the number of samples in the batch.

[0177] The third step is weight allocation: the attention output is a 512-dimensional fused temporal feature vector, in which the weights of ICG image temporal features, TCM syndrome element temporal features, and Western medicine pathological features are adaptively allocated by the attention mechanism. For example, when the duration of ICG reflux in the input patient is significantly prolonged and the rate of change of blood stasis syndrome elements is increased, the attention mechanism automatically increases the weight coefficients of these two temporal features.

[0178] The system adopts a multi-stream Transformer fusion architecture: (1) Image stream: ResNet50 or SwinTransformer is used as the backbone, and ImageNet or general medical image dataset pre-trained weights are loaded to extract ICG image temporal feature embedding; (2) Text / structured stream (Tabular Transformer): TCM syndrome element temporal scores and Western medicine structured clinical data are processed to generate structured feature vectors; (3) Cross-Attention fusion layer: The model automatically learns cross-modal associations through the cross-attention mechanism. For example, when ICG shows severe lymph reflux, the association weight of blood stasis syndrome elements is automatically increased to simulate the doctor's comprehensive judgment process.

[0179] When the sample size is insufficient, the image stream is loaded with encoder weights pre-trained on a general medical imaging dataset and fine-tuned on the ICG image data. At the same time, a GAN network is used to generate simulated ICG images and corresponding symptom labels for data augmentation to alleviate the risk of overfitting under small sample size and improve the model's predictive generalization ability on new patients.

[0180] Example 8

[0181] The edema management decision-making module includes an incubation period early warning unit, a risk early warning information output unit, a management suggestion information generation unit, and a prognostic risk prediction information output unit.

[0182] The specific early warning units for the incubation period are as follows:

[0183] Existing lymphedema diagnostic systems all rely on thresholds for significant structural abnormalities or volume differences. This unit implements the following early warning logic for patients with early-stage, occult BCRL:

[0184] The first step is to read the difference in upper limb circumference and volume output from the upper limb circumference and volume measurement module. When the difference in upper limb circumference is less than 2 cm and the difference in volume is less than 10%, the assessment branch for the occultation period begins.

[0185] The second step is to read the timing parameters of the lymphatic drainage function output by the ICG dynamic lymphangiography acquisition module. When the imaging delay time T_delay exceeds the normal average by more than 30%, or there is splash / stardust-like imaging (stage greater than or equal to 2), or the reflux duration T_reflux exceeds the normal average by more than 20%, the lymphatic drainage function is determined to be abnormal.

[0186] The third step is a comprehensive assessment: When the volume indicators from the first step are normal but the functional indicators from the second step are abnormal, a subclinical BCRL warning is issued, with the following text: "The patient's current upper limb swelling indicators have not reached the clinical diagnostic threshold, but ICG lymphography shows abnormal lymphatic drainage function. It is recommended to initiate enhanced follow-up monitoring, with ICG lymphography and circumferential volume measurement repeated every 4 weeks."

[0187] This early warning mechanism allows for intervention 6 to 12 months earlier, preventing patients from progressing to moderate to severe lymphedema.

[0188] The specific risk warning information output unit is as follows:

[0189] The risk warning information output unit adopts a hybrid architecture of convolutional neural network and XGBoost. The convolutional neural network uses ResNet50 as the backbone network. After loading ImageNet pre-trained weights, the original fully connected layers are removed and replaced with a classification head optimized for ICG lymphangiography images. The classification head structure is as follows: global average pooling layer (output 2048 dimensions) → Dropout layer (dropout rate 0.5) → fully connected layer (output 256-dimensional image deep feature vector).

[0190] The XGBoost part receives the concatenation result of the 256-dimensional image feature vector and the 18-dimensional non-image feature vector. The 18-dimensional non-image features include: number of lymph nodes dissected (1 dimension), TCM syndrome score (6 dimensions), BMI (1 dimension), surgical procedure code (3 dimensions: breast-conserving surgery / modified radical mastectomy / radical mastectomy), pathological molecular subtype code (4 dimensions: Luminal A / Luminal B / HER-2 positive / triple negative), age (1 dimension), postoperative time (1 dimension), and frequency of functional exercises (1 dimension).

[0191] The XGBoost parameters were configured as follows: learning rate 0.05, maximum tree depth 6, subsampling rate 0.8, column sampling rate 0.8, 500 iterations, and a binary logistic regression objective function. The model training employed a five-fold cross-validation strategy, dividing the retrospective dataset into training, validation, and test sets in a 7:1:2 ratio. In each fold of training, the training set was used to update the ResNet50 and XGBoost parameters, the validation set was used for early stopping (stopping if there was no improvement in AUC after 20 consecutive validation rounds), and the test set was used for final performance evaluation.

[0192] The unit outputs the BCRL diagnostic probability value and the staging recommendations from the International Society of Lymphoma. A diagnostic probability greater than or equal to 0.5 is considered positive, and the probability range is mapped to the staging recommendations: 0.5 to 0.6 is mapped to stage I, 0.6 to 0.75 to stage II, 0.75 to 0.9 to stage III, and greater than 0.9 to the advanced stage.

[0193] The management suggestion information generation unit is as follows:

[0194] The management suggestion information generation unit performs differentiated recommendations based on the stratification results of the edema nature stratification module:

[0195] When the stratified assessment unit determines the condition to be fluid edema, conservative treatment is the preferred option. Conservative treatment includes: Western medicine physical therapy sub-actions (manual lymphatic drainage, pressure bandage therapy, rehabilitation training program) and traditional Chinese medicine external therapy sub-actions (cupping along meridians, external application of herbal hot compresses, acupuncture). The system selects the optimal combination from the action space; for example, for patients with fluid edema due to spleen deficiency and dampness retention, the recommended combination is "manual lymphatic drainage + pressure bandage + spleen-strengthening and dampness-eliminating decoction + cupping along meridians".

[0196] When the stratified assessment unit determines the edema to be solid, a surgical intervention combined with comprehensive postoperative anti-edema treatment is recommended. Surgical intervention sub-actions include lymphangiovenous anastomosis, vascularized lymph node transplantation, and liposuction. The system recommends specific surgical procedures based on ICG staging and swelling rate: lymphangiovenous anastomosis is recommended for stages 2 and below with a swelling rate less than 100%; liposuction combined with vascularized lymph node transplantation is recommended for stages 3 and above with a swelling rate greater than 100%. Postoperative comprehensive anti-edema treatment includes pressure bandages, rehabilitation training, and traditional Chinese medicine external therapies.

[0197] The management suggestion information generation unit adopts a reinforcement learning framework based on deep Q-networks. The state space is a fusion of temporal feature vectors (512 dimensions); the action space includes legal permutations and combinations of Western medicine physical therapy sub-actions (2 types), Western medicine surgical treatment sub-actions (3 types), traditional Chinese medicine external treatment sub-actions (3 types), and traditional Chinese medicine internal treatment sub-actions (2 types), for a total of 20 combinations (excluding unreasonable combinations of surgery + surgery).

[0198] The reward function is defined as a weighted sum of three metrics:

[0199]

[0200] Where R represents the immediate reward value of reinforcement learning, which ranges from [0, 10] and is dimensionless; ΔQ represents the edema reduction rate, which is the ratio of the difference in total swelling rate before and after treatment to the baseline swelling rate, expressed as a percentage; ΔQ represents the improvement in quality of life, which is the difference between the total score of the Upper Limb Lymphedema Quality of Life Scale (ULLQoL) or the Functional Assessment of Breast Cancer Treatment (FACT-B) after treatment and the total score before treatment, expressed as a score. The treatment safety score is represented by 10 points for no adverse events, 7 points for minor adverse events, and 0 points for serious adverse events. It is dimensionless. The weighting coefficient representing the edema reduction rate. =0.4; The weighting coefficients represent the degree of improvement in quality of life. = 0.3; The weighting coefficients representing the treatment safety score. = 0.3; + + = 1.0, to ensure that the reward value is normalized.

[0201] The network structure employs two fully connected hidden layers, each with 256 neurons, using ReLU activation. The output layer dimension equals the action space size (20-dimensional). During training, the experience replay pool is set to 10,000 transition samples, and an ε-greedy strategy is used for exploration. The exploration probability decay formula is:

[0202]

[0203] in, Let represent the exploration probability in the t-th training round, with a value range of [0, 1], and be dimensionless; Indicates the initial exploration probability. = 0.3; δ represents the decay rate, δ = 0.05, meaning a 5% decay per cycle; t represents the current training round, t belongs to [0, ], The maximum number of training epochs; T represents the decay period, T = 1000 epochs; ⌊·⌋ represents the floor function, i.e., calculating the integer part of t / T; when When it is less than 0.01, take a fixed value. = 0.01, to ensure that the minimum exploration capability is maintained in the later stages of training.

[0204] The target network updates the weights of the main network every 500 rounds. During the model inference phase, the top 3 actions with the highest Q-values ​​are output.

[0205] The contraindication screening subunit is as follows:

[0206] Establish a contraindication knowledge graph: History of allergy to Chinese medicine → filter out internal / external Chinese medicine treatment plans containing the corresponding Chinese medicine; Coagulation dysfunction (International Normalized Ratio (INR) greater than 1.5) → filter out all Western medicine surgical treatments; Severe heart failure (New York Heart Association functional class III or above) → filter out pressure bandage therapy.

[0207] When the patient's electronic medical record contains the above-mentioned contraindications, the system automatically filters the corresponding action plan and displays the unselectable plan and the reason for disabling it in gray on the interface.

[0208] The evidence-based generation of sub-units is as follows:

[0209] Case database retrieval uses temporal similarity rather than static similarity:

[0210] +

[0211] Where Similarity represents temporal similarity, with a value range of [0, 1], and is dimensionless; Represents the similarity weight of ICG time series features. = 0.4; This represents the weight of the temporal similarity of TCM syndrome elements. = 0.4; Indicates the weight of the decay factor during follow-up time. = 0.2; This represents the current patient's ICG time-series feature vector; ICG time-series feature vectors representing historical cases; This represents the time-series score vector of the TCM syndrome elements of the current patient; The vector represents the time-series score of TCM syndrome elements in historical cases; cosine represents the cosine similarity function. This indicates the difference in follow-up time between current patients and historical cases, expressed in months. This is a time-decrease factor; the smaller the follow-up time difference, the higher the weight.

[0212] After retrieving similar cases, the system extracts the treatment plan and efficacy data (edema reduction rate, improvement in the Breast Cancer Treatment Functional Assessment Scale, adverse event records) used in those cases, generating evidence-based text. For example, for a patient with spleen deficiency and dampness retention, stage 2, and a volume edema rate of 75%, the system outputs:

[0213] The first recommended regimen is "manual lymphatic drainage + pressure bandage + spleen-strengthening and dampness-eliminating decoction + meridian cupping". The evidence is based on the following: "In 50 similar cases in history, this regimen achieved a 78% edema reduction rate, improved the Breast Cancer Treatment Functional Assessment Scale (FACT-B) score by 12 points, and had no serious adverse events".

[0214] The second recommended treatment plan is: "Rehabilitation training program + pressure bandage + blood circulation promoting and stasis removal formula + traditional Chinese medicine hot compress external application". Evidence-based basis: "In 30 similar cases in history, this plan has a 65% edema reduction rate, a 10-point improvement on the breast cancer treatment functional assessment scale, and 1 case of mild skin allergy".

[0215] The third recommended treatment plan is "manual lymphatic drainage + rehabilitation training + spleen-strengthening and dampness-eliminating decoction + acupuncture". The evidence-based basis is: "In 20 similar cases in history, this plan achieved a 70% edema reduction rate, improved the breast cancer treatment functional assessment scale score by 11 points, and had no adverse events".

[0216] The specific output unit for prognostic risk prediction information is as follows:

[0217] The prognostic risk prediction information output unit adopts a hybrid architecture of Cox proportional hazards regression and long short-term memory network.

[0218] The Long Short-Term Memory (LSTM) network receives time-series data from patient follow-up assessment scales. The sequence length is fixed at 6 time points (baseline, 1 month, 2 months, 3 months, 4 months, and 6 months). Each time point's input dimensions include the FACT-B physiological dimension score, DASH score, and Barthel index, totaling 3 dimensions. The LSTM network's hidden layer has 128 units, and dynamic trend features are extracted after time step expansion. The network structure is: Input layer (3 dimensions) → LSTM layer 1 (128 units, returning the sequence) → LSTM layer 2 (128 units, returning the last time step) → Output layer (128-dimensional hidden state vector).

[0219] The Cox proportional hazards regression part uses the 128-dimensional hidden state vector output by the long short-term memory network as a time-varying covariate, combined with patient baseline covariates including age, BCRL stage, sum of TCM syndrome element scores, and pathological molecular subtyping codes, to fit a proportional hazards function:

[0220]

[0221] Where h(t|X) represents the instantaneous hazard rate at time t under the condition of covariate X, in units of 1 / time; h0(t) represents the baseline hazard function, i.e. the hazard rate when all covariates take the value of 0, in units of 1 / time; t represents the follow-up time, in units of months; X represents the covariate vector, which includes the output features of the long short-term memory network and baseline clinical indicators; The value represents the comprehensive temporal feature value of the 128-dimensional hidden state vector output by the Long Short-Term Memory Network after dimensionality reduction, and is dimensionless; A represents the patient's age in years; S represents the BCRL staging code, with values ​​of 0, 1, 2, and 3, corresponding to stages 0 to III of the International Lymphatic Society staging system, respectively. β1 represents the total score of TCM syndrome elements, which is the sum of the scores of six items: Qi stagnation, blood stasis, dampness, Qi deficiency, spleen deficiency, and kidney deficiency, with a value range of 0 to 60; M represents the pathological molecular subtype code, with values ​​of 1, 2, 3, and 4, corresponding to Luminal A, Luminal B, HER-2 positive, and triple negative types, respectively; β1 represents the regression coefficient of the long short-term memory network characteristics, which is solved by maximum partial likelihood estimation; β2 represents the regression coefficient of age; β3 represents the regression coefficient of BCRL stage; β4 represents the regression coefficient of the total score of TCM syndrome elements; β5 represents the regression coefficient of the pathological molecular subtype; exp represents the natural exponential function, which is the power of e, and e is approximately equal to 2.71828.

[0222] The unit outputs the 1-year recurrence probability and the 2-year recurrence probability. When the 1-year recurrence probability exceeds 30%, the system triggers the enhanced follow-up instruction of the follow-up management module, adjusting the follow-up frequency from every 3 months to every 1 month.

[0223] In this embodiment, for scenarios with a limited sample size of BCRL patients, the system adopts the following small sample optimization and interpretability enhancement strategies:

[0224] (1) Transfer learning pre-training: The image stream encoder is pre-trained on ImageNet or public medical image datasets (such as CheXpert, MedMNIST) to obtain general medical image feature extraction capabilities. Then, it is fine-tuned end-to-end on the ICG lymphography dataset, freezing the parameters of the bottom convolutional layers and only updating the top classification head to prevent overfitting.

[0225] (2) Data augmentation: Random rotation (±15°), scaling (0.9-1.1 times), horizontal flipping, elastic deformation and brightness perturbation are performed on the original ICG fluorescence images to expand the diversity of training samples; SMOTE or Tab-GAN is used to generate simulated patient records for structured data to balance the sample distribution of different stages and syndrome types.

[0226] (3) Interpretability module: SHAP (SHapley Additive exPlanations) value analysis is introduced to quantify the contribution of each modality feature (such as ICG reflux duration, blood stasis syndrome score, DASH score) to the probability of BCRL diagnosis, treatment plan recommendation and prognosis recurrence probability; Attention MAP visualization is generated simultaneously to highlight the lymphatic vessel area or reflux area that has the greatest impact on the current stage determination on ICG images, providing clinicians with transparent evidence of "why the model makes such a decision".

[0227] (4) Enhanced time series modeling: The LSTM network in the prognostic assessment unit treats the patient's treatment timeline as sequential data. The time steps include t0 (surgery), t1 (chemotherapy), t2 (edema), and t3 (traditional Chinese medicine intervention), predicting the degree of edema and the risk of complications at the next time point t4, and dynamically adjusting the follow-up frequency.

[0228] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A postoperative upper limb lymphedema management system based on ICG angiography multimodal fusion, characterized in that, include: The ICG dynamic lymphangiography acquisition module is used to acquire dynamic image sequences at three time points: 15 minutes, 30 minutes and 60 minutes after the injection of indocyanine green contrast agent using an infrared fluorescence imager, and to extract the timing parameters of lymphatic drainage function, including lymphatic filling speed, imaging delay time and reflux duration. The upper limb circumference volume measurement module is used to collect segmental circumference measurement data along the affected and healthy upper limbs at preset intervals, and calculate the edema volume of each segment of the upper limb and the total volume swelling rate based on the frustum volume formula. The TCM multimodal data acquisition module is used to collect quantitative data including the color values ​​of the three primary colors of tongue appearance (red, green, and blue), pulse frequency and amplitude, and generate TCM syndrome differentiation labels and time-series score data of six syndrome elements: Qi stagnation, blood stasis, dampness, Qi deficiency, spleen deficiency and kidney deficiency. The Western medicine clinical data acquisition module is used to collect basic patient information, surgical and treatment-related information, auxiliary treatment information, and time-series scoring data of the Patient Reported Outcome Scale (PTS). Among them, the time-series scoring data of the PTS includes the Breast Cancer-Related Lymphedema Symptom Index Scale, the Arm-Shoulder-Hand Dysfunction Scale, the Upper Limb Lymphedema Quality of Life Scale, the Hospital Anxiety and Depression Scale, and the Herth Hope Scale. The edema nature stratification determination module is used to determine whether the main pathological component of edema is liquid edema or solid edema based on the time sequence parameters of lymphatic drainage function output by the ICG dynamic lymphangiography acquisition module, the total volume swelling rate output by the upper limb circumference volume measurement module, and externally input physical examination / MRI / color Doppler ultrasound data. The TCM and Western medicine time-series data fusion module is used to fuse the time-series parameters of the lymphatic drainage function, the total volume swelling rate, the time-series score data of the syndrome elements, the time-series score data of the patient reported outcome scale, and the Western medicine clinical data at the time-series feature level to generate a fused time-series feature vector. The time-series feature-level fusion adopts an attention mechanism that embeds a time-series correlation matrix. The time-series correlation matrix is ​​pre-trained based on the time-series correlation of existing diagnosis and treatment data of TCM syndrome element scores and ICG dynamic feature parameters. The edema management decision module is used to output lymphedema risk warning information, multidisciplinary collaborative management suggestions combining traditional Chinese and Western medicine, and prognostic risk prediction information based on the fusion time-series feature vector and the stratification results of the edema nature stratification judgment module.

2. The postoperative upper limb lymphedema management system based on ICG angiography multimodal fusion according to claim 1, characterized in that: The ICG dynamic lymphangiography acquisition module includes: The contrast agent injection unit is used to inject indocyanine green water-based solution into the lymphatic-rich area of ​​the second finger web or the dorsum of the foot of the affected limb, at a dose of 0.1 to 0.3 ml. Near-infrared excitation imaging unit, used to continuously acquire dynamic fluorescence images at a rate of 30 frames per second in the 750 to 1000 nanometer wavelength band; The temporal frame extraction unit is used to extract static keyframes at 15 minutes, 30 minutes and 60 minutes after contrast agent injection from the dynamic fluorescence image, which correspond to the initial stage of lymphatic vessel filling, dynamic drainage period and delayed imaging period, respectively. The lymphatic function quantification unit is used to calculate the lymphatic vessel continuity index, imaging delay time, and reflux area ratio based on the static keyframe, and output the stage 0 to stage 4 results according to the ICG lymphography staging standard.

3. The postoperative upper limb lymphedema management system based on ICG angiography multimodal fusion according to claim 1, characterized in that: The upper limb circumference volume measuring module is spaced 8 cm apart along the upper limb. The formula for the volume of the frustum is: Where V represents the volume of the frustum, in cubic centimeters; h represents the distance between two adjacent measurement points, in centimeters; C1 represents the circumference of the upper limb of the upper measurement point between two adjacent measurement points, in centimeters; C2 represents the circumference of the upper limb of the lower measurement point between two adjacent measurement points, in centimeters; 0.0265 is the reciprocal correction factor of pi, calculated from 1 / (12π), and is dimensionless.

4. The postoperative upper limb lymphedema management system based on ICG angiography multimodal fusion according to claim 1, characterized in that: The edema nature stratification determination module executes the following determination logic: Determination of fluid edema: When the patient complains of milder symptoms in the morning and more severe symptoms in the evening, physical examination reveals obvious pitting with a pressure depth greater than or equal to 6 mm, MRI shows subcutaneous soft tissue swelling and widening of the interstitial spaces, and color Doppler ultrasound shows homogeneous subcutaneous tissue echoes, the fluid edema stratification result is output. Determination of solid edema: When physical examination reveals non-pitting edema with a pressure depth of less than 6 mm, or accompanied by acanthosis, verrucous hyperplasia, hyperkeratosis, MRI showing fat deposition and fibrosis with high signal on T1-weighted images and low signal on T2-weighted images, and color Doppler ultrasound showing thickening of the skin and fat layers with enhanced echogenicity, the solid edema stratification result is output.

5. The postoperative upper limb lymphedema management system based on ICG angiography multimodal fusion according to claim 1, characterized in that: The traditional Chinese and Western medicine time-series data fusion module includes: The temporal correlation matrix construction unit is used to extract multiple historical diagnosis and treatment data from the encrypted medical record database. Pearson correlation analysis is used to calculate the correlation coefficient between the time-series scores of TCM syndrome elements and the dynamic feature parameters of ICG. The prior probability association rule mining algorithm is used to extract frequent itemsets of the temporal changes of syndrome elements and the temporal changes of ICG. The statistically significant association rules are encoded into a temporal correlation matrix. The attention-series fusion unit is used to map the ICG image time-series feature vector, the TCM syndrome element time-series feature vector, the patient-reported outcome scale time-series feature vector, and the Western medicine pathology feature vector to the same hidden layer dimension. It uses a scaled dot product attention mechanism to calculate the time-series attention weights, superimposes the time-series correlation matrix as a bias term, and outputs the fused time-series feature vector.

6. The postoperative upper limb lymphedema management system based on ICG angiography multimodal fusion according to claim 1, characterized in that: The edema management decision module includes: The latent period early warning unit is used to output a subclinical BCRL early warning prompt when the patient's upper limb circumference difference is less than 2 cm and the volume difference is less than 10%, but the timing parameters of the lymphatic drainage function are abnormal. The risk warning information output unit is used to output the BCRL diagnosis probability and the staging recommendation of the International Society of Lymphoma based on the fused temporal feature vector. The management suggestion information generation unit is used to prioritize conservative treatment when the edema nature stratification judgment module outputs liquid edema, and recommend surgical intervention combined with postoperative comprehensive edema reduction treatment when it outputs solid edema. The treatment recommendation unit adopts a multi-objective optimization decision model based on reinforcement learning. The action space includes permutations and combinations of Western medicine physical therapy sub-actions, Western medicine surgical treatment sub-actions, traditional Chinese medicine external treatment sub-actions, and traditional Chinese medicine internal treatment sub-actions. The reward function is a weighted sum of edema regression rate, patient-reported outcome scale quality of life improvement, and treatment safety. The quality of life improvement is quantified by the difference in total score before and after treatment of the upper limb lymphedema quality of life scale or breast cancer treatment function assessment scale. The unit outputs the top three treatment options in descending order of decision confidence and their corresponding evidence-based text. The prognostic risk prediction information output unit is used to output the 1-year and 2-year recurrence probability prediction based on time-series follow-up data.

7. The postoperative upper limb lymphedema management system based on ICG angiography multimodal fusion according to claim 6, characterized in that: The management suggestion information generation unit includes: The contraindication screening subunit is used to automatically filter action plans containing corresponding Chinese medicine or surgery when there is a history of allergy to Chinese medicine or coagulation dysfunction in the patient's electronic medical record. The evidence-based evidence generation subunit is used to call the historical multidisciplinary collaborative case library of integrated traditional Chinese and Western medicine, retrieve the efficacy data of cases similar to the current patient's syndrome element time series score and ICG dynamic feature parameters, and generate the top three treatment plans and corresponding evidence-based evidence texts in descending order of decision confidence. The decision confidence is determined based on the decision quality value output by the reinforcement learning model.

8. The postoperative upper limb lymphedema management system based on ICG angiography multimodal fusion according to claim 6, characterized in that: The prognostic risk prediction information output unit adopts a hybrid architecture of Cox proportional hazards regression and long short-term memory network. The long short-term memory network receives time-series score data of the Breast Cancer Treatment Functional Assessment Scale, Arm, Shoulder and Hand Dysfunction Scale, Upper Limb Lymphedema Quality of Life Scale, Hospital Anxiety and Depression Scale, Herth Hope Scale and Daily Living Ability Index during the patient's follow-up period to capture dynamic trends. The Cox proportional hazards regression fits a risk proportional function with the dynamic trend characteristics and the patient's baseline covariates, and outputs the predicted values ​​of recurrence probability at 1 year and 2 years.

9. The postoperative upper limb lymphedema management system based on ICG angiography multimodal fusion according to claim 1, characterized in that: The system is deployed on a local hospital server. The ICG dynamic lymphangiography acquisition module, the upper limb circumference and volume measurement module, the traditional Chinese medicine multimodal data acquisition module, and the Western medicine clinical data acquisition module are connected to the traditional Chinese and Western medicine time-series data fusion module through a standard data interface. The edema management decision module outputs structured prediction results to the application layer through a representation state transition application programming interface. The application layer includes a data entry module, an image analysis module, a diagnostic suggestion module, a treatment recommendation module, a prognostic assessment module, and a follow-up management module.

Images