A method for precise localization of breast tumors based on multi-level dual-layer coding

Abstract

This invention relates to a method for precise localization of breast tumors based on multi-level double-layer coding, comprising the following steps: acquisition of scattered microwave signals; signal processing and feature construction; multi-level data label coding based on a breast model containing the tumor target: each level contains double-layer coding, the first layer of coding is coarse-grained localization, and the second layer of coding is fine-grained localization; the localization interval of the initial level is determined based on the breast model containing the tumor target; given the constraints of the termination coding, the localization interval of the next level is determined based on the current level's localization interval after data label double-layer coding; model training and coding prediction output.

Description

Technical Field

[0001] This invention belongs to the field of medical signal processing and machine learning technology, and specifically relates to a method for precise localization of breast tumors based on ultra-wideband microwave. Background Technology

[0002] Breast cancer is a major public health issue worldwide, and its early diagnosis and effective intervention still face many challenges. Breast cancer has a high incidence rate and a heavy disease burden, and in recent years, influenced by factors such as population structure and lifestyle, its incidence rate has shown a continuous upward trend and exhibits a certain trend towards affecting younger people. Practice has shown that early detection and early intervention of breast cancer are of great significance for improving treatment outcomes and reducing the risk of death; therefore, there is an urgent need to develop safe, accessible, and repeatable early detection and localization technologies.

[0003] Current clinical methods for breast cancer detection mainly include mammography, ultrasound imaging, and magnetic resonance imaging (MRI). While these methods are widely used, they still have limitations. For example, mammography involves ionizing radiation and has limited sensitivity in some individuals with dense breasts; ultrasound imaging is highly dependent on operator experience, and repeatability and consistency are easily affected; MRI is costly and requires significant equipment resources, making it difficult to promote as a routine screening method on a larger scale. Therefore, there is a real need to develop novel breast cancer detection technologies that are non-ionizing, cost-effective, comfortable to use, and suitable for repeated testing.

[0004] Ultra-wideband microwave detection, as a non-ionizing detection method, has attracted widespread attention in recent years due to its advantages such as low radiation power, low cost, repeatable detection, and high comfort. Its principle lies in the fact that different tissues within the breast have different dielectric properties in the microwave frequency band. When the ultra-wideband microwave signal emitted by the antenna propagates through the tissue, it produces different scattering and reflection responses due to the dielectric differences, thus enabling the received signal to contain characteristic information that can characterize tissue differences. By processing and analyzing the scattered signals, abnormal tissues can be detected, and this can be further used for tumor localization.

[0005] In tumor localization applications, existing methods mostly use quadrant-level results as the main form of localization output. Although this can provide coarse-grained spatial guidance, its output usually corresponds to a large candidate interval, making it difficult to form a "precise localization interval" that can be directly used in clinical operations. If the localization interval is narrowed down by refining the partitions, it is easy to introduce problems such as class size expansion, uneven sample distribution, and difficulty in unifying the annotation standards. This results in the localization results lacking a stable and consistent convergence path across different samples and systems, which in turn affects the stability, reproducibility, and interpretability of the localization output.

[0006] Therefore, it is necessary to propose a precise method for locating breast tumors that is simple in calculation, has clear judgment rules, is scalable, and can balance positioning accuracy and stability. Summary of the Invention

[0007] The purpose of this invention is to overcome the shortcomings of existing technologies in breast tumor localization, which rely solely on quadrant layer results, leading to excessively large localization intervals, difficulty in expressing intra-quadrant differences, and limited localization resolution. This invention proposes a precise breast tumor localization method based on multi-level, dual-layer coded tags. The technical solution of this invention is as follows.

[0008] A method for precise localization of breast tumors based on multi-level dual-layer coding includes the following steps: Step 1, Acquisition of scattered microwave signals: A breast model containing the tumor target is constructed, an antenna array is arranged on the surface of the breast model, ultra-wideband microwaves are emitted through the antennas and the echo signals scattered by the breast tissue are received to acquire scattered microwave signal data; Step 2, Signal Processing and Feature Construction: The scattered microwave signal is preprocessed and features are constructed to obtain a feature representation of the scattered microwave signal, which is used to support subsequent position coding prediction.

[0009] Step 3: Perform multi-level data label encoding based on a breast model containing the tumor target. Each level includes two layers of encoding: the first layer is for coarse-grained localization, and the second layer is for fine-grained localization. The localization range of the initial level is determined based on the breast model containing the tumor target. Given the termination encoding conditions, the localization range of the next level is determined based on the current level's localization range after the two-layer data label encoding. The two-layer data label encoding method for the current level is as follows: (1) Determination of the first layer of coding: The dividing line of the first layer of coding is the quadrant frame. The first layer of location coding of the tumor is determined under the quadrant frame, which is used to characterize the rough spatial location of the tumor. When the tumor crosses the dividing line and the location is not unique, the main region coding is determined according to whether the area of ​​the tumor region in the corresponding quadrant when the tumor crosses the dividing line is greater than the preset coverage ratio threshold. The specific method is as follows: Let P be the area of ​​the tumor in each quadrant. k , The coverage threshold is η, and P is compared sequentially for each quadrant. k With threshold η, when P k When P > η, the tumor is located in that quadrant, and labeled 1 in that quadrant. k <η indicates that the tumor is not located in this quadrant, and this quadrant is marked as 0; the number of quadrants marked as 1 is counted. The first-level code, i.e. the main region code, is output according to the following rules: 1) If N1=1, it means that the tumor is mainly located in a certain quadrant. The main region code takes the corresponding quadrant number, which is called the quadrant code; 2) If N1=2, it means that the tumor crosses a certain dividing line and covers two adjacent quadrants at the same time. The main region code takes the corresponding boundary code; 3) If N1≥3, it means that the tumor covers multiple quadrants at the same time. The main region code takes the corresponding center code. At this time, the localization termination condition is triggered, and the current multi-level double-layer code position label is output.

[0010] (2) Second layer coding determination: When the first layer coding does not trigger the termination condition, the candidate region is determined according to the main region coding. The candidate region is used as a constraint, and the same coverage ratio threshold as the first layer is used to determine the fine-grained positioning sub-region coding of each candidate region.

[0011] (3) Generation of multi-level double-layer coding location labels: The tumor location labels are constructed by using a multi-level double-layer coding method. Each level and each layer is spliced ​​or mapped into the final multi-level double-layer coding location label according to the preset combination rules, so as to realize the multi-level double-layer description of the tumor location.

[0012] Step 4: Model training and encoding prediction output.

[0013] Furthermore, in step three, the preset coverage percentage threshold is 15%.

[0014] Furthermore, in step three, Let the total area of ​​the tumor be S. T The tumor region T is assigned to a quadrant, and the area of ​​the tumor falling into each quadrant is...

[0015] The corresponding tumor area percentages in each quadrant are as follows: .

[0016] Furthermore, in step three (2), the method for generating the fine-grained localization sub-region code of the second layer using the candidate region as a constraint is as follows: when the main region code of the candidate region is a certain quadrant code, the two dividing axes used for quadrant division are used as references, and a reference dividing line parallel to the dividing axis is set inside the quadrant to map the inside of the quadrant into several sub-region categories, and the sub-region code of the tumor is determined accordingly; when the main region code of the candidate region is a certain boundary code, the dividing axis corresponding to the boundary code is mapped into several sub-region categories, and the sub-region code of the tumor is determined accordingly; the same coverage ratio threshold determination method as the first layer is used to count the number of covered sub-regions N2. If N2≥3, it is determined that a "sub-center code" is generated and the termination condition is triggered. Otherwise, according to the sub-region code, the localization interval for the next level of coding is determined, and the data label double-layer coding of the next level continues.

[0017] Furthermore, the method in step four is as follows: using the scattered microwave signal features obtained in step two as the model input, and using the multi-level double-layer coded location labels generated in step three as the supervision signal to train the machine learning model, so that the model learns the mapping relationship between "scattered signal features - multi-level double-layer coding"; in the inference stage, the model outputs the multi-level double-layer coding results and back-maps them to the breast model space according to the preset "coding-spatial interval" mapping relationship, thereby obtaining the corresponding minimum localization candidate interval and outputting the accurate localization result.

[0018] The beneficial effects of this invention are as follows: This invention proposes a method for precise localization of breast tumors based on multi-level dual-layer coding. Using ultra-wideband microwave scattering signals as the information carrier, it acquires multi-channel scattering echoes by deploying an antenna array on the breast surface. A multi-level dual-layer position coding mechanism is introduced into the localization framework, breaking through the resolution limitations of traditional fixed grid division. In each level, the candidate region output from the previous level serves as the input constraint for the next level. The target range is continuously narrowed through coarse and fine spatial division within each level, and the generation of a "center code" or "sub-center code" serves as an adaptive termination condition. A maximum allowed number of coding levels can also be set as the termination. Through a dynamically progressive convergence method, the localization results can adaptively and continuously converge to the smallest candidate interval that accurately matches the physical scale of the tumor, thereby significantly improving localization resolution and directionality. This method achieves dynamic, step-by-step approximation and final precise localization of breast tumors, providing a more directional spatial reference for subsequent clinical decision-making. Attached Figure Description

[0019] Figure 1 Schematic diagrams of four types of breast models; Figure 2 This diagram illustrates how to classify a tumor based on an area threshold when it is located at the boundary. Figure 3 This is a schematic diagram of a certain level of the coding system of the multi-level dual-layer coding method proposed in this invention, wherein: (1) it shows the main region coding set formed by "quadrant + decision boundary / center" on the basis of the traditional four-quadrant framework; (2) it shows the sub-region coding set obtained by applying different refinement rules to different main regions; (3) it shows the single-level dual-layer label set formed by the combination of main region coding and sub-region coding (example is 49 types of labels), which reflects the response results of coding rules to different spatial situations, and is not a fixed geometric block of breast space; Figure 4 This is a schematic diagram of the breast tumor precise localization process based on multi-level double-layer coding proposed in this invention (taking a single-level double-layer coding as an example). Figure 5This is a schematic diagram comparing the minimum localization interval of performing one "first-level double-layer" encoding in the multi-level double-layer encoding method of the present invention with that of traditional 4-partition localization, wherein: (1) is a schematic diagram of a realistic two-dimensional cross-sectional sample of the breast and a tumor constructed based on MRI segmentation; (2) is a schematic diagram of the minimum localization interval output by the scheme of the present invention (example label is L). 2-6 (3) is a schematic diagram of the minimum positioning interval (quadrant layer interval) output by the traditional 4-partition scheme, which is used to compare with the effect of 49 coding regions generated by the "first-level double-layer" coding in the embodiment. Detailed Implementation

[0020] To make the objectives, technical solutions, and beneficial effects of this invention clearer, the invention will be further described below with reference to the accompanying drawings and embodiments. It should be understood that the following embodiments are only for explaining the invention and do not constitute a limitation on the scope of protection of this invention. Other implementation methods obtained by those skilled in the art based on the embodiments of this invention without creative effort should all fall within the scope of protection of this invention.

[0021] This method uses ultra-wideband microwave detection as the signal acquisition means. By emitting and receiving scattered microwave signals through an antenna array deployed on the surface of the breast, multi-channel scattered signals reflecting the dielectric differences of the tissue are obtained. In the data annotation stage, a reusable multi-level dual-layer label system is adopted, so that each level of label is constrained by the previous level of candidate regions to further converge the localization interval, thereby providing higher resolution, more consistent and interpretable spatial supervision information under the premise of controllable category size. In the localization implementation stage, the scattered microwave signals are preprocessed and feature constructed as necessary, and then used together with the corresponding multi-level dual-layer coded position labels to train the machine learning model. Then, during inference, the model outputs multi-level dual-layer codes and directly outputs the smallest localization candidate interval corresponding to the code according to the preset "code-interval" mapping, as the accurate localization result of the tumor. Through the above integrated process of "microwave signal acquisition - multi-level dual-layer coding annotation - model learning and coding prediction - interval mapping output", this invention effectively refines the localization result from "quadrant level" to "intra-quadrant sub-region level" and smaller range without significantly increasing the annotation complexity, thus achieving accurate localization of breast tumors.

[0022] This embodiment constructs a realistic two-dimensional cross-sectional model of the breast based on the segmentation results of real MRI images, and randomly inserts tumor targets of different sizes and positions into the breast model to generate data samples for precise localization of early breast tumors. The technical solution of this invention is described in detail below with an actual workflow: 1. Acquisition of scattered microwave signals: Figure 1This paper illustrates cross-sectional views of four representative realistic breast models, each containing a skin layer, fat layer, glandular layer, and tumor target. This embodiment uses a Type II breast model, with an antenna array uniformly arranged on the breast surface. A pre-defined transmit-receive strategy is used to acquire multi-channel scattered echo signals, forming raw observation data for precise tumor localization. Specifically, this embodiment uses 12 antennas operating in a "single-transmit, multiple-receive" manner: one antenna transmits at a time, while the other 11 antennas simultaneously receive the scattered echoes. After all 12 antennas complete one transmission, a total of 12 × 11 = 132 channels of signal are obtained. Each channel signal is a discrete-time sequence, with 2000 samples per signal. Regarding the transmitted signal type, this embodiment uses a Gaussian first-order guided pulse with a center frequency of 6 GHz as the excitation signal. The above acquisition process is repeated for breast models with different tumor locations / sizes, resulting in 4500 sets of sample data.

[0023] Therefore, the input of step (1) is the breast model and antenna arrangement / excitation configuration, and the output is a multi-channel scattered echo signal set containing 132 channels and 2000 points per channel, providing a unified data foundation for subsequent feature construction and coding prediction. 2. Signal processing and feature construction: The multi-channel scattered microwave signal obtained in step (1) is subjected to time-frequency feature analysis and feature dimensionality reduction to obtain feature vectors for subsequent coding prediction, and used for subsequent positioning inference of the minimum positioning candidate interval.

[0024] Specifically, this embodiment performs dual-tree complex wavelet transform (DTCWT) decomposition on each channel signal, using an 8-level decomposition to obtain wavelet coefficients for each level. Subbands with concentrated energy are selected based on energy proportions, and statistical features are extracted from the selected subband coefficients and the original time-domain waveform. These statistical features include, but are not limited to, maximum, minimum, mean, variance, kurtosis, skewness, integral, and information entropy, thereby obtaining the feature vector corresponding to each group of channel signals. Further, the feature vectors are subjected to dimensionality reduction processing (e.g., principal component analysis or feature selection), and the dimensionality reduction mapping is kept consistent during the training and inference stages to obtain a consistent and reproducible localization input feature representation.

[0025] Therefore, the input of step (2) is a set of multi-channel echo signals, and the output is a feature vector representation used for machine learning model training / inference.

[0026] 3. Multi-level two-layer position tag construction (taking coarse and fine two-layer encoding within a single-level loop as an example): (1) Establishment of coordinate system and quadrant division. A relative coordinate system is established in the plane of the breast model. Two pairs of antennas are selected that are opposite each other on the surface of the breast and approximately perpendicular to each other. The direction of their connecting line is taken as the direction of the two coordinate axes. The intersection of the two coordinate axes is taken as the geometric center or centroid of the breast region, which is used to divide the breast region into four quadrants. For consistency, the two axes are denoted as I and I below. x and I y The four quadrants are designated as Q1, Q2, Q3, and Q4 (quadrant numbering follows the numbering conventions agreed upon in this manual).

[0027] (2) First-layer coding determination (coarse-grained localization). Since tumor locations are randomly distributed, tumor regions may cross quadrant boundaries. To ensure stable and reproducible coding output for samples crossing boundaries, this invention introduces consistency processing such as "boundary code / center code" on top of quadrant coding, and determines the quadrant affiliation using a coverage percentage threshold. The tumor region T is assigned to a quadrant. Let the total tumor area be S. T The area of ​​the tumor falling into each quadrant is

[0028] The corresponding area proportions are:

[0029] In this embodiment, the coverage percentage threshold η is set to 15%. P is compared sequentially for each quadrant. k With threshold η, when P k When P > η, the tumor is located in that quadrant, and labeled 1 in that quadrant. k If the value is less than η, the tumor is not located in that quadrant, and that quadrant is marked with 0. After marking all four quadrants, count the number of quadrants marked with 1. The first-level code is output according to the following rules: 1) If N1=1, it means the tumor is mainly located in a certain quadrant, and the main region code takes the corresponding quadrant code; 2) If N1=2, it means the tumor crosses a certain dividing line and covers two adjacent quadrants at the same time, and the main region code takes the corresponding boundary code; 3) If N1≥3, it means the tumor covers multiple quadrants at the same time, usually corresponding to the tumor being located near the center of the model or crossing two dividing lines, and the main region code takes the corresponding "center code". At this time, the localization termination condition is triggered, and all current and subsequent subdivision operations are directly exited, and the current multi-level two-layer encoded position label is output.

[0030] Figure 2The diagrams illustrating the three scenarios—"unique quadrant," "boundary crossing," and "center crossing"—are provided to explain the output set and decision logic of the first-level encoding. Type 1 is the "unique quadrant" scenario, where the main region encoding takes the corresponding quadrant code. In this case, N1=1, meaning the main region encoding is L1. Type 2 is another "unique quadrant" type. Since the tumor is mainly located in the first quadrant, and the area in the fourth quadrant is less than 15% of the total area, it indicates that the tumor is mainly located in the first quadrant. N1=1, and the main region encoding still takes the corresponding quadrant code, the same as Type 1, also L1. Type 3 is the "boundary crossing" scenario. Since the area of ​​the tumor in both the first and fourth quadrants is greater than 15% of the total area, N1=2, meaning the tumor crosses a certain dividing line and simultaneously covers two adjacent quadrants. The main region encoding is determined based on the corresponding boundary code. This invention specifies the following rules for determining the boundary code: if it spans both the first and second quadrants, the boundary code is 7; if it spans both the second and third quadrants, the boundary code is 5; if it spans both the third and fourth quadrants, the boundary code is 8; and if it spans both the fourth and first quadrants, the boundary code is 6. Specifically... Figure 2 Type 3 uses a main region code of 8. Type 4 is the "center-crossing" case, where the center code is specified as 0. Specifically... Figure 2 Type four: The tumor spans four quadrants, the main region is encoded as 0, and the termination condition is triggered.

[0031] In this embodiment, the tumor diameter is d=5. mm, its total area is approximately 19.63 mm². 2 After determining the main region, the main region code of the sample is L2 (the termination condition was not triggered, and the second layer of coding is entered).

[0032] (3) Second-layer encoding determination (fine-grained localization). When the first-layer main region encoding is determined and the termination condition is not triggered, the candidate region corresponding to the first-layer encoding is used as a constraint, and reference dividing lines are introduced only within the candidate region for refinement, thereby obtaining the second-layer sub-region encoding. When the main region encoding is a quadrant code, I is used as the basis for further refinement. x and I y For reference, a reference dividing line parallel to the coordinate axis is set within the quadrant, and the quadrant is mapped into 9 sub-regions; when the main region is encoded as a boundary code, the coordinate line corresponding to the boundary code is subdivided into 3 sub-regions. Figure 3 (2) The refined sets shown are L3, L5 and L6. In the second layer of refinement, the same coverage ratio threshold determination method as the first layer is used to count the number of covered sub-regions N2. If N2≥3, then a "sub-center code" is generated and the termination condition is triggered.

[0033] Figure 3It shows the set of encoding results and the corresponding refinement relationship (i.e. the set of encoding responses "from coarse to fine") of the multi-level two-layer coding system under the condition of two-layer partitioning, which is used to illustrate the composition and hierarchical rules of position labels.

[0034] After the primary region of the first layer is determined to be L2, the candidate region corresponding to the first layer encoding is the second quadrant, such as... Figure 3 As shown, each quadrant code corresponds to 9 sub-regions. Based on the sub-regions divided in the second quadrant, the tumor location label is determined as L. 2-6 .

[0035] To fully and vividly illustrate the "single-pole two-layer, multi-level iterative" dynamic coding system of this invention, the above embodiments specifically demonstrate the execution process of the first-level loop (i.e., the first-level two-layer). In practical applications, this invention adopts a multi-level nested iterative mechanism: if the "center code" does not appear in the first-level second-layer coding (e.g., L2-6 is output), the system does not stop, but uses the physical candidate region corresponding to L2-6 as the "initial positioning interval" of the second-level loop (i.e., equivalent to magnifying and mapping the tumor candidate sub-region), re-establishes the local relative coordinate system within the constraint boundary, and repeats the above-mentioned "first-level coarse-grained positioning" and "second-level fine-grained positioning" in the second-level loop. This multi-level loop will continue to iterate, with the sub-region output by the previous level always serving as the constraint, until the "center code" determination is triggered at a certain level, or the maximum allowed coding level preset by the system is reached, at which point the calculation is completely terminated and the final accurate positioning result is output.

[0036] 4. Model Training and Encoding Prediction Output: Using the scattered microwave signal features obtained in step 2 as model input, and the multi-level double-layer encoded location labels constructed in step 3 as supervisory signals, the machine learning model is trained to learn the mapping relationship between "scattered signal features - multi-level double-layer encoding". During the inference phase, the scattered microwave signal to be tested is input into the model according to the same processing flow, outputting the double-layer multi-level encoding result. Based on the preset "encoding-spatial interval" mapping relationship, the encoding is back-mapped to the breast model space to obtain the corresponding minimum localization candidate interval and output the precise localization result. For example... Figure 4 As shown, a flowchart of the overall process for precise localization of breast tumors based on two-layer multi-level coding is presented to illustrate the integrated localization path from "signal acquisition - feature construction - two-layer coding - coding prediction - interval mapping output".

[0037] 5. Data Analysis. To quantify the multi-level, two-layer effect, let the area of ​​the breast region be B, and the area of ​​the output region at the m-th level be R. m Define the interval contraction ratio of the corresponding boundary as:

[0038] Using the quadrant candidate interval area R1 as a benchmark, and the second layer output candidate interval area R2, then the shrinkage ratio ρ2 = R2 / R1. In this example, the second layer encoding satisfies this condition in general samples. That is, the area of ​​the minimum localization candidate interval output by the second layer does not exceed 25% of the candidate interval of the quadrant layer, corresponding to a candidate interval area reduced by at least 4 times; when the tumor is located near the segmentation line or other complex situations, the candidate interval can be further converged through a consistent judgment rule, which can achieve This means the minimum localization interval can be as low as approximately 10% of the quadrant layer, corresponding to a reduction of the candidate interval area by about 10 times. Simultaneously, this scheme introduces a consistent decision rule for a small number of samples crossing the segmentation line during the main region and sub-region determination process to unify the encoding output. This makes the multi-level, two-layer location label and interval mapping results stable and reproducible, reducing label ambiguity and inconsistency caused by boundary samples, thereby improving label consistency, interpretability, and overall robustness in subsequent model training and evaluation. Figure 5 As shown, where Figure 5 (2) The minimum positioning interval output by the two-level double-layer encoding method used in this invention (example label L) 2-6 (The rectangle represents the corresponding candidate interval) Figure 5 (3) is the minimum positioning interval of the quadrant layer output by the traditional four-partition scheme. The comparison between the two shows that the present invention can significantly reduce the positioning range and improve the positioning directionality while maintaining clear and reproducible rules.

[0039] The above embodiments are merely illustrative of the technical concept and features of the present invention and should not be construed as limiting the scope of protection of the present invention. All equivalent changes or modifications made in accordance with the spirit and essence of the present invention should be covered within the scope of protection of the present invention.

Claims

1. A method for precise localization of breast tumors based on multi-level dual-layer coding, comprising the following steps: Step 1, Acquisition of scattered microwave signals: A breast model containing the tumor target is constructed, an antenna array is arranged on the surface of the breast model, ultra-wideband microwaves are emitted through the antennas and the echo signals scattered by the breast tissue are received to acquire scattered microwave signal data; Step 2, Signal Processing and Feature Construction: The scattered microwave signal is preprocessed and features are constructed to obtain a feature representation of the scattered microwave signal, which is used to support subsequent position coding prediction; Step 3: Perform multi-level data label encoding based on a breast model containing the tumor target. Each level includes a two-layer encoding: the first layer is for coarse-grained localization, and the second layer is for fine-grained localization. The initial localization range is determined based on the breast model containing the tumor target. Given the termination encoding constraints, the next level's localization range is determined after the current level's localization range through the two-layer data label encoding. The current level data label two-layer encoding method is as follows: (1) Determination of the first layer of coding: The dividing line of the first layer of coding is the quadrant frame. The first layer of location coding of the tumor is determined under the quadrant frame, which is used to characterize the rough spatial location of the tumor. When the tumor crosses the dividing line and the location is not unique, the main region coding is determined according to whether the area of ​​the tumor region in the corresponding quadrant when the tumor crosses the dividing line is greater than the preset coverage ratio threshold. The specific method is as follows: Let P be the area of ​​the tumor in each quadrant. k , The coverage threshold is η, and P is compared sequentially for each quadrant. k With threshold η, when P k When P > η, the tumor is located in that quadrant, and labeled 1 in that quadrant. k <η indicates that the tumor is not located in this quadrant, and this quadrant is marked as 0; the number of quadrants marked as 1 is counted. The first-level code, i.e., the main region code, is output according to the following rules: 1) If N1=1, it means that the tumor is mainly located in a certain quadrant. The main region code takes the corresponding quadrant number, which is called the quadrant code; 2) If N1=2, it means that the tumor crosses a certain dividing line and covers two adjacent quadrants at the same time. The main region code takes the corresponding boundary code; 3) If N1≥3, it means that the tumor covers multiple quadrants at the same time. The main region code takes the corresponding center code. At this time, the localization termination condition is triggered, and the current multi-level double-layer code position label is output. (2) Second layer coding determination: When the first layer coding does not trigger the termination condition, the candidate region is determined according to the main region coding. The candidate region is used as a constraint, and the same coverage ratio threshold as the first layer is used to determine the fine-grained positioning sub-region coding of each candidate region. (3) Generation of multi-level double-layer coding location labels: The tumor location labels are constructed by using a multi-level double-layer coding method. Each level and each layer is spliced ​​or mapped into the final multi-level double-layer coding location label according to the preset combination rules, so as to realize the multi-level double-layer description of the tumor location. Step 4: Model training and encoding prediction output.

2. The method for precise localization of breast tumors according to claim 1, characterized in that, In step three, the preset coverage percentage threshold is 15%.

3. The method for precise localization of breast tumors according to claim 1, characterized in that, In step three, Let the total area of ​​the tumor be S. T The tumor region T is assigned to a quadrant, and the area of ​​the tumor falling into each quadrant is... Correspondingly, the area proportion of the tumor in each quadrant is as follows: 。 4. The method for precise localization of breast tumors according to claim 1, characterized in that, In step 3(2), the method for generating the fine-grained localization sub-region code of the second layer using the candidate region as a constraint is as follows: When the main region code of the candidate region is a certain quadrant code, the two dividing axes used for quadrant division are used as references, and a reference dividing line parallel to the dividing axis is set inside the quadrant to map the inside of the quadrant into several sub-region categories, and the sub-region code of the tumor is determined accordingly; When the main region code of the candidate region is a certain boundary code, the dividing axis corresponding to the boundary code is mapped into several sub-region categories, and the sub-region code of the tumor is determined accordingly; The number of covered sub-regions N2 is counted using the same coverage ratio threshold determination method as the first layer. If N2≥3, it is determined that a "sub-center code" is generated and the termination condition is triggered. Otherwise, the localization interval for the next level of coding is determined according to the sub-region code, and the data label double-layer coding of the next level continues.

5. The method for precise localization of breast tumors according to claim 1, characterized in that, The method in step four is as follows: using the scattered microwave signal features obtained in step two as the model input, and using the multi-level double-layer coded location labels generated in step three as the supervision signal to train the machine learning model, so that the model learns the mapping relationship between "scattered signal features - multi-level double-layer coding"; in the inference stage, the model outputs the multi-level double-layer coding results and back-maps them to the breast model space according to the preset "coding-spatial interval" mapping relationship, thereby obtaining the corresponding minimum localization candidate interval and outputting the accurate localization result.

Images