An interactive image segmentation method
Patent Information
- Application Number
- CN202610430429.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-04-02
- Publication Date
- 2026-08-14
AI Technical Summary
[0003]目前的图像分割模型存在的问题在于，其并未关注影像中的不确定区域，对于一些情景下尤其是乳腺BI-RADS 4A影像的灰区病灶的敏感性较弱，最终导致分割精度不佳
本发明提出一种交互式影像分割方法，通过预建立的交互式影像分割模型进行医学影像的病灶与背景分割，交互式影像分割模型共分为三个阶段，1）伪标签生成阶段：基于nnU-Net的精确分割与不确定区域提取；2）多尺度双模型交互阶段：nnU-Net×SAM3互补蒸馏；3）拓扑与置信度约束的开放世界分割阶段：不确定区域作为新类进行联合优化。本发明所提出的基于交互式影像分割模型能够在对医学影像进行分割时，在“病灶与背景”两类主要任务中实现高精度分割，同时自动发现、标记并学习不确定区域，以提升对灰区病灶的敏感性和临床诊断效能。
Smart Images
Figure CN122573994A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of machine learning technology, and in particular to an interactive image segmentation method. Background Technology
[0002] Currently, image segmentation often involves preprocessing acquired medical images, such as denoising, before directly segmenting them using a pre-established image segmentation model. These models are typically built by training and validating labeled lesions and background data to separate the lesions from the background.
[0003] The current image segmentation model has a problem that it does not pay attention to uncertain areas in the image. It is less sensitive to gray area lesions in some scenarios, especially in breast BI-RADS 4A images, which ultimately leads to poor segmentation accuracy. Summary of the Invention
[0004] The purpose of this invention is to at least address one of the shortcomings of the prior art and provide an interactive image segmentation method.
[0005] To achieve the above objectives, the present invention adopts the following technical solution: Specifically, an interactive image segmentation method is proposed, including the following: Acquire target image data; The target image data is input into a pre-established interactive image segmentation model, and the interactive image segmentation model outputs the segmentation result. The operation process of the interactive image segmentation model includes: Pseudo-label generation stage: The target image data is initially segmented and uncertain regions are extracted using nnU-Net; Multi-scale dual-model interaction stage: Complementary distillation of nnU-Net and SAM3; Open-world segmentation stage with topology and confidence constraints: Uncertain regions are treated as new classes for joint optimization, and the segmentation results are finally output.
[0006] Furthermore, specifically, the target image data is initially segmented and uncertain regions are extracted using nnU-Net, including: First, the target image data is processed using nnU-Net to output a lesion prediction mask. Missed detection mask With false detection perception mask ; Then according to , and The results were further expanded to include three lesion prediction masks, including integration. of , integrate of fusion and of , in, , , .
[0007] Furthermore, specifically, complementary distillation of nnU-Net and SAM3 is performed, including, Will , , and After calculating the detection bounding boxes covering the mask, the detection bounding boxes are then enlarged by a factor of one with the center as the origin. , , and ; Each of the following: , ]、[ , ]、[ , ]and[ , The prompt word is input into SAM3, which outputs four lesion prediction probability maps. , , and ; , The predicted probability of the lesion category output by the first branch of nnU-Net; The mean of the probability prediction maps for these four lesions is calculated and defined as: ; in, This indicates the pre-defined weights of the augmented prediction results. Finally, the lesion prediction mask based on the SAM3 mean probability value is defined as follows: .
[0008] Furthermore, specifically, uncertain regions are treated as new classes for joint optimization, ultimately outputting segmentation results, including: use The results are used to optimize the prediction results of the nnUNet network. First use To optimize and The result is defined as: ; ; Finally, the segmentation mask optimized by nnUNet is: .
[0009] Furthermore, specifically, The value is set to 0.7, and the first multiple is set to 1.05.
[0010] Furthermore, specifically, the training phase of the interactive image segmentation model includes the generation of labeled pseudo-labels and the generation of unlabeled pseudo-labels.
[0011] Furthermore, specifically, there is the generation of pseudo-labels for labeled data, including, Define the image data segmentation training set as ,in This represents the number of images in the image data set of the training set. and Let i and i represent the image data of the i-th image and the corresponding lesion segmentation and annotation image, respectively; define The size is Where H and W represent the length and width of the labeled image, respectively; similarly, define and The validation and test sets are for lesion segmentation. This refers to unlabeled image data; Next, build The Uncertain Pseudo-Label Set includes both missed detection pseudo-labels and false detection pseudo-labels; firstly, using... To generate detection boxes , Represents the x-coordinate, y-coordinate, height, and width of the detection box; Just cover The area; then use Inputting SAM3 as a prompt word generates single-channel images with values of 0 and 1. ; indicates that it will be added next. Generated prompt words Input SAM3 to get , The size gradually decreases as the epoch increases, scaling the dimensions. from The rate decreased from 1.5 times to within 1.2 times. The width and height are defined as follows: , The horizontal and vertical coordinates are defined as follows:
[0012]
[0013] That is, when the x-coordinate and y-coordinate of the expanded detection box are greater than 0, their values remain unchanged; otherwise, they are equal to 0. Then, the calculation... and and The difference set generates the missed detection and false detection regions, denoted as follows:
[0014]
[0015] in, , , and express The missed detection and false detection areas are identified using the coarse and fine prediction results generated by SAM3; subsequently, { , }and{ , } Each image is integrated into a single image to obtain and , and It includes three categories: missed detections, false detections, and background; finally, and Used as pseudo-labels to guide the training of nnuNet.
[0016] Furthermore, specifically, the generation of pseudo-labels for unlabeled data includes, In use model After training to convergence, add the unlabeled dataset. Begin the second phase of training; In enter The predicted probabilities of branch one and branch two are obtained from this. and Subsequently and Perform Argmax calculations separately to obtain the lesion prediction masks for branch one and branch two. and error perception mask Next, calculate the high-confidence lesion prediction mask. and missed detection mask With false detection perception mask Defined as:
[0017] in, , and The threshold values are set to 0.7, 0.8, and 0.8 respectively. Then, the calculation is performed. and The union of the sets is denoted as ;for and After calculating the detection bounding boxes covering the mask, the detection bounding boxes are then magnified 1.1 times with the center as the origin. and Finally, with As the input image, and [ , ]and[ , The terms ] are used as clues for lesions and false positives, respectively. After passing through SAM3, a probability graph related to the clues is output, denoted as . and Then, calculate the mask for the lesion area and the false positive area, and denote it as:
[0018] in, and They respectively indicate the use of [ , The output of lesion segmentation regions and high-confidence segmentation regions are used as prompt words. This indicates the use of [ , The high-confidence false positive region is used as the prompt word output; then the unlabeled dataset is calculated. The pseudo-tag is defined as:
[0019] in, , and These represent unlabeled images. The corresponding lesion pseudo-labels, false-detection pseudo-labels, and false-detection pseudo-labels are then calculated; next, the pseudo-labels for the two branch background regions are calculated, first by extracting... and The high-confidence region in the background is denoted as:
[0020] Therefore, the background pseudo-tag is defined as:
[0021]
[0022] Therefore, unlabeled data The pseudo-label of the first branch is defined as:
[0023] Where 0, 1, and 2 represent the category values of the background region, the lesion region, and the unlabeled region, respectively.
[0024] The pseudo-label for the second branch is defined as follows:
[0025] Where 0, 1, 2, and 3 represent the category values of background area, missed detection area, false detection area, and unlabeled area, respectively.
[0026] Furthermore, specifically, the loss calculation of the interactive image segmentation model includes, First, nnU-Net is used as the backbone network for the first stage of training to obtain the segmentation model. The model outputs two branches; the first branch outputs the predicted probabilities for two categories: lesion and background, denoted as . The second branch outputs the predicted probabilities for three categories: missed detection, false detection, and background, denoted as . During training Optimization is performed using Softmax cross-entropy loss and Dice loss. The softmax cross-entropy loss is used for optimization, and different weights are applied to optimize the missed and false positive labels in the coarse and fine prediction results. The segmentation loss of the first branch is defined as follows: ; in, and Each represents and The calculated Softmax cross-entropy loss and Dice loss, where B represents the batch size. The segmentation loss of the second branch is defined as: ; Similarly, and They represent respectively with and Calculate the Softmax cross-entropy loss; and They represent and The weight, Next, the interaction loss between the two branches is calculated, including two types of interaction loss: one is to use the error perception of the second branch to weight the first branch, and the other is to calculate the category logic consistency constraint between branch one and branch two. First, calculate the error perception weight of the second branch, defined as: ; in, and This is used to weigh the probability of missed detection and the probability of false detection; then the error-weighted loss is calculated, defined as: ; The error perception of the second branch does not change the predicted category of the first branch, but spatially reweights the binary classification loss by the error probability; Next, the category logical consistency constraint is calculated and defined as: ; In the second branch, in areas with high FN, the binary classification of the first branch will be more biased towards lesions; in areas with high FP, the binary classification of the first branch will be more biased towards background. The overall loss during training is then defined as: .
[0027] This invention also proposes a breast BI-RADS 4A segmentation method, which applies the steps of the aforementioned interactive image segmentation method, including: A pre-trained interactive image segmentation model is obtained by training an interactive image segmentation model based on breast BI-RADS 4A. Then, the target breast BI-RADS 4A is segmented using the pre-trained interactive image segmentation model.
[0028] The beneficial effects of this invention are as follows: This invention proposes an interactive image segmentation method. It segments lesions and background in medical images using a pre-established interactive image segmentation model. The interactive image segmentation model consists of three stages: 1) pseudo-label generation stage: accurate segmentation and extraction of uncertain regions based on nnU-Net; 2) multi-scale dual-model interaction stage: nnU-Net×SAM3 complementary distillation; 3) open-world segmentation stage with topological and confidence constraints: uncertain regions are treated as new classes for joint optimization. The interactive image segmentation model proposed in this invention achieves high-precision segmentation in the two main tasks of "lesion and background" in medical image segmentation, while automatically discovering, labeling, and learning uncertain regions to improve sensitivity to gray-area lesions and clinical diagnostic efficacy. Attached Figure Description
[0029] The above and other features of this disclosure will become more apparent from the detailed description of the embodiments illustrated in conjunction with the accompanying drawings. In the accompanying drawings, the same reference numerals denote the same or similar elements. Obviously, the drawings described below are merely some embodiments of this disclosure. For those skilled in the art, other drawings can be obtained from these drawings without any creative effort. In the drawings: Figure 1 The diagram shown is a schematic diagram of an interactive image segmentation method according to the present invention. Detailed Implementation
[0030] The following will provide a clear and complete description of the concept, specific structure, and technical effects of the present invention in conjunction with embodiments and accompanying drawings, so as to fully understand the purpose, solution, and effects of the present invention. It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. The same reference numerals used throughout the accompanying drawings indicate the same or similar parts.
[0031] Example 1, referring to Figure 1 This invention proposes an interactive image segmentation method, comprising the following: Acquire target image data; The target image data is input into a pre-established interactive image segmentation model, and the interactive image segmentation model outputs the segmentation result. The operation process of the interactive image segmentation model includes: Pseudo-label generation stage: The target image data is initially segmented and uncertain regions are extracted using nnU-Net; Multi-scale dual-model interaction stage: Complementary distillation of nnU-Net and SAM3; Open-world segmentation stage with topology and confidence constraints: Uncertain regions are treated as new classes for joint optimization, and the segmentation results are finally output.
[0032] In this preferred embodiment 1, high-precision segmentation is achieved in the two main tasks of "lesion and background", while automatically discovering, marking and learning uncertain regions to improve the sensitivity to gray area lesions and clinical diagnostic efficacy.
[0033] This invention also proposes a breast BI-RADS 4A segmentation method, which applies the steps of the aforementioned interactive image segmentation method, including: A pre-trained interactive image segmentation model is obtained by training an interactive image segmentation model based on breast BI-RADS 4A. Then, the target breast BI-RADS 4A is segmented using the pre-trained interactive image segmentation model.
[0034] The interactive image segmentation method proposed in this invention is described below for application in breast BI-RADS 4A segmentation: A new framework for gray-zone tumor segmentation in breast BI-RADS 4A is proposed, based on open-world understanding and multi-model interactive distillation. This framework aims to achieve high-precision segmentation in two main tasks: lesion and background. Simultaneously, it automatically discovers, labels, and learns uncertain regions to improve sensitivity for gray-zone lesions and clinical diagnostic efficacy. The framework consists of three stages: 1) Pseudo-label generation stage: accurate segmentation and uncertain region extraction based on nnU-Net; 2) Multi-scale dual-model interaction stage: nnU-Net×SAM3 complementary distillation; 3) Open-world segmentation stage with topology and confidence constraints: uncertain regions are treated as new classes for joint optimization.
[0035] Pseudo-label generation with labeled data: Define the breast BI-RADS 4A segmentation training set as follows ,in This represents the number of breast ultrasound images in the training set. and Let represent the i-th breast ultrasound image and the corresponding breast tumor segmentation and annotation image (Ground-Truth), respectively. The size is Where H and W represent the length and width of the labeled image, respectively. Similarly, define... and The validation and test sets are for breast tumor segmentation. This is unlabeled breast 4A data.
[0036] Next, build The Uncertain pseudo-label set includes both missed detection pseudo-labels and false detection pseudo-labels. First, using... To generate detection boxes , This represents the x-coordinate, y-coordinate, height, and width of the detection box. Just cover The area. Then use Inputting SAM3 as a prompt word generates single-channel images with values of 0 and 1. . Indicates that you will then join. Generated prompt words Input SAM3 to get . The size gradually decreases as the epoch increases, scaling the dimensions. from The rate decreased from 1.5 times to within 1.2 times. The width and height are defined as follows: , The horizontal and vertical coordinates are defined as follows:
[0037]
[0038] That is, when the x-coordinate and y-coordinate of the expanded detection box are greater than 0, their values remain unchanged; otherwise, they are equal to 0. Then calculate... and and The difference set generates the missed detection and false detection regions, denoted as follows:
[0039]
[0040] in, , , and express The missed detection areas and false detection areas are identified using the coarse and fine prediction results generated by SAM3. Then, { , }and{ , The data are integrated into a single image, categorized into three types: missed detections, false detections, and background. and Finally, and Used as pseudo-labels to guide the training of nnuNet.
[0041] Loss calculation: First, nnU-Net is used as the backbone network for the first stage of training to obtain the segmentation model. The model outputs two branches. The first branch outputs the predicted probabilities for two categories: lesion and background, denoted as . The second branch outputs the predicted probabilities for three categories: missed detection, false detection, and background, denoted as . During training Optimization is performed using Softmax cross-entropy loss and Dice loss. We use Softmax cross-entropy loss for optimization, and employ different weights for missed and false positives in the coarse and fine prediction results. The segmentation loss of the first branch is defined as:
[0042] in, and Each represents and The calculated Softmax cross-entropy loss and Dice loss, where B represents the batch size. The segmentation loss of the second branch is defined as:
[0043] Similarly, and They represent respectively with and Calculate the Softmax cross-entropy loss. and They represent and The weights, during training, are set to... Equals 0.2, It equals 0.8.
[0044] Next, the interaction loss between the two branches is calculated, including two types of interaction loss: one is to use the error perception of the second branch to weight the first branch, and the other is to calculate the category logic consistency constraint between branch one and branch two.
[0045] First, calculate the error perception weight of the second branch, defined as:
[0046] in, and This is used to weigh the probability of false negatives against the probability of false positives. Then, the error-weighted loss is calculated, defined as:
[0047] The error perception in the second branch does not change the predicted category of the first branch, but rather spatially reweights the binary classification loss by applying the error probability.
[0048] Next, the category logical consistency constraint is calculated and defined as:
[0049] In the second branch, in areas with high FN, the binary classification of the first branch will be more biased towards the lesion; in areas with high FP, the binary classification of the first branch will be more biased towards the background.
[0050] The overall loss during training is then defined as:
[0051] Generating pseudo-labels for unlabeled data: When using model After training to convergence, add the unlabeled dataset. Begin the second phase of training. In enter The predicted probabilities of branch one and branch two are obtained from this. and Subsequently and Perform Argmax calculations separately to obtain the lesion prediction masks for branch one and branch two. and error perception mask Next, the high-confidence lesion prediction mask is calculated. and missed detection mask With false detection perception mask Defined as:
[0052] in, , and The threshold values are set to 0.7, 0.8, and 0.8 respectively. Then, the calculation is performed. and The union of the sets is denoted as .for and After calculating the detection bounding boxes covering the mask, the detection bounding boxes are then magnified 1.1 times with the center as the origin. and Finally, with As the input image, and [ , ]and[ , The terms ] are used as clues for lesions and false positives, respectively. After passing through SAM3, a probability graph related to the clues is output, denoted as . and Then, the masks for the lesion area and the false positive area are calculated and denoted as:
[0053] in, and They respectively indicate the use of [ , The output of lesion segmentation regions and high-confidence segmentation regions are used as prompt words. This indicates the use of [ , The high-confidence false positive region is used as the output of the prompt word. Next, the unlabeled dataset is calculated. The pseudo-tag is defined as:
[0054] in, , and These represent unlabeled images. Corresponding lesion pseudo-labels, false positive lesion labels, and false negative lesion labels. In calculation... Consider extracting high-confidence lesion prediction labels from SAM3 as pseudo-labels, while calculating... and In such cases, the standard lesion prediction labels are removed. Next, pseudo-labels for the two branch background regions are calculated, first by extracting... and The high-confidence region in the background is denoted as:
[0055] Therefore, the background pseudo-tag is defined as:
[0056] Therefore, unlabeled data The pseudo-label of the first branch is defined as:
[0057] Where 0, 1, and 2 represent the category values of the background region, the lesion region, and the unlabeled region, respectively.
[0058] The pseudo-label for the second branch is defined as follows:
[0059] Where 0, 1, 2, and 3 represent the category values for background area, missed detection area, false detection area, and unlabeled area, respectively. In the analysis of... When calculating loss using data, use and The data loss is the same. The difference is that... The pseudo-labels for the data failed to cover all areas of the image, so Loss is calculated only in regions with pseudo-labels.
[0060] Interactive segmentation prediction: During the testing phase, the mask is first predicted by the lesion prediction branch of nnUNet. Missed detection mask With false detection perception mask Then according to , and The results can be expanded to include three more lesion prediction masks, including integration. of , integrate of fusion and of . As defined above, and It can be defined as:
[0061] Then , , and After calculating the detection bounding boxes covering the mask, the detection bounding boxes are magnified by 1.05 times with the center as the origin. , , and Then separately [ , ]、[ , ]、[ , ]and[ , The prompt word is input into SAM3, which outputs four lesion prediction probability maps. , , and Subsequently, the mean of the probability maps of these four lesions was calculated and defined as: . in, The weighting for the augmented prediction results is set to 0.7. Finally, the lesion prediction mask based on the SAM3 mean probability value is defined as:
[0062] Finally utilize The results are used to optimize the prediction results of the nnUNet network. First, the results are used... To optimize and The result is defined as:
[0063] Finally, the segmentation mask optimized by nnUNet is defined as:
[0064] Furthermore, the functional modules in the various embodiments of the present invention can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module. The integrated modules described above can be implemented in hardware or as software functional modules.
[0065] If the integrated module is implemented as a software functional module and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the above embodiments of the present invention can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include: any entity or system capable of carrying the computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc.
[0066] Although the description of the invention has been quite detailed and particularly of several described embodiments, it is not intended to limit it to any of these details or embodiments or any particular embodiment, but should be considered as providing a broad possible interpretation of the claims by referring to the appended claims and taking into account the prior art, thereby effectively covering the intended scope of the invention. Furthermore, the invention has been described above with respect to embodiments foreseeable by the inventors in order to provide a useful description, and non-substantial modifications to the invention that have not yet been foreseen may still represent equivalent modifications.
[0067] The above description is merely a preferred embodiment of the present invention. The present invention is not limited to the above-described embodiments. Any embodiment that achieves the technical effects of the present invention using the same means should fall within the protection scope of the present invention. Within the protection scope of the present invention, various modifications and variations can be made to the technical solutions and / or implementation methods.
Claims
1. An interactive image segmentation method, characterized in that, Including the following: Acquire target image data; The target image data is input into a pre-established interactive image segmentation model, and the interactive image segmentation model outputs the segmentation result. The operation process of the interactive image segmentation model includes: Pseudo-label generation stage: The target image data is initially segmented and uncertain regions are extracted using nnU-Net; Multi-scale dual-model interaction stage: Complementary distillation of nnU-Net and SAM3; Open-world segmentation stage with topology and confidence constraints: Uncertain regions are treated as new classes for joint optimization, and the segmentation results are finally output.
2. The interactive image segmentation method according to claim 1, characterized in that, Specifically, the target image data is initially segmented and uncertain regions are extracted using nnU-Net, including: First, the target image data is processed using nnU-Net to output a lesion prediction mask. Missed detection mask With false detection perception mask ; Then according to , and The results were further expanded to include three lesion prediction masks, including integration. of , integrate of fusion and of , in, , , .
3. The interactive image segmentation method according to claim 2, characterized in that, Specifically, complementary distillation of nnU-Net and SAM3 is performed, including: Will , , and After calculating the detection bounding boxes covering the mask, the detection bounding boxes are then enlarged by a factor of one with the center as the origin. , , and ; Each of the following: , ]、[ , ]、[ , ]and[ , The prompt word is input into SAM3, which outputs four lesion prediction probability maps. , , and ; , The predicted probability of the lesion category output by the first branch of nnU-Net; The mean of the probability prediction maps for these four lesions is calculated and defined as: ； in, This indicates the pre-defined weights of the augmented prediction results. Finally, the lesion prediction mask based on the SAM3 mean probability value is defined as follows: 。 4. The interactive image segmentation method according to claim 3, characterized in that, Specifically, uncertain regions are treated as new classes for joint optimization, and the final output segmentation results include: use The results are used to optimize the prediction results of the nnUNet network. First use To optimize and The result is defined as: ； ； Finally, the segmentation mask optimized by nnUNet is: 。 5. The interactive image segmentation method according to claim 3, characterized in that, Specifically, The value is set to 0.7, and the first multiple is set to 1.
05.
6. The interactive image segmentation method according to claim 1, characterized in that, Specifically, the training phase of the interactive image segmentation model includes the generation of pseudo-labels for labeled data and the generation of pseudo-labels for unlabeled data.
7. The interactive image segmentation method according to claim 6, characterized in that, Specifically, there is the generation of pseudo-labels for labeled data, including, Define the image data segmentation training set as ,in This represents the number of images in the image data set of the training set. and Let i and y represent the i-th image data and the corresponding lesion segmentation and annotation image, respectively; define... The size is Where H and W represent the length and width of the labeled image, respectively; similarly, define and The validation and test sets are for lesion segmentation. This refers to unlabeled image data; Next, build The Uncertain Pseudo-Label Set includes both missed detection pseudo-labels and false detection pseudo-labels; firstly, using... To generate detection boxes , Represents the x-coordinate, y-coordinate, height, and width of the detection box; Just cover The area; then use Inputting SAM3 as a prompt word generates single-channel images with values of 0 and 1. ; indicates that it will be added next. Generated prompt words Input SAM3 to get , The size gradually decreases as the epoch increases, scaling the dimensions. from The rate decreased from 1.5 times to within 1.2 times. The width and height are defined as follows: , The horizontal and vertical coordinates are defined as follows: That is, when the x-coordinate and y-coordinate of the expanded detection box are greater than 0, their values remain unchanged; otherwise, they are equal to 0. Then, the calculation... and and The difference set generates the missed detection and false detection regions, denoted as follows: in, , , and express The missed detection and false detection areas are identified using the coarse and fine prediction results generated by SAM3; subsequently, { , }and{ , } Each image is integrated into a single image to obtain and , and It includes three categories: missed detections, false detections, and background; finally, and Used as pseudo-labels to guide the training of nnuNet.
8. The interactive image segmentation method according to claim 7, characterized in that, Specifically, the generation of pseudo-labels for unlabeled data includes, In use model After training converges, add the unlabeled dataset. Begin the second phase of training; In enter The predicted probabilities of branch one and branch two are obtained from this. and Subsequently and Perform Argmax calculations separately to obtain the lesion prediction masks for branch one and branch two. and error perception mask Next, calculate the high-confidence lesion prediction mask. and missed detection mask With false detection perception mask Defined as: in, , and The threshold values are set to 0.7, 0.8, and 0.8 respectively; then, the calculation is performed. and The union of the sets is denoted as ;for and After calculating the detection bounding boxes covering the mask, the detection bounding boxes are then magnified 1.1 times with the center as the origin. and Finally, with As the input image, and [ , ]and[ , The terms ] are used as clues for lesions and false positives, respectively. After passing through SAM3, a probability graph related to the clues is output, denoted as . and Then calculate the mask for the lesion area and the false detection area, and denote it as: in, and They respectively indicate the use of [ , The output of lesion segmentation regions and high-confidence segmentation regions are used as prompt words. This indicates the use of [ , The high-confidence false positive region is used as the prompt word output; then the unlabeled dataset is calculated. The pseudo-tag is defined as: in, , and These represent unlabeled images. The corresponding lesion pseudo-labels, false-detection pseudo-labels, and false-detection pseudo-labels are then calculated; next, the pseudo-labels for the two branch background regions are calculated, first by extracting... and The high-confidence region in the background is denoted as: Therefore, the background pseudo-tag is defined as: Therefore, unlabeled data The pseudo-label of the first branch is defined as: Where 0, 1, and 2 represent the category values of the background region, the lesion region, and the unlabeled region, respectively; The pseudo-label for the second branch is defined as follows: Where 0, 1, 2, and 3 represent the category values of background area, missed detection area, false detection area, and unlabeled area, respectively.
9. The interactive image segmentation method according to claim 8, characterized in that, Specifically, the loss calculation of the interactive image segmentation model includes, First, nnU-Net is used as the backbone network for the first stage of training to obtain the segmentation model. The model outputs two branches; the first branch outputs the predicted probabilities for two categories: lesion and background, denoted as . The second branch outputs the predicted probabilities for three categories: missed detection, false detection, and background, denoted as . During training Optimization is performed using Softmax cross-entropy loss and Dice loss. The softmax cross-entropy loss is used for optimization, and different weights are applied to optimize the missed and false positive labels in the coarse and fine prediction results. The segmentation loss of the first branch is defined as follows: ； in, and Each represents and The calculated Softmax cross-entropy loss and Dice loss, where B represents the batch size; The segmentation loss of the second branch is defined as: ； Similarly, and They represent respectively with and Calculate the Softmax cross-entropy loss; and They represent and The weight, Next, the interaction loss between the two branches is calculated, including two types of interaction loss: one is to use the error perception of the second branch to weight the first branch, and the other is to calculate the category logic consistency constraint between branch one and branch two. First, calculate the error perception weight of the second branch, defined as: ； in, and This is used to weigh the probability of missed detection and the probability of false detection; then the error-weighted loss is calculated, defined as: ； The error perception of the second branch does not change the predicted category of the first branch, but spatially reweights the binary classification loss by the error probability; Next, the category logical consistency constraint is calculated and defined as: ； In the second branch, in areas with high FN, the binary classification of the first branch will be more biased towards lesions; in areas with high FP, the binary classification of the first branch will be more biased towards background. The overall loss during training is then defined as: 。 10. A method for segmenting breast tissue using BI-RADS 4A, characterized in that, The steps of applying the method according to any one of claims 1-9 include: A pre-trained interactive image segmentation model is obtained by training an interactive image segmentation model based on breast BI-RADS 4A. Then, the target breast BI-RADS 4A is segmented using the pre-trained interactive image segmentation model.