Weakly supervised medical image segmentation method and system based on learnable threshold mechanism

Abstract

This invention discloses a weakly supervised medical image segmentation method and system based on a learnable threshold mechanism. The method constructs a medical image segmentation model for segmenting the tested medical image. Within this model, the input medical image is processed through a multi-level segmentation network structure, with independent learnable thresholds set for different prediction levels in the network for binarization thresholding. Simultaneously, the invention utilizes eye-tracking data from non-expert observers browsing the medical image as a source of weakly supervised information for model training. An adaptive threshold mechanism and a threshold diversity mechanism loss are constructed, allowing the threshold parameters to be dynamically adjusted according to the prediction probability distribution during training and maintaining reasonable threshold intervals at different levels. This better adapts to the uncertainty of non-expert eye-tracking data, improving the stability and reliability of pseudo-labels.

Description

Technical Field

[0001] This invention belongs to the field of image segmentation technology, specifically relating to a weakly supervised medical image segmentation method and system based on a learnable threshold mechanism. Background Technology

[0002] Medical image segmentation is a fundamental core technology in the field of intelligent medical image analysis. Current mainstream segmentation methods are primarily based on deep learning models, trained pixel-level through fully supervised learning. However, these methods rely on a large amount of high-precision pixel-level labeled data, while medical image annotation typically requires experts with specialized medical knowledge to draw each pixel individually. This process is costly, time-consuming, and difficult to scale, becoming a significant bottleneck restricting the widespread application of medical artificial intelligence systems.

[0003] To reduce annotation costs, weakly supervised medical image segmentation methods have gained increasing attention. Existing weakly supervised methods include image-level label supervision, point annotation supervision, bounding box annotation supervision, and doodle-style annotation supervision. While these methods reduce the workload of fine-grained annotation to some extent, they still require manual active annotation. Furthermore, due to the lack of complete spatial structural information, they often suffer from problems such as blurred boundaries, unstable pseudo-label quality, and strong dependence on post-processing. In addition, when non-professionals participate in annotation, these traditional weakly supervised methods struggle to guarantee accuracy, further limiting their application among a large number of general users.

[0004] With the development of eye-tracking technology, utilizing human visual attention information as a weak supervisory signal has become a new technological approach. Eye-tracking data can objectively record the gaze area and visual search trajectory of an observer during image browsing, obtaining information on potential target areas without explicit manual annotation. Compared with traditional manual annotation methods, it has advantages such as strong naturalness, high acquisition efficiency, and low intervention.

[0005] In existing technologies, some methods have attempted to utilize expert eye-tracking data to assist in medical image segmentation training. However, these methods still primarily rely on medical experts for data collection, which suffers from high collection costs and limited sample size.

[0006] In contrast, non-expert observers rely primarily on basic visual mechanisms for target localization when browsing medical images, such as regional contrast differences, structural edge features, and visual saliency. Non-expert eye-tracking data possesses stronger universal cognitive characteristics and has the potential for large-scale collection, providing a low-cost source of attentional information for weakly supervised learning. However, non-expert eye-tracking data typically exhibits characteristics such as dispersed fixation areas, significant individual differences, and large fluctuations in confidence levels, leading to high uncertainty and noise interference in the generated eye-tracking heatmaps.

[0007] Existing weakly supervised methods based on eye-tracking data typically employ a fixed threshold to binarize eye-tracking heatmaps to generate pseudo-labels. However, this fixed threshold strategy struggles to adapt to differences in confidence distribution among different samples, observers, and network layers. Setting the threshold too high can lead to underactivation of the target region; setting it too low can introduce a large amount of background noise, affecting the training stability and generalization performance of the segmentation model. Furthermore, in multi-layered network structures, the confidence distribution of feature maps at different layers varies significantly, making it difficult to simultaneously address the needs of multi-scale feature representation with a uniform fixed threshold.

[0008] Therefore, when using non-expert eye-tracking data for weakly supervised medical image segmentation, how to construct a threshold mechanism that can dynamically adjust according to the data distribution to improve the quality of pseudo-labels, enhance the stability of the supervision signal, and adapt to the differences in features of multi-level networks has become a key technical problem that urgently needs to be solved in the current field of technology. Summary of the Invention

[0009] The purpose of this invention is to provide a weakly supervised medical image segmentation method and system based on a learnable threshold mechanism.

[0010] In a first aspect, the present invention provides a weakly supervised medical image segmentation method based on a learnable threshold mechanism, the method comprising:

[0011] Construct a dataset containing different medical images and preprocess the medical images in the dataset;

[0012] A medical image segmentation model is constructed, comprising a multi-scale feature extraction module, a threshold learning module, and a fusion module. The multi-scale feature extraction module processes the input medical image to generate feature extraction results at different scales. The threshold learning module includes multiple threshold learning layers, with the number of threshold learning layers being the same as the number of feature extraction results. Learnable thresholds are introduced in each threshold learning layer for segmentation based on the feature extraction results. The fusion module fuses the output feature maps of multiple threshold learning layers to obtain the segmentation result output by the medical image segmentation model.

[0013] A medical image segmentation model is trained using a dataset; during training, a total loss is constructed to guide model updates; the total loss includes a supervised loss and an adaptive threshold mechanism loss; the adaptive threshold mechanism loss dynamically divides the positive and negative regions of pseudo-labels based on the prediction confidence.

[0014] The medical image to be tested is preprocessed and then input into the trained medical image segmentation model to obtain the target segmentation result of the medical image to be tested.

[0015] Preferably, the method for constructing the adaptive threshold mechanism loss is as follows:

[0016] Based on the true probabilities of different pixels on pseudo-labels, different pixels are divided into positive and negative subsets. Based on the confidence of the positive and negative subsets, an adaptive threshold mechanism loss with boundary m is constructed. for:

[0017]

[0018] in, For the first Learnable thresholds introduced in the layer threshold learning layer; and These are the confidence scores for the positive and negative subsets, respectively. Indicates the non-negative part.

[0019] Preferably, the confidence levels of the positive and negative subsets are the average predicted probabilities of all pixels in the corresponding subsets.

[0020] Preferably, the total loss also includes a threshold diversification mechanism loss; the threshold diversification mechanism loss is used to maintain the difference in thresholds between levels.

[0021] Preferably, the threshold diversification mechanism loss The construction method is as follows:

[0022]

[0023] in, The threshold learning layer number is m; m is the minimum margin value. and The learnable thresholds for different threshold learning layers.

[0024] Preferably, the multi-scale feature extraction module uses a multi-level U-Net network, with each layer independently outputting a predicted probability map, which is then input into the corresponding threshold learning layer.

[0025] Preferably, the supervision loss adopts the cross-entropy loss function.

[0026] Preferably, the dataset uses eye-tracking data as pseudo-labels for medical images; before using the dataset for model training, the eye-tracking data is refined using a fully connected conditional random field, and the refined eye-tracking data is thresholded to generate binary pseudo-labels, which serve as supervision signals for weakly supervised learning.

[0027] Preferably, the preprocessing procedure is as follows: all medical images are resized and data augmentation is performed using translation, random flipping, and affine transformation.

[0028] Secondly, this invention provides a weakly supervised medical image segmentation system based on a learnable threshold mechanism, which is used to execute the aforementioned weakly supervised medical image segmentation method. The weakly supervised medical image segmentation system includes a data acquisition module, a preprocessing module, a model building module, a model training module, and a real-time segmentation module. The data acquisition module is used to acquire a medical image dataset. The preprocessing module is used to resize and augment the medical image dataset. The model building module is used to construct a medical image segmentation model and transmit the data information from the medical image segmentation model to the model training module. The model training module is used to train the medical image segmentation model. The real-time segmentation module is used to perform real-time segmentation of the tested medical image using the trained medical image segmentation model.

[0029] The beneficial effects of this invention are:

[0030] 1. This invention addresses the problem that non-expert eye-tracking data often suffers from scattered attention distribution, significant individual differences, and high noise levels, leading to unstable pseudo-label quality when directly used for weakly supervised training. This is achieved by introducing a learnable threshold mechanism at different levels and dynamically adjusting the threshold parameters based on the predicted probability distribution during training. This better adapts to the uncertainty of non-expert eye-tracking data, improving the stability and reliability of pseudo-labels. Furthermore, by constructing an adaptive threshold optimization mechanism that considers the differences in confidence distribution among different samples, the invention allows the threshold to be dynamically adjusted based on the statistical characteristics of the predicted probability of positive and negative regions. This effectively avoids overactivation or underactivation problems caused by a fixed threshold, thus improving the quality of pseudo-label generation.

[0031] 2. In this invention, independent learnable thresholds are set for different prediction levels in a multi-level segmentation network structure, and a threshold difference constraint mechanism is introduced to enable different levels to maintain a reasonable threshold interval, thereby better adapting to the probability distribution differences of features at different scales and enhancing the model's ability to identify complex structural boundaries and multi-scale target regions.

[0032] 3. This invention can utilize eye-tracking data from non-expert observers browsing medical images as a source of weakly supervised information. It can obtain attention information of the target region without requiring medical experts to perform pixel-level fine annotation. Compared with traditional fully supervised segmentation methods that rely on a large amount of manually labeled data, this invention can significantly reduce data acquisition costs and improve data acquisition efficiency, providing a low-cost solution for large-scale data construction for medical image segmentation tasks. Attached Figure Description

[0033] Figure 1 This is a schematic diagram illustrating the acquisition of eye-tracking data in this invention.

[0034] Figure 2This is a schematic diagram of the binary pseudo-label in this invention.

[0035] Figure 3 This is a schematic diagram of the medical image segmentation model in this invention.

[0036] Figure 4 This is a schematic diagram comparing the medical image segmentation results of the present invention with those of the classic weakly supervised method.

[0037] Figure 5 This diagram illustrates the performance comparison of models trained using eye-tracking data from professional and non-professional observers before and after the introduction of a learnable threshold. Detailed Implementation

[0038] The present invention will be further described below with reference to the accompanying drawings.

[0039] A weakly supervised medical image segmentation method based on a learnable threshold mechanism is proposed. The weakly supervised medical image segmentation system includes a data acquisition module, a preprocessing module, a model building module, a model training module, and a real-time segmentation module. The data acquisition module includes a medical image acquisition module and an eye-tracking data acquisition module. The medical image acquisition module acquires medical images; the eye-tracking data acquisition module acquires eye-tracking data from non-professional observers. The preprocessing module resizes and augments the medical image dataset. The model building module constructs a medical image segmentation model and transfers the data from the model to the model training module. The model training module trains the medical image segmentation model. The real-time segmentation module uses the trained medical image segmentation model to perform real-time segmentation of the tested medical image.

[0040] In this embodiment, the eye-tracking data acquisition module adopts one of a desktop eye tracker, a head-mounted eye-tracking device, and a camera-based visual estimation eye-tracking system.

[0041] This weakly supervised medical image segmentation method includes the following steps:

[0042] Step 1: Building the dataset

[0043] This embodiment uses medical images from the publicly available Kvasir-SEG dataset to construct the dataset. The original Kvasir-SEG dataset contains 1000 medical images and high-precision pixel-level masks annotated by medical experts. All images are uniformly scaled to a resolution of 224×224, and translation, random flipping, and affine transformations are used as data augmentation techniques to improve the model's generalization ability. The processed dataset is then split into 900 training images and 100 test images.

[0044] In this embodiment, the medical image is an endoscopic polyp image.

[0045] Step 2: Collect eye-tracking data

[0046] like Figure 1 As shown, an eye-tracking signal acquisition device was placed below the computer screen to collect the observer's annotation information. The sampling rate was 60Hz. A hardware-software hybrid approach was used, based on the Python programming language and the PyQt5 graphical user interface framework, combined with an eye-tracking software development kit (SDK), to build an eye-tracking observation data acquisition system for posture calibration, real-time acquisition, visualization, and storage of eye-tracking data. This provided high-quality and structured input for subsequent attention modeling and data analysis.

[0047] Three non-professional observers (without training in eye tracking or image annotation) were randomly selected to participate. Before formal data collection, each observer underwent a rigorous gaze calibration process to ensure spatial accuracy and data robustness during the acquisition phase. During data collection, individual target images were presented to the observers sequentially. Participants were instructed to focus on the target object in the image; after quickly locating the object through gaze, they performed a purposeful and comprehensive scan: typically starting by observing along the object's boundary contours to form a spatial concept, then gradually moving towards the central area for detailed examination. Throughout the scan, the observer's gaze coordinates were captured in real time and automatically mapped to their relative screen position. The average sampling time for each target image was approximately 10 to 12 seconds. After collecting 100 images, participants rested for 1 to 2 minutes to prevent visual fatigue from affecting sampling quality. After the rest period, they needed to readjust their sampling posture before continuing further data collection.

[0048] Each observer's gaze coordinates were mapped to an image size (224×224), and an initial gaze heatmap was generated using Gaussian kernel smoothing. The heatmaps of the three observers were averaged to obtain the fused gaze heatmap, thus reducing individual variability.

[0049] Step 3: Data Processing

[0050] To further improve the semantic accuracy of the gaze heatmap, a fully connected conditional random field (Dense CRF) is used to refine it. Guided by the original image, the CRF refines the edges of the gaze heatmap based on the color and spatial similarity between pixels, making the gaze response more closely match the target boundary. The refined gaze heatmap is then thresholded (fixed or adaptive threshold) to generate binary pseudo-labels, which serve as supervisory signals for weakly supervised segmentation. Figure 2 As shown.

[0051] In some embodiments, attention information generated by mouse trajectories, click behavior, or visual saliency prediction models is used as a supervision signal.

[0052] Step 4: Construct a medical image segmentation model

[0053] like Figure 3 As shown, the medical image segmentation model includes a multi-scale feature extraction module, a threshold learning module, and a fusion module. The multi-scale feature extraction module performs multi-scale feature extraction and segmentation prediction on the input medical image, obtaining feature extraction results at different scales. The threshold learning module includes multiple threshold learning layers. A learnable and flexible threshold mechanism is introduced in the threshold learning layers, assigning independent learnable thresholds to each layer and binarizing the predicted probabilities based on these thresholds. The threshold learning layers dynamically divide the positive and negative regions of the pseudo-labels according to the prediction confidence and maintain the difference in thresholds between layers through diversity constraints. The fusion module fuses the output feature maps of multiple threshold learning layers to obtain the segmentation result output by the medical image segmentation model.

[0054] In the threshold learning module, each threshold learning layer A learnable threshold is assigned. (Initialized to 0.5), the first The predicted probability at pixel k in the threshold learning layer for:

[0055]

[0056] in, For the first The input features of the threshold learning layer; σ This represents the sigmoid activation function.

[0057] In this embodiment, the multi-scale feature extraction module adopts a multi-level U-Net network, with each layer independently outputting a predicted probability map, and constrained within the range of [0,1] by a threshold learning layer.

[0058] In some embodiments, the multi-scale feature extraction module employs one of the following: a segmentation model based on a convolutional neural network, a segmentation model based on an attention mechanism, and a visual segmentation model based on a Transformer structure.

[0059] Step 5: Model Training

[0060] A medical image segmentation model was trained using a dataset, and a total loss was constructed to guide model parameter updates. Since gaze behavior is influenced by inter-individual differences, expertise levels, and the complexity of visual content, a fixed threshold cannot handle dynamic changes between feature layers or samples, thus limiting the model's robustness and generalization ability. To address this issue, a total loss was constructed based on an adaptive thresholding mechanism and a threshold diversity mechanism. The former ensures that the threshold receives reliable gradients and is adaptively updated. The latter ensures that thresholds at different levels maintain a certain degree of diversity. Total Loss It is expressed as follows:

[0061]

[0062] in, For adaptive threshold mechanism loss; For threshold diversity mechanism loss; To monitor losses; and This is a hyperparameter.

[0063] In this embodiment, the supervision loss uses the cross-entropy loss function. During training, a total of 12,000 training iterations are performed, with a batch size of 8. The optimizer used is SGD (Stochastic Gradient Descent); the learning rate scheduling employs a cosine annealing (Cosine AnnealingLR) strategy, with an initial learning rate of 1e⁻² (0.01) and a minimum learning rate of 2e⁻². 4 (0.0002).

[0064] The methods for constructing the adaptive threshold mechanism loss and the threshold diversity mechanism loss are as follows:

[0065] (1) Adaptive threshold mechanism loss

[0066] An adaptive thresholding mechanism allows each level to adjust its confidence criterion accordingly. By analyzing the model's predicted probabilities on pseudo-labels, pixels are divided into positive and negative classes, and the threshold is optimized based on the average confidence of the two classes. Pseudo-labels are used. Divide the pixels into positive and negative subsets:

[0067]

[0068]

[0069] in, For positive subsets; It is a negative subset.

[0070] Calculate the average predicted probability within each subset to capture the confidence distribution of pseudo-labels as follows:

[0071]

[0072]

[0073] in, The average predicted probability of the positive subset; The average predicted probability of the negative subset.

[0074] To encourage threshold To effectively distinguish between positive and negative regions, an adaptive threshold mechanism loss with boundary m is constructed. for:

[0075]

[0076] in, Indicates the non-negative part. .

[0077] The parameter m acts as a buffer, helping to stabilize the threshold adaptation process. Intuitively, this loss penalizes thresholds that are too close to the average confidence of any pseudo-labeled pixels of the positive class (µ + i) or negative class (µ - i). Specifically, it ensures a learnable threshold t. i A minimum spacing of at least m is maintained between the predicted probabilities and the average predicted probabilities of the positive and negative pixel sets to prevent premature convergence and promote stable threshold optimization. During training, the predicted probabilities... Treating it as a constant ensures that the gradient only flows toward the learnable threshold. This ensures that threshold learning remains stable independently of network parameter updates, without affecting the parameters of the backbone network.

[0078] (2) Threshold diversification mechanism loss

[0079] The threshold diversification mechanism loss is used to reasonably divide the thresholds across different levels, thereby enhancing the complementarity and expressiveness of multi-level pseudo-labels. The threshold diversification mechanism loss is expressed as:

[0080]

[0081] Where m is the minimum interval value, used to specify the required separation distance between thresholds.

[0082] The threshold diversification mechanism loss penalizes thresholds that are too close to each other, thereby encouraging thresholds in each layer to maintain a certain distance, which can enhance the complementarity and expressive power of multi-layer pseudo-labels.

[0083] Step 5: After preprocessing the medical image to be tested, input it into the trained medical image segmentation model to perform image segmentation and obtain the target segmentation result of the medical image to be tested.

[0084] Step Six: Evaluation Indicators and Experimental Results Analysis

[0085] This invention uses mDice (mean Dice coefficient) as the evaluation metric. The Dice coefficient is one of the most commonly used evaluation metrics in medical image segmentation tasks, used to quantify the similarity between the model's predicted segmentation results and the ground truth annotations. Its value ranges from [0,1]. The Dice coefficient is essentially a measure of the overlapping region between two sets, equivalent to a pixel-level version of the F1 score. The specific formula is as follows:

[0086]

[0087] in, Represents the predicted segmentation region (binary mask, set of foreground pixels); This represents the labeled segmented region (Ground Truth, the set of foreground pixels labeled by experts). Indicates the correct number of foreground pixels (intersection); and These represent the total number of pixels in the predicted foreground and the actual foreground, respectively.

[0088] The present invention was compared with the classical weak supervision method, and the comparison results are as follows. Figure 4As shown, Vanilla is a fully supervised control, BoxTeacher (Box), BoxInst (Box), AGMM (Point), PointSup (Point), and AGMM (Scribble) are other weakly supervised methods, GazeMedSeg is a framework that removes the expert-labeled data after learning threshold adjustment, and Ours (Expert) and Ours (Non-expert) represent the different performances of the medical image segmentation model of this invention on expert and non-expert datasets, respectively. Without the learning threshold strategy, the mDice value for non-experts is 0.7472, which is 4.1% lower than the expert-viewing performance of GazeMedSeg (0.7780). However, after introducing the learning threshold strategy, mDice improved to Ours (Non-expert) 0.7742, narrowing the gap with Ours (Expert) under expert-viewing supervision to only 1.9%, and improving performance by 3.6% and 1.4% on the two datasets, respectively. In terms of annotation efficiency, this invention requires only 3.0 hours—just 0.8 hours more than the 2.2 hours required to collect expert gaze data—demonstrating that non-expert gaze annotation can achieve considerably high efficiency under the medical segmentation model constructed in this invention. Furthermore, the weakly supervised strategy based on non-expert gaze points outperforms other sparse annotation methods in terms of segmentation accuracy while maintaining a comparable annotation time. Notably, this invention achieves 94% of the performance of a fully supervised (Vanilla) model, highlighting its effectiveness and practicality in scenarios with limited expert resources.

[0089] Figure 5 This demonstrates how the segmentation accuracy of ordinary observers and expert annotators changes with increasing dataset size before and after removing the learnable thresholding mechanism. Figure 5 It can be seen that when the number of learnable samples is small, the segmentation accuracy of the learnable threshold strategy (ours) is slightly lower than that of not using the learnable segmentation strategy (GazeMedSeg). However, with the use of more data, the accuracy of the learnable threshold strategy (ours) gradually surpasses that of not using the learnable segmentation strategy (GazeMedSeg). This is because as the amount of training data increases, the model is able to gradually learn a better segmentation threshold.

Claims

1. A weakly supervised medical image segmentation method based on a learnable threshold mechanism, characterized in that: The method includes: Construct a dataset containing different medical images and preprocess the medical images in the dataset; A medical image segmentation model is constructed, comprising a multi-scale feature extraction module, a threshold learning module, and a fusion module. The multi-scale feature extraction module processes the input medical image to generate feature extraction results at different scales. The threshold learning module includes multiple threshold learning layers, with the number of threshold learning layers being the same as the number of feature extraction results. Learnable thresholds are introduced in each threshold learning layer for segmentation based on the feature extraction results. The fusion module fuses the output feature maps of multiple threshold learning layers to obtain the segmentation result output by the medical image segmentation model. A medical image segmentation model is trained using a dataset; during training, a total loss is constructed to guide model updates; the total loss includes a supervised loss and an adaptive threshold mechanism loss; the adaptive threshold mechanism loss dynamically divides the positive and negative regions of pseudo-labels based on the prediction confidence. The medical image to be tested is preprocessed and then input into the trained medical image segmentation model to obtain the target segmentation result of the medical image to be tested.

2. The weakly supervised medical image segmentation method based on a learnable threshold mechanism according to claim 1, characterized in that: The method for constructing the adaptive threshold mechanism loss is as follows: Based on the true probabilities of different pixels on pseudo-labels, different pixels are divided into positive and negative subsets. Based on the confidence of the positive and negative subsets, an adaptive threshold mechanism loss with boundary m is constructed. for: in, For the first Learnable thresholds introduced in the layer threshold learning layer; and These are the confidence scores for the positive and negative subsets, respectively. Indicates the non-negative part.

3. The weakly supervised medical image segmentation method based on a learnable threshold mechanism according to claim 2, characterized in that: The confidence scores of the positive and negative subsets are the average predicted probabilities of all pixels in the corresponding subsets.

4. The weakly supervised medical image segmentation method based on a learnable threshold mechanism according to claim 1, characterized in that: The total loss also includes a threshold diversification mechanism loss, which is used to maintain the differences in thresholds between levels.

5. The weakly supervised medical image segmentation method based on a learnable threshold mechanism according to claim 4, characterized in that: The threshold diversification mechanism loss The construction method is as follows: in, The threshold learning layer number is m; m is the minimum margin value. and The learnable thresholds for different threshold learning layers.

6. The weakly supervised medical image segmentation method based on a learnable threshold mechanism according to claim 1, characterized in that: The multi-scale feature extraction module uses a multi-level U-Net network, with each layer independently outputting a predicted probability map, which is then input into the corresponding threshold learning layer.

7. The weakly supervised medical image segmentation method based on a learnable threshold mechanism according to claim 1, characterized in that: The supervision loss uses the cross-entropy loss function.

8. The weakly supervised medical image segmentation method based on a learnable threshold mechanism according to claim 1, characterized in that: The dataset uses eye-tracking data as pseudo-labels for medical images; Before using the dataset for model training, the eye-tracking data is refined using a fully connected conditional random field. The refined eye-tracking data is then thresholded to generate binary pseudo-labels, which serve as supervision signals for weakly supervised learning.

9. The weakly supervised medical image segmentation method based on a learnable threshold mechanism according to claim 1, characterized in that: The preprocessing process involves unifying the size of all medical images and performing data augmentation using translation, random flipping, and affine transformation.

10. A weakly supervised medical image segmentation system based on a learnable threshold mechanism, characterized in that: This system is used to execute a weakly supervised medical image segmentation method based on a learnable threshold mechanism as described in claim 1. The weakly supervised medical image segmentation system includes a data acquisition module, a preprocessing module, a model building module, a model training module, and a real-time segmentation module. The data acquisition module acquires a medical image dataset. The preprocessing module resizes and augments the medical image dataset. The model building module constructs a medical image segmentation model and transmits the data information from the medical image segmentation model to the model training module. The model training module trains the medical image segmentation model. The real-time segmentation module uses the trained medical image segmentation model to perform real-time segmentation of the tested medical image.

Images