A physical intelligent agent physical examination palpation result judgment algorithm

Abstract

This invention discloses an embodied intelligent physical examination palpation result judgment algorithm, belonging to the field of medical robot technology. This invention is applied to an embodied intelligent agent comprising a tactile sensor array, a force sensor, and an end effector. The algorithm includes: acquiring the initial scanning path of the palpation area; controlling the end effector to perform initial screening palpation at each palpation point along the initial scanning path. The advantages are: this invention achieves intelligent palpation through a multi-layered design, including online estimation of tissue impedance and adaptive adjustment of palpation application parameters; autonomous encrypted rescanning of suspicious areas to accurately confirm abnormal boundaries; assignment of confidence labels based on the change in confidence before and after rescanning; and inputting structured palpation data into a medical large language model to generate a formatted report. This results in improved palpation accuracy, reduced the risk of missed or misdiagnosed diagnoses, and reduced the paperwork burden on doctors.

Description

Technical Field

[0001] This invention relates to the field of medical robot technology, and in particular to an algorithm for judging the results of embodied intelligent physical examination palpation. Background Technology

[0002] Physical examination and palpation are basic clinical skills. Doctors use their fingers to press and feel the hardness, elasticity, and boundaries of subcutaneous tissue to detect abnormalities such as lumps and nodules. This process is highly dependent on the doctor's experience and real-time adjustment of techniques. There are differences in judgment between different doctors, and it is difficult to implement in telemedicine scenarios.

[0003] With the development of medical robot technology, palpation robots have begun to be used to replace manual palpation in standardized palpation. Existing palpation robots usually use fixed force values ​​or displacement control to collect tactile signals point by point along a preset grid path and classify abnormalities through simple stiffness thresholds.

[0004] The above solution has the following shortcomings: With fixed parameters, it is impossible to adjust the pressure and speed according to the real-time response of the tissue, making it difficult to simulate the doctor's adaptive palpation technique; It only performs a single coverage scan and does not automatically encrypt and re-examine after discovering suspicious areas, lacking a detailed investigation process; The judgment result is output as a binary label of normal or abnormal, which cannot reflect the degree of certainty of the judgment itself.

[0005] Therefore, it is necessary to design an embodied intelligent physical examination palpation result judgment algorithm. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides an algorithm for judging the results of embodied intelligent physical examination palpation, which solves the problems mentioned in the background.

[0007] To achieve the above objectives, the present invention adopts the following technical solution: An algorithm for judging the palpation results of an embodied intelligent physical examination, applied to an embodied intelligent agent comprising a tactile sensor array, a force sensor, and an end effector, includes: The initial scanning path of the palpation area is obtained, and the end effector is controlled to perform initial screening palpation on each palpation point along the initial scanning path, while real-time acquisition of tactile and force signal data. Using a preset online tissue impedance estimation model, the equivalent tissue impedance parameters are estimated based on the tactile and force signal data of the current palpation point, and the palpation application parameters of the next palpation point are adaptively adjusted based on the equivalent tissue impedance parameters. Using a pre-defined abnormality probability prediction model, an abnormality probability heatmap of the palpation area is constructed based on the deviation between the tissue equivalent impedance parameters of each palpation point and the pre-defined normal tissue impedance benchmark value. When the anomaly score of any spatial unit in the anomaly probability heatmap exceeds the preset rescan trigger threshold, a rescan encrypted sampling path is automatically generated with that spatial unit as the center, and rescan palpation is performed on the rescan area to obtain encrypted tactile signal data and force signal data. Based on the encrypted data obtained from the rescan palpation, the abnormality judgment result of the abnormal area is obtained, and the boundary of the abnormal area is segmented and the quantitative descriptive parameters of the abnormal area are extracted. For each abnormal region that has been rescanned, the change in confidence level before and after rescanning is calculated. Based on the magnitude and direction of the change in confidence level, a preset confidence level label is assigned to the abnormal judgment result. The tissue equivalent impedance parameters of each palpation point, the quantitative description parameters of abnormal areas, and the confidence labels are transformed into a structured natural language description sequence, which is then input into a pre-trained medical multimodal large language model to generate a formatted physical examination palpation report.

[0008] Furthermore, the tissue equivalent impedance parameter includes at least one of tissue equivalent stiffness, tissue equivalent damping coefficient, and tissue equivalent thickness; the palpation application parameter includes at least one of applied force amplitude, pressing speed, dwell time, and kneading motion trajectory radius.

[0009] Furthermore, the adaptive adjustment of the palpation application parameters at the next palpation point based on the tissue equivalent impedance parameter specifically includes: When the tissue's equivalent stiffness is higher than the preset stiffness reference value, the applied force amplitude is reduced and the pressing speed is decreased; when the tissue's equivalent stiffness is lower than the preset stiffness reference value, the applied force amplitude is increased and the pressing speed is increased. When the tissue's equivalent damping coefficient is higher than the preset damping reference value, the residence time is increased; when the tissue's equivalent damping coefficient is lower than the preset damping reference value, the residence time is shortened.

[0010] Furthermore, the anomaly probability prediction model constructs an anomaly probability heatmap of the palpation area, specifically including: The tissue equivalent impedance parameters of each palpation point are compared with the preset normal tissue impedance benchmark value. Based on the comparison results, the deviations of each parameter are weighted and fused to generate an abnormal score for that palpation point. The abnormal scores of all palpation points are spatially interpolated and normalized to generate an abnormal probability heatmap covering the palpation area.

[0011] Furthermore, the autonomously generated rescan encrypted sampling path includes: Centered on the spatial unit that triggers rescanning, an encrypted sampling path with a higher sampling point density than the initial scanning path is generated within a preset radius, following at least one of the preset inward spiral path, outward radial path, or reciprocating broken line path along the main direction of the abnormal score gradient; the applied force resolution of the rescanning palpation is higher than that of the initial screening palpation, and the pressing speed is lower than that of the initial screening palpation.

[0012] Furthermore, the calculation of the change in judgment confidence before and after rescanning includes: The initial screening anomaly score of the spatial cell before rescanning is extracted from the anomaly probability heatmap as the initial screening confidence level; the encrypted anomaly score of the anomaly region is calculated from the rescanned encrypted data as the rescan confidence level; the difference between the rescan confidence level and the initial screening confidence level is used as the change in judgment confidence level.

[0013] Furthermore, the step of assigning a preset level of confidence label to the anomaly judgment result based on the magnitude and direction of the change in confidence level specifically includes: A high confidence label is assigned when the change in confidence level is positive and its absolute value exceeds the preset significance threshold; a medium confidence label is assigned when the change in confidence level is positive but its absolute value does not exceed the preset significance threshold, or when the change in confidence level is negative but its absolute value does not exceed the preset significance threshold; and a low confidence label is assigned when the change in confidence level is negative and its absolute value exceeds the preset significance threshold.

[0014] Furthermore, the quantitative descriptive parameters of the abnormal region include at least one of the following: the sequence of contour boundary points of the abnormal region, area, length in the major axis direction, length in the minor axis direction, mean of tissue equivalent stiffness, variance of tissue equivalent stiffness, and texture uniformity index.

[0015] Furthermore, the online tissue impedance estimation model employs a pre-trained deep learning model, using the pressure distribution matrix collected by the tactile sensor array and the three-dimensional force vector collected by the force sensor as input, and outputs the tissue equivalent stiffness, tissue equivalent damping coefficient, and tissue equivalent thickness of the current contact area.

[0016] Furthermore, all model inference tasks deployed on the embodied intelligent agent are executed on a local edge computing platform; the initial screening palpation stage only runs the tissue impedance online estimation model and the anomaly probability prediction model, the rescanning palpation stage dynamically activates the boundary segmentation model, and the inference computing resources of the initial screening palpation stage and the rescanning palpation stage are allocated on demand.

[0017] Compared with existing technologies, the advantages of this invention are: 1. By using an online tissue impedance estimation model to obtain tissue equivalent impedance parameters in real time, and by using an admittance control model to adaptively adjust the applied force amplitude, pressing speed and dwell time at the next palpation point, the palpation technique is dynamically adjusted according to tissue characteristics. This achieves the effect of simulating the palpation skills of doctors tailored to individuals and locations, and avoids the problem of excessive soft tissue deformation or omission of nodules caused by fixed parameters.

[0018] 2: By constructing a heat map through an anomaly probability prediction model and autonomously generating an encrypted sampling path to perform rescanning when the anomaly score exceeds the threshold, and simultaneously improving force resolution and reducing pressing speed, a layered focused palpation effect from coarse screening to fine inspection is achieved. This has the effects of accurately segmenting anomaly boundaries, extracting multi-dimensional quantitative descriptive parameters, and reducing missed diagnoses and misdiagnoses.

[0019] 3: By calculating the change in confidence level before and after rescanning, and assigning high, medium, and low confidence labels to the abnormal judgment results based on the magnitude and direction of the change, the design achieves a quantitative assessment of the reliability of the judgment conclusion itself. This has the benefits of assisting doctors in prioritizing high-confidence abnormalities, suggesting further imaging examinations for low-confidence areas, and optimizing the priority of clinical decisions.

[0020] 4. By converting palpation impedance parameters, abnormality description parameters, and confidence labels into structured natural language sequences, and inputting them into a medical multimodal large language model, a design is automatically generated that includes palpation findings, pathological correlation analysis, and suggestions for further examination. This design achieves an end-to-end mapping effect from numerical features to medical semantics, which has the benefits of improving report standardization and traceability, and reducing the paperwork burden on doctors.

[0021] In summary, this invention achieves intelligent palpation through a multi-layered design: online estimation of tissue impedance and adaptive adjustment of palpation application parameters; autonomous encrypted rescanning of suspicious areas to accurately confirm abnormal boundaries; assignment of confidence labels based on the change in confidence before and after rescanning; and input of structured palpation data into a medical big language model to generate formatted reports. This enables the stratification of initial screening perception and focused re-examination, the quantification of the reliability of judgment conclusions, and end-to-end mapping of data to semantics. This improves the accuracy of palpation, reduces the risk of missed or misdiagnosed cases, and alleviates the paperwork burden on doctors. Attached Figure Description

[0022] Figure 1 This is a flowchart of an algorithm for judging the palpation results of an embodied intelligent physical examination proposed in this invention; Figure 2 This is a diagram illustrating the on-demand allocation mechanism for edge computing resources proposed in this invention. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the embodiments of the present invention. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0024] Example 1 Reference Figure 1 The present invention provides an embodied intelligent physical examination palpation result judgment algorithm, which is applied to an embodied intelligent agent including a tactile sensor array, a force sensor and an end effector. The hardware platform involved in the algorithm is a local edge computing node, which includes at least one central processing unit, a graphics processing unit or an embedded neural network processor and a shared memory unit. All model inference and control tasks are executed locally and do not rely on cloud computing power.

[0025] First, the initial scanning path of the palpation area is obtained. This initial scanning path is pre-planned and generated by the physician or automatically generated by the vision system. The vision system acquires three-dimensional point cloud data of the area to be examined through a depth camera, uses a human topology recognition algorithm to lock the target anatomical position, and automatically plans a serpentine or grid-like initial screening coverage path. The path is discretized into a series of palpation point coordinates, and the end effector performs initial screening palpation point by point along the initial scanning path. At each touch point, the tactile sensor array acquires a high-dimensional pressure distribution matrix in real time, and the force sensor synchronously acquires the three-dimensional force vector of the end effector, forming tactile signal data and force signal data.

[0026] The online tissue impedance estimation model takes the single-frame pressure distribution matrix and three-dimensional force vector collected at the current moment as input and estimates the tissue equivalent impedance parameters of the current contact area in real time. The model is constructed using a pre-trained deep learning model. In order to effectively capture the strong spatiotemporal correlation of tactile data, the model specifically includes a spatial feature extraction subnetwork and a temporal feature fusion subnetwork. The spatial feature extraction subnetwork uses a graph convolutional neural network to treat each tactile unit in the pressure distribution matrix as a node in the graph. Based on the spatial arrangement relationship of the tactile units, an adjacency matrix is ​​constructed to extract the spatial topological features of the single-frame pressure distribution. The temporal feature fusion subnetwork utilizes a long short-term memory network to receive a continuous multi-frame spatial feature vector sequence output by the spatial feature extraction subnetwork, and fuses it with the three-dimensional force vector at the corresponding time step. Through a gating mechanism, the sequence information is selectively memorized and forgotten, and finally, the three equivalent tissue impedance parameters of the current contact area are regressed and output: tissue equivalent stiffness, tissue equivalent damping coefficient, and tissue equivalent thickness.

[0027] After obtaining the tissue equivalent impedance parameters, the system adaptively adjusts the end effector's palpation application parameters at the next palpation point based on parameter changes, including the applied force amplitude, compression velocity, and dwell time. This adaptive adjustment is achieved through an admittance control model. The contact force and displacement relationship of the end effector is controlled based on a second-order linear system model, the expression of which is: In the formula, For contact force vectors, It is a displacement vector. The velocity vector of the end effector. The acceleration vector of the end effector; For virtual quality parameters, For virtual damping parameters, This is a virtual stiffness parameter.

[0028] The virtual damping parameters are updated in real time based on the tissue's equivalent impedance parameters. and virtual stiffness parameters The specific rules are as follows: When the equivalent stiffness of the organization is higher than the preset stiffness benchmark value, the virtual stiffness parameter is reduced proportionally. At the same time, the target movement speed of the end effector is reduced to reduce the applied force amplitude and the pressing speed. When the equivalent stiffness of the organization is lower than the preset stiffness benchmark value, the virtual stiffness parameter is increased proportionally. At the same time, the target motion speed of the end effector is increased to increase the applied force amplitude and increase the pressing speed; When the equivalent damping coefficient of the tissue is higher than the preset damping benchmark value, a virtual damping parameter is added to the admittance control model. The value of is chosen to extend the response time for the force to reach a steady state, thereby increasing the dwell time. When the equivalent damping coefficient of the organization is lower than the preset damping benchmark value, reduce the virtual damping parameter. The value of is chosen to shorten the dwell time; Virtual quality parameters Maintain a preset constant value throughout the process to ensure the stability of the system's inertial characteristics.

[0029] By applying adaptive impedance through the aforementioned admittance control model, the mechanical stability of the end effector during the contact transition process on different soft and hard tissues is ensured, avoiding oscillations or unstable contact caused by abrupt changes in tissue characteristics.

[0030] The tissue equivalent impedance parameters of each palpation point are processed, and an anomaly probability heatmap of the palpation area is constructed using an anomaly probability prediction model. The anomaly probability prediction model compares the tissue equivalent impedance parameters of each palpation point with the preset normal tissue impedance benchmark value, calculates the deviation of each parameter, and calculates the anomaly score by weighted fusion of the deviations of each parameter using the following formula: In the formula, The number of tissue equivalent impedance parameters. The current palpation point Measured values ​​of the equivalent impedance parameters of the tissue. The preset baseline values ​​for parameters corresponding to normal tissue are... For the first The preset weighting coefficients correspond to the parameters. The abnormal scores of all palpation points are interpolated spatially to form a continuous distribution covering the entire palpation area. After normalization, an abnormality probability heatmap is generated.

[0031] When the anomaly score of any spatial cell in the anomaly probability heatmap exceeds the preset rescan trigger threshold, a rescan encrypted sampling path is automatically generated with that spatial cell as the center.

[0032] The adaptive selection mechanism of the rescanning path strategy is as follows: calculate the directional distribution characteristics of the abnormal score gradient field in the local region centered on the spatial cell; When the directional dispersion of the abnormal score gradient field exceeds the preset isotropic threshold, that is, when the gradient direction is uniformly distributed in all directions, an inward spiral path is generated. When the directional dispersion of the abnormal score gradient field is lower than the preset anisotropy threshold, i.e., the gradient direction is concentrated along a certain principal axis, a reciprocating broken line path is generated along the principal direction of the abnormal score gradient. The sampling point density of the rescanning encrypted sampling path is higher than that of the initial scanning path, and the encryption factor is preset to a configurable parameter.

[0033] During the rescan palpation phase, the force control of the end effector switches to high-resolution mode. The applied force resolution is increased to the preset magnification of the initial screening palpation phase, and the pressing speed is reduced to the preset ratio of the initial screening palpation phase, so that the transition features from hard to soft at the edge of the abnormal area can be captured more precisely.

[0034] Based on the encrypted tactile and force signal data obtained from the rescan palpation, the abnormality judgment result of the abnormal area is obtained, and the boundary of the abnormal area is segmented. Boundary segmentation adopts a region growth algorithm based on tissue equivalent stiffness gradient. The sampling point with the highest tissue equivalent stiffness in the rescanned region is used as the seed point, and growth is carried outward along the direction of stiffness gradient change. When the tissue equivalent stiffness difference between adjacent sampling points is lower than the preset boundary threshold, growth stops. The boundary point sequence is the line connecting the sampling points at the front edge of growth stopping. Simultaneously, quantitative descriptive parameters of the abnormal region are extracted, including the sequence of contour boundary points, area, length along the major axis, length along the minor axis, mean of the tissue equivalent stiffness, variance of the tissue equivalent stiffness, and texture uniformity index.

[0035] For each anomalous region that has been rescanned, the change in confidence level before and after rescanning is calculated, and the initial screening anomalous score of the spatial unit before rescanning is extracted from the anomalous probability heatmap as the initial screening confidence level. The encryption anomaly score of the abnormal region is recalculated from the rescanned encrypted data as the rescan confidence level, and the difference between the rescan confidence level and the initial screening confidence level is used as the change in confidence level.

[0036] Based on the magnitude and direction of the change in confidence level, a preset level of confidence label is assigned to the anomaly judgment result: when the change in confidence level is positive and the absolute value exceeds the preset significance threshold, it indicates that the rescan confirms and strengthens the anomaly judgment, and a high confidence label is assigned. When the change in confidence level is positive but the absolute value does not exceed the preset significance threshold, or when the change in confidence level is negative but the absolute value does not exceed the preset significance threshold, it indicates that the rescan did not significantly change the initial screening conclusion, and a medium confidence label is assigned. When the change in confidence level is negative and the absolute value exceeds the preset significance threshold, it indicates that the rescan tends to overturn the abnormal judgment of the initial screening, assigns a low confidence label, and recommends that the physician further confirm it through imaging methods.

[0037] The tissue equivalent impedance parameters, quantitative descriptive parameters of abnormal areas, and corresponding confidence labels of each palpation point are obtained, and the above multidimensional data are transformed into a structured natural language description sequence. The transformation rule is: convert numerical parameters such as tissue equivalent stiffness and tissue equivalent damping coefficient into textual descriptions; The geometric parameters such as the sequence of boundary points, area, and length of the major axis of the abnormal region are converted into anatomical location and size descriptions. Transform credibility labels into deterministic descriptions.

[0038] The structured natural language description sequence is used as prompt words and input into a pre-trained medical multimodal large language model deployed on a local edge computing node. This large language model is based on the Transformer architecture and has been pre-trained and fine-tuned with medical instructions from a large corpus of medical literature, clinical guidelines and physical examination reports, and has a built-in medical prior knowledge base.

[0039] The model performs semantic understanding and medical reasoning on the input palpation description sequence and outputs a formatted physical examination palpation report.

[0040] The report contains three parts: a detailed description of the findings made during palpation, listing the tissue characteristics and abnormalities found in each palpation area using standardized medical terminology; Potential pathological association analysis, combined with the built-in medical knowledge base, lists the possible pathological states and their correlation with the palpation findings; Further diagnostic tests are recommended, based on the confidence labels of the abnormal areas and the results of pathological correlation analysis. Ultrasound examination, magnetic resonance imaging, or biopsy are recommended as further diagnostic methods.

[0041] Example 2 Reference Figure 2 Based on Embodiment 1, preferably, the algorithm allocates resources on demand for inference tasks on the edge computing platform in the following manner.

[0042] During the initial screening palpation stage, only the online tissue impedance estimation model and the anomaly probability prediction model reside in the system memory. The inference of the online tissue impedance estimation model is executed on the graphics processor at a normal frame rate. After each inference is completed, the inference result is immediately fed into the anomaly probability prediction model for real-time anomaly scoring.

[0043] When the system determines that a spatial cell exceeding the rescan trigger threshold appears in the anomaly probability heatmap, triggering the rescan mechanism, the system utilizes zero-copy shared memory technology to directly map the weight parameters of the pre-trained high-precision boundary segmentation model from non-volatile storage to the free area of ​​the graphics processor's video memory, completing dynamic loading and initialization. The entire process does not require memory copying through the central processing unit. Simultaneously, the system increases the bus communication synchronization refresh rate of the control system from the normal state of the initial screening stage to a preset high-frequency state to shorten the underlying control cycle. This enables synchronous sampling and alignment of force-position-touch multimodal data between the force sensor, tactile sensor array, and end effector, thereby matching the real-time processing requirements of the high-frequency tactile signals generated by the increased sampling frame rate of the tactile sensor array during the rescan stage.

[0044] After the rescan palpation is completed, if there are no other abnormal areas in the current task that would trigger a rescan, the system releases the space occupied by the boundary segmentation model in the graphics processor memory, restores the system bus communication synchronization refresh rate to the normal state of the initial screening stage, and only retains the tissue impedance online estimation model and the abnormal probability prediction model to continue to reside, ensuring efficient use of computing and power resources.

[0045] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention.

Claims

1. An algorithm for judging the palpation results of an embodied intelligent physical examination, applied to an embodied intelligent agent comprising a tactile sensor array, a force sensor, and an end effector, characterized in that, include: The initial scanning path of the palpation area is obtained, and the end effector is controlled to perform initial screening palpation on each palpation point along the initial scanning path, while real-time acquisition of tactile and force signal data. Using a preset online tissue impedance estimation model, the equivalent tissue impedance parameters are estimated based on the tactile and force signal data of the current palpation point, and the palpation application parameters of the next palpation point are adaptively adjusted based on the equivalent tissue impedance parameters. Using a pre-defined abnormality probability prediction model, an abnormality probability heatmap of the palpation area is constructed based on the deviation between the tissue equivalent impedance parameters of each palpation point and the pre-defined normal tissue impedance benchmark value. When the anomaly score of any spatial unit in the anomaly probability heatmap exceeds the preset rescan trigger threshold, a rescan encrypted sampling path is automatically generated with that spatial unit as the center, and rescan palpation is performed on the rescan area to obtain encrypted tactile signal data and force signal data. Based on the encrypted data obtained from the rescan palpation, the abnormality judgment result of the abnormal area is obtained, and the boundary of the abnormal area is segmented and the quantitative descriptive parameters of the abnormal area are extracted. For each abnormal region that has been rescanned, the change in confidence level before and after rescanning is calculated. Based on the magnitude and direction of the change in confidence level, a preset confidence level label is assigned to the abnormal judgment result. The tissue equivalent impedance parameters of each palpation point, the quantitative description parameters of abnormal areas, and the confidence labels are transformed into a structured natural language description sequence, which is then input into a pre-trained medical multimodal large language model to generate a formatted physical examination palpation report.

2. The embodied intelligent physical examination palpation result judgment algorithm according to claim 1, characterized in that, The tissue equivalent impedance parameter includes at least one of tissue equivalent stiffness, tissue equivalent damping coefficient, and tissue equivalent thickness; the palpation application parameter includes at least one of applied force amplitude, pressing speed, dwell time, and kneading motion trajectory radius.

3. The algorithm for judging the palpation results of an embodied intelligent physical examination according to claim 2, characterized in that, The adaptive adjustment of the palpation application parameters at the next palpation point based on the tissue equivalent impedance parameter specifically includes: When the tissue's equivalent stiffness is higher than the preset stiffness reference value, the applied force amplitude is reduced and the pressing speed is decreased; when the tissue's equivalent stiffness is lower than the preset stiffness reference value, the applied force amplitude is increased and the pressing speed is increased. When the tissue's equivalent damping coefficient is higher than the preset damping reference value, the residence time is increased; when the tissue's equivalent damping coefficient is lower than the preset damping reference value, the residence time is shortened.

4. The algorithm for judging the palpation results of embodied intelligent physical examination according to claim 1, characterized in that, The anomaly probability prediction model constructs an anomaly probability heatmap of the palpation area, specifically including: The tissue equivalent impedance parameters of each palpation point are compared with the preset normal tissue impedance benchmark value. Based on the comparison results, the deviations of each parameter are weighted and fused to generate an abnormal score for that palpation point. The abnormal scores of all palpation points are spatially interpolated and normalized to generate an abnormal probability heatmap covering the palpation area.

5. The embodied intelligent physical examination palpation result judgment algorithm according to claim 1, characterized in that, The self-generated rescan encrypted sampling path includes: Centered on the spatial unit that triggers rescanning, an encrypted sampling path with a higher sampling point density than the initial scanning path is generated within a preset radius, following at least one of the preset inward spiral path, outward radial path, or reciprocating broken line path along the main direction of the abnormal score gradient; the applied force resolution of the rescanning palpation is higher than that of the initial screening palpation, and the pressing speed is lower than that of the initial screening palpation.

6. The embodied intelligent physical examination palpation result judgment algorithm according to claim 1, characterized in that, The calculation of the change in confidence level before and after rescanning includes: The initial screening anomaly score of the spatial cell before rescanning is extracted from the anomaly probability heatmap as the initial screening confidence level; the encrypted anomaly score of the anomaly region is calculated from the rescanned encrypted data as the rescan confidence level; the difference between the rescan confidence level and the initial screening confidence level is used as the change in judgment confidence level.

7. The embodied intelligent physical examination palpation result judgment algorithm according to claim 6, characterized in that, The step of assigning a preset level of confidence label to the anomaly judgment result based on the magnitude and direction of the change in confidence level specifically includes: A high confidence label is assigned when the change in confidence level is positive and its absolute value exceeds the preset significance threshold; a medium confidence label is assigned when the change in confidence level is positive but its absolute value does not exceed the preset significance threshold, or when the change in confidence level is negative but its absolute value does not exceed the preset significance threshold; and a low confidence label is assigned when the change in confidence level is negative and its absolute value exceeds the preset significance threshold.

8. The embodied intelligent physical examination palpation result judgment algorithm according to claim 1, characterized in that, The quantitative descriptive parameters of the abnormal region include at least one of the following: the sequence of contour boundary points of the abnormal region, area, length in the major axis direction, length in the minor axis direction, mean of the tissue equivalent stiffness, variance of the tissue equivalent stiffness, and texture uniformity index.

9. The embodied intelligent physical examination palpation result judgment algorithm according to claim 1, characterized in that, The online tissue impedance estimation model employs a pre-trained deep learning model, using the pressure distribution matrix collected by the tactile sensor array and the three-dimensional force vector collected by the force sensor as input, and outputs the tissue equivalent stiffness, tissue equivalent damping coefficient, and tissue equivalent thickness of the current contact area.

10. The embodied intelligent physical examination palpation result judgment algorithm according to claim 1, characterized in that, All model inference tasks deployed on the embodied intelligent agent are executed on the local edge computing platform; the initial screening palpation stage only runs the tissue impedance online estimation model and the anomaly probability prediction model, the rescanning palpation stage dynamically activates the boundary segmentation model, and the inference computing resources of the initial screening palpation stage and the rescanning palpation stage are allocated on demand.

Images