An active learning method for medical image segmentation based on superpixels and diversity

Abstract

The present invention discloses an active learning medical image segmentation method based on superpixels and diversity, comprising the following steps: splitting images in a medical image dataset into small samples of fixed size, constructing a labeled sample pool and an unlabeled sample pool; establishing a medical image segmentation model and a similarity assessment model, and setting the labeling budget required in the active learning strategy; constructing an active learning strategy based on information content and diversity, evaluating samples in the unlabeled pool, screening out the most valuable unlabeled samples, having them annotated by experts and adding them to the labeled sample pool, deducting the labeling budget, and iterating training; updating the medical image segmentation model parameters, iterating the active model until the labeling budget is zero; and segmenting samples that lack labeling. The present invention overcomes the drawback of traditional active learning query methods that only focus on the amount of sample information, adds a measure of the relationship between samples, and further improves the accuracy of medical image segmentation when using a small number of labeled samples.

Description

Technical Field

[0001] The present invention belongs to the field of artificial intelligence and computer vision technology, and specifically relates to an active learning medical image segmentation method based on superpixels and diversity. Background Art

[0002] In recent years, with the widespread application of artificial intelligence in medicine, medical image analysis processes are increasingly incorporating computer vision analysis techniques. Active learning-based medical image segmentation has garnered significant attention. Fully supervised deep learning methods traditionally rely on large amounts of accurately labeled data with uniform sample distribution. Active learning methods can distinguish between key and redundant samples in unlabeled data, significantly improving labeling efficiency and increasing available data.

[0003] Currently, the medical image segmentation methods based on active learning have the following main problems: 1. The unlabeled sample pool is constructed through a region-based method, and redundant information will appear in the samples; 2. The query strategy only considers the effect of the model on the segmentation results of unlabeled samples, and does not consider the relationship between multiple samples.

[0004] Current medical image segmentation methods based on active learning only measure the uncertainty of the model for samples, ignoring the diversity of unlabeled samples. When constructing samples, medical images are often cropped into non-overlapping regions of fixed size in a region-based manner, resulting in samples containing more redundant information and affecting the final effect of the segmentation model.

[0005] In existing medical image processing methods based on active learning, such as the patent application for a federated learning method, system, device, and medium for medical image segmentation (publication number CN114140478A), the measurement of samples often only considers the amount of information in the sample itself, that is, whether the model can accurately identify the sample. Such methods lack the construction of relationships between multiple samples, making it difficult for active learning queries to obtain out-of-distribution samples. Summary of the Invention

[0006] In order to overcome the shortcomings of the above-mentioned existing technologies, the purpose of the present invention is to provide an active learning medical image segmentation method based on superpixels and diversity, which divides unlabeled samples according to the prior information of the image, reduces the redundancy of random division, and increases the total amount of important samples; at the same time, it introduces a sample diversity mechanism to overcome the defect that the traditional active learning query method only focuses on the amount of sample information, increases the measurement of the relationship between samples, solves the problems existing in the existing technology, and further improves the accuracy of medical image segmentation when using a small number of labeled samples.

[0007] In order to achieve the above object, the technical solution adopted by the present invention is:

[0008] A medical image segmentation method based on active learning of superpixels and diversity, comprising the following steps;

[0009] S1, split the images in the medical image dataset into small samples of fixed size, and construct a labeled sample pool and an unlabeled sample pool;

[0010] S2: Establish a medical image segmentation model and a similarity assessment model, and set the annotation budget required for the active learning strategy. This step requires establishing a medical image segmentation model and a similarity assessment model, and setting the annotation budget required for the active learning strategy. The annotation budget will be gradually consumed during the active learning process. When the annotation budget is consumed to zero, the active learning iteration terminates.

[0011] S3: Build an active learning strategy based on information volume and diversity, evaluate the samples in the unlabeled pool in S1, screen out the most valuable unlabeled samples, have them annotated by experts and add them to the labeled sample pool, deduct the annotation budget, iterate the training, update the parameters of the medical image segmentation model, and iterate the active model until the annotation budget reaches zero;

[0012] S4, uses the updated iterative medical image segmentation model to segment samples that lack annotations.

[0013] In step S1, the medical image sample pool Includes annotated pools and unlabeled sample pool

[0014]

[0015] in, is the labeled sample pool, I i represents the i-th sample with annotation, L i Indicates I i Corresponding annotated diagram; is a set of unlabeled medical image samples, u={U1,U2,…}, U j represents the jth unlabeled sample.

[0016] The step S1 is specifically as follows:

[0017] S11, split the labeled image samples to obtain small-size samples for training the model, forming a labeled sample pool; for a pair of labeled images I and corresponding annotations L, divide the image and samples into several non-overlapping small areas with width w and height h in a unified way These small samples constitute the labeled sample pool

[0018] S12, split the unlabeled image and obtain multiple small areas with similar characteristics in color and texture through superpixel algorithm. According to the center position of the area and the fixed width w and height h, the image is divided into multiple small samples {U1, U2, ...} of fixed size. All unlabeled small samples constitute the entire unlabeled sample pool.

[0019] The step S2 is specifically as follows:

[0020] S21, constructing a medical image segmentation model M, which consists of an encoder E for compressing information to extract features and a decoder D for restoring scale. The medical image segmentation model M is used to segment unlabeled medical image samples. Multiple active learning iterations are required to obtain accurate segmentation results. During the active learning iterations, the medical image segmentation model M participates in two main processes. The first process is model training, which takes labeled medical image samples as input and outputs segmentation results. The loss function between the segmentation results and the annotations is calculated, and the model parameters are optimized through backpropagation. The second process is model inference, which takes unlabeled medical image samples as input and obtains segmentation results.

[0021] S22, build similarity evaluation model Model sim , Model sim The input is two samples, and the output is the sample similarity distance. By calculating the similarity distance between a sample and all other samples, the representative score of this sample to the entire sample pool is calculated, which is used to calculate the diversity of the sample.

[0022] S23, set the labeling budget. The labeling budget is an important condition for determining whether the active learning iteration is terminated. In each round of active learning iteration, the number of labeled samples will be increased through expert labeling, and a certain labeling budget will be deducted each time. When the labeling budget is consumed to zero, the active learning iteration stops and the final medical image segmentation model is obtained.

[0023] The step S3 is specifically performed according to the following steps:

[0024] S31, using a labeled sample pool In the active learning strategy, the medical image segmentation model M undergoes t active learning iterations, denoted as M t , the model that has not undergone active learning iterative training is recorded as M0;

[0025] S32, use the trained model to infer all samples in the unlabeled pool to obtain a set of sample segmentation probability maps and the segmentation result graph collection

[0026] S33, after obtaining the probability map and segmentation result map of all unlabeled samples, the information query function Query info (·) Evaluate the sample; Query info (·) is realized by calculating the segmentation probability map and the result map, and the unlabeled sample U i The query is as follows:

[0027] Query info (U i )=Q class (Y i )×Q uncertainy (P i )

[0028] Q class (Y i )=Σ c∈C I c (Y i )

[0029]

[0030] Among them, U i Refers to the i-th unlabeled sample, Y i and P i They are the corresponding segmentation result map and probability map, Q class (Y i ) means Sample U i The category score, C is the total number of categories on the sample segmentation map, x represents the position, I_c(·) is the indicator function, Y i (x) represents the segmentation map Y i The value at pixel location x; Q uncertainy (P i ) means Sample U i The uncertainty score, y represents the position coordinate of the segmentation probability map, P i (y) represents the probability map P i The probability value of the upper position being y;

[0031] S34, query the unlabeled sample U according to the feature set of the unlabeled sample i With the marked pool and the unlabeled pool except U i Similarity of other samples; Similarity query function Query sim (·) is determined by the unlabeled similarity S U Similarity S with the marked L It consists of two parts, for sample U i The similarity query is as follows:

[0032] Query sim (U i )=S U (U i )-S L (U i )

[0033]

[0034]

[0035] Among them, Model sim For the similarity measurement network, input two samples and get their similarity evaluation. Here, all labeled samples and samples except U are evaluated. i The unlabeled samples outside the test are measured and the results are fed into the similarity query function Query sim (·), get the sample U i Similarity score of

[0036] S35, sum the information score and diversity score of the sample to obtain the sample value score:

[0037] Value(U i )=Query info (U i )+Query sim (U i )

[0038] Sort all unlabeled samples in descending order of sample value scores and select the top K samples. As samples that need to be labeled in the tth round of active learning iteration;

[0039] S36, K samples Hand it over to experts for labeling, use the labeling cost, and get K labeled samples Update sample pool and mark budget:

[0040]

[0041]

[0042] Budget t+1 =Budget t -Cost t

[0043] Wherein, the superscript t represents the tth active learning iteration;

[0044] Furthermore, before the labeling budget is exhausted, the active learning strategy cycles in the order of S31 to S37 until the labeling budget reaches 0.

[0045] The medical image segmentation model in step S4 is obtained from the segmentation model that is updated after multiple trainings after the annotation budget in step S3 is exhausted.

[0046] Beneficial effects of the present invention:

[0047] (1) In the sample construction step, the present invention segments the unlabeled image through the superpixel method, clusters the regions with similar texture, color and other features, and obtains fixed-size samples on this basis. While removing the redundant parts of the image, it improves the problem that the randomly divided image cannot effectively focus on the target object and cannot effectively capture the key areas, so that the unlabeled samples can contain more of the same content features, thereby providing rich guiding prior information for image segmentation.

[0048] (2) The present invention decomposes the query in the active learning strategy into the measurement of sample diversity and inter-sample diversity. The former evaluates the characteristics of the sample itself, and the latter measures the relationship between multiple samples. It solves the problem that the traditional query method only focuses on the quality of the model's sample segmentation results and lacks the measurement of the relationship between samples. At the same time, the diversity measurement is dynamically updated as the active learning iteration is performed, and it can maintain high efficiency and accuracy during the sample pool update process.

[0049] (3) The present invention can use the latest segmentation model and similarity measurement model, and has good flexibility and scalability. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 This is a flowchart of an active learning medical image segmentation method based on superpixels and diversity according to an embodiment of the present invention.

[0051] Figure 2 2 is a diagram of the sample segmentation model architecture of an embodiment of the present invention.

[0052] Figure 3 4 is a diagram of a sample similarity measurement network architecture according to an embodiment of the present invention.

[0053] Figure 4 Schematic diagram of the active learning iterative strategy according to an embodiment of the present invention. DETAILED DESCRIPTION

[0054] The present invention will be described in further detail below with reference to the accompanying drawings.

[0055] The embodiment of the present invention provides an active learning medical image segmentation method based on superpixels and diversity, such as Figure 1 As shown, the following steps are included:

[0056] S1, split the images in the medical image dataset into small samples of fixed size, build the sample pool required for active learning, medical image sample pool Includes annotated pools and unlabeled sample pool

[0057]

[0058] in, is the labeled sample pool, I i represents the i-th sample with annotation, L i express I i corresponds to the labeled image; u is the set of unlabeled medical image samples, U = {U1, U2, ...}, U j represents the jth unlabeled sample.

[0059] In particular, the sample pool has been labeled The pool of unlabeled samples U will be updated with the iteration of active learning. Here, the superscript t represents the number of iterations of active learning, for example: represents the labeled sample pool of the tth active learning iteration, represents the unlabeled sample pool of the tth active learning iteration.

[0060] S11, split the labeled image to obtain small-sized samples that can be used to train the model, forming a labeled sample pool; for a pair of labeled images I and corresponding annotations L, divide the image and samples into several non-overlapping small areas with a width of w and a height of h in a unified way These small samples constitute the labeled sample pool

[0061] In a specific embodiment, the width and height of the sample are set to w = 128 and h = 128 respectively. For a medical image I, cropping is performed starting from the upper left corner with coordinates (0, 0), and the width is divided every time w = 128 and the height every time h = 128 pixels to obtain the following area:

[0062]

[0063] The superscript [0:127], [0:127] indicates that the pixels come from the image coordinate area with a width of 0 to 127 and a height of 0 to 127. The pixels in each divided area do not appear repeatedly. Similarly, the same division is performed on the label L, resulting in:

[0064]

[0065] For image region R I and marked area RL Pairing, dividing the same pixel areas into a group to obtain a labeled sample pool

[0066]

[0067] S12, split the unlabeled image and obtain multiple small areas with similar characteristics in color and texture through superpixel algorithm. According to the center position of the area and the fixed width w and height h, the image is divided into multiple small samples {U1, U2, ...} of fixed size. All unlabeled small samples constitute the entire unlabeled sample pool.

[0068] In a specific embodiment, the width and height of the sample are set to w = 128 and h = 128 respectively, and the superpixel SLIC algorithm is used to cluster the pixels in the unlabeled medical image U to obtain a set of regions R U = {R1, R2, ...}. Since the regions are of different sizes, they are then clipped. Each region can be considered a connected region. For the connected region R1, the coordinates of the center C(R1) can be calculated:

[0069]

[0070] Among them, |R1| represents the area of ​​region R1, that is, the number of pixels; according to the center coordinates C(R1) and the set width w and height h, the sample can be cropped:

[0071]

[0072] Among them, p i,j is a pixel point in the image U, the subscripts i and j represent the position coordinates, and the sample U1 is composed of rectangular pixels with width w and height h respectively, centered at C(R1).

[0073] S2, initializes the key components in active learning, including building a medical image segmentation model and a similarity assessment model, and setting the annotation budget required in the active learning strategy;

[0074] S21, construct a medical image segmentation model M, which consists of an encoder E that compresses information to extract features and a decoder D that restores scale; in particular, the segmentation model will be updated with active learning iterations, and the superscript t represents the number of active learning iterations, for example, M t represents the medical image segmentation model in the tth active learning iteration;

[0075] In a specific embodiment, UNet is used as the architecture of the medical image segmentation model, such as Figure 2The encoder part consists of four stages, each of which includes three convolutional modules (ConvBlock). The convolutional module uses 3×3 convolution, batch normalization (BN), and linear rectification (ReLU) to extract features. The pooling layer is used between each stage to compress the feature map. The number of convolution channels between each stage increases according to [64, 128, 256, 512]. The convolutional module is shown in the following formula:

[0076] ConvBlock(x)=ReLU(BN(Conv 3×3 (x)))

[0077] The decoder of the model consists of three stages. Each stage uses the same convolutional module. Each stage is connected by a transpose convolution to restore the scale of the compressed feature map. At the same time, the number of channels is gradually reduced from [256, 128, 64]. Fine-grained features are transmitted between symmetrical stages through skip connections and splicing. At the end of the model, there is an output layer that uses 1×1 convolution and softmax classifier to obtain the probability score of each category. The specific formula is as follows:

[0078]

[0079] Among them, the subscripts i and j represent the coordinate position on the image, c is a specific category, is the eigenvalue with spatial position (i, j) and channel dimension c after 1×1 convolution, and e is a natural constant with a value of approximately 2.718.

[0080] In a specific embodiment, the loss function during segmentation model training uses the Dice Similarity Coefficient, which is formulated as follows:

[0081]

[0082] Among them, y lable and y pred They represent the binarized sample annotations and model prediction results respectively, C represents the total number of categories, and the superscript i represents the i-th category and also the i-th channel.

[0083] S22, build similarity evaluation model Model sim ; In particular, the similarity evaluation model will be updated with the active learning iterations, and the superscript t represents the number of active learning iterations, for example represents the similarity evaluation model in the t-th iteration;

[0084] In a specific embodiment, a Siamese Neural Network is used to measure the similarity between two samples, such as Figure 3 The twin neural network consists of two networks with the same structure. The network reuses the encoder part of UNet and adopts a 4-layer convolutional neural network. Each layer consists of three convolution modules (ConvBlock). Features are extracted through 3×3 convolution, batch normalization (Batch Normalization, BN), and linear rectification function (Rectified Linear Unit, ReLU). The weights are shared between the two networks, and the final similarity is measured using the pre-selected similarity distance:

[0085]

[0086] Among them, f1 and f2 represent the two features output by the twin neural network respectively.

[0087] S23, setting a labeling budget Budget, which is used to determine whether the active learning cycle should be stopped; in a specific embodiment, the labeling budget is represented by the number of labeled pixels, and the overall labeling budget is set to 50% of the total number of pixels of all unlabeled samples.

[0088] S3, build an active learning strategy based on information volume and sample diversity, evaluate the samples in the unlabeled pool, screen out the most valuable unlabeled samples, have them annotated by experts and add them to the labeled sample pool, deduct the annotation budget, iterate training, update the medical image segmentation model parameters, iterate the active model until the annotation budget is zero, the active learning iteration is as follows Figure 4 shown.

[0089] S31, using a labeled sample pool In the active learning strategy, the model undergoes t active learning iterations, denoted as M t ,In particular, the model that has not been trained through active learning iterations is denoted as M0;

[0090] S32, use the trained model to infer all samples in the unlabeled pool to obtain a set of sample segmentation probability maps and the segmentation result graph collection

[0091] S33, after obtaining the probability map and segmentation result map of all unlabeled samples, the information query function Query info (·) Evaluate the sample; Query info (·) is realized by calculating the segmentation probability map and the result map, and the unlabeled sample Ui The query is as follows:

[0092] Query info (U i )=Q class (Y i )×Q uncertainy (P i )

[0093] Q class (Y i )=Σ c∈C I c (Y i )

[0094]

[0095] Among them, U i Refers to the i-th unlabeled sample, Y i and P i They are the corresponding segmentation result map and probability map, Q class (Y i ) means Sample U i The category score, C is the total number of categories on the sample segmentation map, x represents the position, I_c(·) is the indicator function, Y i (x) represents the segmentation map Y i The value at pixel location x; Q uncertain y(P i ) means Sample U i The uncertainty score, y represents the position coordinate of the segmentation probability map, P i (y) represents the probability map P i The probability value of the upper position being y.

[0096] S34, calculate the diversity score of samples in the unlabeled pool, query the unlabeled samples U i With the marked pool and unlabeled pool except U i Similarity to other samples; Diversity query function Query div (·) is determined by the unlabeled similarity S U Similarity S with the marked L It consists of two parts, for sample U i The similarity query is as follows:

[0097] Query div (U i )=S U (U i )-S L (U i )

[0098]

[0099]

[0100] Among them, Model sim For the similarity measurement network, input two samples and get their similarity evaluation. Here, all labeled samples and samples except U are evaluated. i The unlabeled samples outside the test are measured and the results are fed into the similarity query function Query div (·), get the sample U i diversity score.

[0101] S35, sum the information score and diversity score of the sample to obtain the sample value score:

[0102] Value(U i )=Query info (U i )+Query div (U i )

[0103] Sort all unlabeled samples in descending order of sample value scores. Select the top K samples. As the samples that need to be labeled in the tth round of active learning iteration.

[0104] S36, K samples Hand it over to experts for labeling, use the labeling cost, and get K labeled samples Update sample pool and mark budget:

[0105]

[0106]

[0107] Budget t+1 =Budget t -Cost t

[0108] The superscript t represents the tth active learning iteration; in a specific embodiment, the cost required for labeling is obtained by counting the number of pixels labeled by the expert during labeling.

[0109] Furthermore, before the labeling budget is exhausted, the active learning strategy cycles in the order of S31 to S37 until the labeling budget reaches 0.

[0110] S4, segmentation is performed using a medical image segmentation model based on active learning.

[0111] During the query phase, this method considers both the sample's inherent information content and diversity, effectively capturing outliers and improving the model's segmentation accuracy for small samples. Furthermore, when constructing a sample pool, while traditional methods often rely on random sampling, this method uses the image's inherent prior information to split the sample, increasing the distribution of similar features within the sample, improving the overall proportion of important samples, and reducing the redundant information that may arise from random partitioning.

Claims

1. A medical image segmentation method based on active learning of superpixels and diversity, characterized by: The following steps are included: S1, split the images in the medical image dataset into small samples of fixed size, and construct a labeled sample pool and an unlabeled sample pool; S2, establishes a medical image segmentation model and a similarity assessment model, and sets the annotation budget required in the active learning strategy; S3: Build an active learning strategy based on information volume and diversity, evaluate the samples in the unlabeled pool in S1, screen out the most valuable unlabeled samples, have them annotated by experts and add them to the labeled sample pool, deduct the annotation budget, iterate the training, update the parameters of the medical image segmentation model, and iterate the active model until the annotation budget reaches zero; S4, using the updated iterative medical image segmentation model to segment samples that lack annotations; The step S3 is specifically performed according to the following steps: S31, using a labeled sample pool In the active learning strategy, the medical image segmentation model M undergoes t active learning iterations, denoted as M t , the model that has not undergone active learning iterative training is recorded as M0; S32, use the trained model to infer all samples in the unlabeled pool to obtain a set of sample segmentation probability maps and the segmentation result graph collection S33, after obtaining the probability map and segmentation result map of all unlabeled samples, the information query function Query info (·) Evaluate the sample; Query info (·) is realized by calculating the segmentation probability map and the result map, and the unlabeled sample U i The query is as follows: Query info (U i )=Q class (Y i )×Q uncertainy (P i ) Q class (Y i )=∑ c∈C I c (Y i ) Among them, U i Refers to the i-th unlabeled sample, Y i and P i They are the corresponding segmentation result map and probability map, Q class (Y i ) means Sample U i The category score, C is the total number of categories on the sample segmentation map, x represents the position, I_c(·) is the indicator function, Y i (x) represents the segmentation map Y i The value at pixel location x; Q uncertainy (P i ) means Sample U i The uncertainty score, y represents the position coordinate of the segmentation probability map, P i (y) represents the probability map P i The probability value of the upper position being y; S34, query the unlabeled sample U according to the feature set of the unlabeled sample i With the marked pool and the unlabeled pool except U i Similarity of other samples; Similarity query function Query sim (·) is determined by the unlabeled similarity S U Similarity S with the marked L It consists of two parts, for sample U i The similarity query is as follows: Query sim (U i )=S U (U i )-S L (U i ) Among them, Model sim For the similarity measurement network, input two samples and get their similarity evaluation. Here, all labeled samples and samples except U are evaluated. i The unlabeled samples outside the test are measured and the results are fed into the similarity query function Query sim (·), get the sample U i Similarity score of S35, sum the information score and diversity score of the sample to obtain the sample value score: Value(U i )=Query info (U i )+Query sim (U i ) Sort all unlabeled samples in descending order of sample value scores and select the top K samples. As samples that need to be labeled in the tth round of active learning iteration; S36, K samples Hand it over to experts for labeling, use the labeling cost, and get K labeled samples Update sample pool and mark budget: Budget t+1 =Budget t -Cost t Wherein, the superscript t represents the tth active learning iteration; Furthermore, before the labeling budget is exhausted, the active learning strategy cycles in the order of S31 to S37 until the labeling budget reaches 0.

2. The active learning medical image segmentation method based on superpixels and diversity according to claim 1, characterized in that: In step S1, the medical image sample pool Includes annotated pools and unlabeled sample pool in, is the labeled sample pool, I i represents the i-th sample with annotation, L i Indicates I i Corresponding annotated diagram; is a set of unlabeled medical image samples, U j represents the jth unlabeled sample.

3. The active learning medical image segmentation method based on superpixels and diversity according to claim 1, characterized in that: The step S1 is specifically as follows: S11, split the labeled image samples to obtain small-size samples for training the model, forming a labeled sample pool; for a pair of labeled images I and corresponding annotations L, divide the image and samples into several non-overlapping small areas with width w and height h in a unified way These small samples constitute the labeled sample pool S12, split the unlabeled image and obtain multiple small areas with similar characteristics in color and texture through superpixel algorithm. According to the center position of the area and the fixed width w and height h, the image is divided into multiple small samples {U1, U2, ...} of fixed size. All unlabeled small samples constitute the entire unlabeled sample pool.

4. The active learning medical image segmentation method based on superpixels and diversity according to claim 1, characterized in that: The step S2 is specifically as follows: S21, construct a medical image segmentation model N, which consists of an encoder E that compresses information to extract features and a decoder D that restores scale. The medical image segmentation model M is used to segment unlabeled medical image samples. Multiple active learning iterations are required to obtain accurate segmentation results. During the active learning iterations, the medical image segmentation model M participates in two main processes. The first process is model training, which takes labeled medical image samples as input and outputs segmentation results. The loss function between the segmentation results and the annotations is calculated, and the model parameters are optimized through backpropagation. The second process is model inference, which takes unlabeled medical image samples as input and obtains segmentation results. S22, build similarity evaluation model Model sim , Model sim The input is two samples, and the output is the sample similarity distance. By calculating the similarity distance between a sample and all other samples, the representative score of this sample to the entire sample pool is calculated, which is used to calculate the diversity of the sample. S23, set the labeling budget. The labeling budget is an important condition for determining whether the active learning iteration is terminated. In each round of active learning iteration, the number of labeled samples will be increased through expert labeling, and a certain labeling budget will be deducted each time. When the labeling budget is consumed to zero, the active learning iteration stops and the final medical image segmentation model is obtained.

5. The active learning medical image segmentation method based on superpixels and diversity according to claim 1, characterized in that: The medical image segmentation model in step S4 is obtained from the segmentation model that is updated after multiple trainings after the annotation budget in step S3 is exhausted.

Images