Online learning based sketch model construction method and related device
By generating multiple adaptive delineation models through online learning, the problems of poor consistency and long processing time in target delineation results are solved, achieving efficient and accurate target delineation that adapts to clinical guidelines of different hospitals and supports version traceability and visualization assistance.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- MEVION MEDICAL EQUIPMENT CO LTD
- Filing Date
- 2025-11-17
- Publication Date
- 2026-07-21
AI Technical Summary
Existing target delineation techniques suffer from poor consistency in delineation results, long processing times, reliance on manual operation, and difficulty in adapting to clinical guidelines followed by different hospitals, especially in complex cases where generalization ability is insufficient.
By using an online learning-based approach, multiple adaptive delineation models are generated, supporting adaptation to multiple target region delineation standards, providing multiple candidate segmentation results, and updating model parameters online through user interaction, thus achieving traceability of delineation versions and visual assistance.
It improves the compatibility, accuracy and efficiency of target delineation, reduces human bias, supports flexible selection and comparison of different standards, reduces fatigue error, and achieves low-cost long-term performance maintenance and personalized adaptation.
Smart Images
Figure CN121505263B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of radiotherapy technology, and in particular to a method and related equipment for constructing delineation models based on online learning. Background Technology
[0002] In clinical radiotherapy practice, accurate target delineation is crucial for ensuring treatment effectiveness and minimizing damage to normal tissues. However, current target delineation procedures and techniques still face numerous challenges:
[0003] There are various delineation guidelines and standards both domestically and internationally. These standards differ in details such as the definition of target boundaries and the extent of clinical target expansion, resulting in poor consistency in delineation results for the same case among different medical institutions and even different doctors. This poses difficulties for multi-center clinical research and treatment standardization.
[0004] Currently, the mainstream delineation method still heavily relies on doctors manually drawing layers on medical images, especially for tumors with complex anatomical structures such as the head and neck, and pelvis, where the entire process can take several hours. Prolonged work can easily lead to doctor fatigue, introducing significant subjective judgment bias and affecting the stability and repeatability of the delineation quality.
[0005] Although AI-based automated segmentation techniques have been applied, most existing systems are trained on single, fixed delineation standard datasets, making it difficult to flexibly adapt to the different clinical guidelines followed by different hospitals. Furthermore, when encountering complex cases with postoperative anatomical changes, abnormal tumor morphology, or other accompanying diseases, the models' generalization ability is insufficient, and segmentation accuracy drops significantly. Summary of the Invention
[0006] To address the aforementioned issues, the present invention aims to provide a method and related equipment for constructing a delineation model based on online learning. This method supports adaptation to multiple target delineation standards, generates multiple candidate segmentation results, and updates model parameters online through user interaction. It also enables traceability of delineation versions and provides visual assistance, thereby improving the compatibility, accuracy, and efficiency of target delineation.
[0007] In a first aspect, the present invention provides a method for constructing a delineation model based on online learning, the method comprising:
[0008] The pre-trained base segmentation model is tuned on different delineation standard datasets to generate multiple adaptive models, each of which corresponds to a delineation standard.
[0009] Based on multiple adaptive models, the target medical image is segmented to generate multiple candidate target region segmentation results corresponding to various delineation standards.
[0010] Based on the user's selection and / or correction of the candidate target region segmentation results, a target segmentation model is determined, and the target segmentation model is trained online to update its parameters; wherein, the target segmentation model is one selected from the plurality of adaptive models, or a newly created independent model with the same structure as the base segmentation model.
[0011] Preferably, the basic segmentation model is a cue-based segmentation model, which includes an image encoder, a cue encoder, and a mask decoder.
[0012] Preferably, the basic segmentation model is the MedSAM2 model or a variant of the MedSAM2 model.
[0013] Preferably, generating multiple adaptive models includes:
[0014] The basic segmentation model was trained using different delineation standard datasets to optimize all its parameters and obtain an initial optimized model.
[0015] The parameters of the image encoder of the initial optimization model are fixed, and the parameters of its decoding-related units are adjusted using different delineation standard datasets to obtain multiple delineation standard-specific models; the decoding-related units include memory attention units and mask decoders.
[0016] Preferably, the segmentation process of the target medical image based on multiple adaptive models includes:
[0017] The target medical image is subjected to feature extraction based on a shared image encoder;
[0018] The extracted image features are input into decoders of multiple different delineation standard dedicated models for processing to generate the multiple candidate target region segmentation results.
[0019] Preferably, the segmentation process of the target medical image based on multiple adaptive models includes:
[0020] Acquire the target medical image data and parse its format;
[0021] The target medical image is processed using multiple segmentation models configured with different delineation standards to generate the segmentation results of the multiple candidate target regions.
[0022] Preferably, during the segmentation process, different candidate target region segmentation results are generated by providing different prompts to the prompt-based segmentation model; wherein the prompts include at least one of points, bounding boxes, or text descriptions.
[0023] Preferably, the method further includes:
[0024] When displaying the segmentation results of the multiple candidate target regions, the delineation criteria on which each result is based are automatically labeled;
[0025] And / or, calculate the difference between each result and the corresponding standard template;
[0026] And / or, the segmentation results of the multiple candidate target regions are visualized and overlaid.
[0027] Preferably, the method further includes:
[0028] It stores multiple versions of data during the target delineation process and supports switching and comparing multiple versions based on delineation standards, timestamps, or user identifiers.
[0029] Preferably, the method further includes: the user's correction operation on the candidate target region segmentation result includes at least one of adding regions, deleting regions, or modifying the contour boundary of the target region contour.
[0030] Secondly, the present invention provides a delineation model construction system based on online learning, the system comprising:
[0031] The model generation module is configured to adjust the pre-trained base segmentation model on different delineation standard datasets to generate multiple adaptive models, each of which corresponds to a delineation standard.
[0032] The segmentation processing module is configured to segment the target medical image based on multiple adaptive models, generating multiple candidate target region segmentation results corresponding to multiple delineation criteria.
[0033] The online learning module is configured to determine a target segmentation model based on the user's selection and / or correction operations on the candidate target region segmentation results, and to perform online learning on the target segmentation model to update its parameters; wherein the target segmentation model is one selected from the plurality of adaptive models, or a newly created independent model with the same structure as the base segmentation model.
[0034] Thirdly, the present invention provides a radiotherapy system, including a processing unit and a memory, wherein the memory stores a computer program, and when the computer program is executed by the processing unit, it implements the steps of any method of the embodiments of the present invention.
[0035] Fourthly, the present invention provides an electronic device, the electronic device including a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the steps of any of the methods of the present invention or the functions of the system of the present invention.
[0036] Fifthly, the present invention provides a computer-readable storage medium storing computer instructions, wherein when a computer reads the computer instructions, the computer executes the steps of any of the methods described in the present invention.
[0037] Compared with the prior art, the beneficial effects of the present invention include at least the following: by pre-setting multiple dedicated models corresponding to different international / domestic guidelines, it is possible to generate candidate results that meet different standards with one click, which not only unifies the delineation specifications under specific standards and reduces the deviation of human interpretation, but also gives doctors the ability to flexibly select and compare different standards, fundamentally solving the problem of delineation consistency caused by the diversity of standards and inconsistent implementation.
[0038] The parallel generation of multiple high-quality initial candidate target areas frees doctors from the tedious manual delineation layer by layer, transforming them into efficient reviewers and selectors. This reduces delineation time from hours to minutes, significantly improving clinical work efficiency and reducing errors caused by fatigue.
[0039] By introducing prompt-based interactive segmentation and online learning mechanisms, doctors can accurately and quickly correct unsatisfactory results with simple clicks (dots, boxes), and the system can absorb these corrections in real time and dynamically optimize the model. This allows the system to continuously adapt to complex anatomical variations, special cases, and the personalized sketching habits of specific hospitals or doctors, becoming smarter with use and possessing extremely high practical value.
[0040] By fully recording the entire process of data from initial candidates, user selection, interactive corrections to the final version, it supports multi-dimensional version comparison and backtracking based on standards, time, and users. This not only provides a solid data foundation for treatment quality control and auditing, but also provides valuable information for clinical teaching, training, and scientific research analysis.
[0041] The online learning mechanism enables the model to be continuously fine-tuned using correction data generated in daily clinical work, without the need for frequent, costly, and time-consuming centralized offline retraining. This allows the evolution of the AI model to be seamlessly integrated with the clinical workflow, achieving low-cost long-term performance maintenance and personalized adaptation. Attached Figure Description
[0042] Figure 1 This is a schematic diagram of the online learning-based delineation model construction method according to an embodiment of the present invention;
[0043] Figure 2 This is a schematic diagram of the online learning-based delineation model construction process according to an embodiment of the present invention. Detailed Implementation
[0044] Exemplary embodiments will now be described more fully with reference to the accompanying drawings. However, these exemplary embodiments can be implemented in many forms and should not be construed as limited to the embodiments set forth herein; rather, they are provided to make the invention more comprehensive and complete, and to fully convey the concept of the exemplary embodiments to those skilled in the art. The same reference numerals in the drawings denote the same or similar structures, and therefore repeated descriptions of them will be omitted.
[0045] The terms used to express position and direction in this invention are illustrated with reference to the accompanying drawings, but changes can be made as needed, and all such changes are included within the scope of protection of this invention.
[0046] Example 1, refer to Appendix Figure 1 and attached Figure 2 This embodiment provides a method for constructing a delineation model based on online learning, the method including:
[0047] The pre-trained base segmentation model was adjusted on different delineation standard datasets to generate multiple adaptive models, each of which corresponds to a delineation standard.
[0048] Based on multiple adaptive models, the target medical image is segmented to generate multiple candidate target region segmentation results corresponding to various delineation standards.
[0049] Based on the user's selection and / or correction of the candidate target region segmentation results, a target segmentation model is determined, and the target segmentation model is trained online to update its parameters; wherein, the target segmentation model is one selected from the plurality of adaptive models, or a newly created independent model with the same structure as the base segmentation model.
[0050] In one possible implementation, the underlying segmentation model is a cue-based segmentation model, which includes an image encoder, a cue encoder, and a mask decoder.
[0051] In one possible implementation, the underlying segmentation model is the MedSAM2 model or a variant of the MedSAM2 model.
[0052] In one possible implementation, the multiple delineation standards include at least one of the following: domestic expert consensus, international RTOG standards, ESTRO standards, NCCN standards, or CSCO guidelines.
[0053] The working principle and effects of the above technical solution are as follows:
[0054] A robust base segmentation model is used as the foundation, such as the MedSAM2 model or its variants. Using delineation standard datasets collected from various international and domestic guidelines, this base model is fine-tuned to generate a series of adaptive models, each internalizing a specific delineation standard. When a new target medical image is input, all these adaptive models are invoked in parallel, instantly generating multiple candidate target region segmentation results that conform to different standards. By pre-setting models to solidify different delineation standards, the inconsistency problem caused by different doctors' interpretations of the same standard is fundamentally solved, ensuring the standardization and comparability of treatment plans.
[0055] Doctors are no longer the starting point for manually sketching out each layer; instead, they become reviewers and decision-makers of the results. They can intuitively compare multiple candidate outcomes and select the one that best meets the current clinical needs as the basis. For the selected outcome, if there are minor deviations, doctors do not need complex operations; they can simply make corrections through simple interactive prompts (such as clicking on the foreground or background, or drawing a box). The system's built-in prompt-based segmentation capability can understand the doctor's intentions in real time and instantly generate accurate corrected results.
[0056] Each selection and correction by the doctor is used to fine-tune the target segmentation model (i.e., the base model chosen by the doctor or a newly created independent model) online, dynamically updating its model parameters. This allows the model to quickly learn and adapt to complex anatomical variations, special cases, and the delineation preferences and habits of specific hospitals or individual doctors. The system fully records the entire process data from initial candidate generation, user selection, each step of interactive correction to the final draft, forming a traceable version chain.
[0057] In one possible implementation, the basic segmentation model is capable of receiving and processing multiple modal images from CT, MRI, and PET, and integrating multimodal information through its internal feature fusion mechanism.
[0058] It can simultaneously receive imaging data from multiple modalities, such as CT, MRI, and PET, from the same patient.
[0059] Before processing the model, image registration technology is used to precisely align these images of different modalities in space to ensure that the same anatomical location corresponds on different images.
[0060] The image encoder of the basic segmentation model (such as the enhanced MedSAM2 or its variants) is designed to process inputs of different modalities in parallel or collaboratively.
[0061] The model extracts unique features from images of different modalities:
[0062] CT: Provides precise information on anatomical structures (such as bones and lung cavities) and electron density (for dose calculation), and outlines clear contours.
[0063] MRI provides excellent soft tissue contrast (such as the boundary between the tumor and the surrounding normal tissue, and neurovascular bundles), which helps to distinguish structures that are difficult to distinguish on CT.
[0064] PET: Provides functional metabolic information, highlighting highly metabolically active tumor areas and helping to distinguish the active parts of the tumor from necrotic tissue.
[0065] The model's internal feature fusion mechanism (such as through attention mechanisms, feature splicing, or convolution) intelligently integrates the feature maps from CT, MRI, and PET. Instead, it allows the model to autonomously learn how to weigh and combine the value of different modal information to generate a more comprehensive and powerful super feature map.
[0066] The fused super feature map, along with user prompts (such as clicks or selections), is input into the mask decoder. The decoder decodes based on this more information-rich feature map, and the final segmentation result (candidate target area) is no longer a guess based on a single image, but rather a comprehensive piece of evidence integrating anatomy, structure, and functional metabolism, resulting in a qualitative improvement in accuracy and reliability.
[0067] The subsequent online learning process is also based on multimodal data. Doctors correct the segmentation results on a view that incorporates multimodal information; during fine-tuning, the model learns how to adjust its internal feature fusion strategy based on these corrections, so that it can better integrate multimodal information to meet the doctor's clinical intent when encountering similar situations in the future.
[0068] It solves the problem of missing target areas or unclear boundaries caused by poor soft tissue contrast when delineating based solely on CT images.
[0069] In complex cases such as tumor recurrence, postoperative anatomical changes, or severe adhesion between lesions and normal tissues, single-modal imaging often provides limited information. Multimodal fusion provides models and physicians with cross-validation and complementary information, greatly reducing the difficulty of decision-making and becoming a powerful tool for treating difficult and complex cases.
[0070] In one possible implementation, generating multiple adaptive models includes:
[0071] The basic segmentation model was trained using different delineation standard datasets to optimize all its parameters and obtain an initial optimized model.
[0072] The parameters of the image encoder of the initial optimization model are fixed, and the parameters of its decoding-related units are adjusted using different delineation standard datasets to obtain multiple delineation standard-specific models; the decoding-related units include memory attention units and mask decoders.
[0073] Specifically, a low-rank adaptation method is used to adjust the parameters of the decoding-related units.
[0074] The working principle of the above technical solution is as follows:
[0075] Using one (or more) delineated standard datasets, perform end-to-end training on all parameters of the MedSAM2 model (including the image encoder, cue encoder, memory attention unit, and mask decoder) to obtain the initial optimized model.
[0076] The image encoder of the initial fine-tuning model is frozen, fine-tuned using LoRA (Low-Rank Adaptation), the memory attention unit and mask decoder are updated, and the model is saved to different delineation standard models.
[0077] The LoRA fine-tuning decoding logic integrates the LoRA adapter into the memory attention unit and the mask decoder. Then, using different delineation standard datasets, LoRA fine-tuning is performed on these two parts separately.
[0078] The memory attention unit is responsible for integrating historical information or context, and LoRA fine-tunes it to learn to focus on image features most relevant to a specific criterion.
[0079] The mask decoder is responsible for generating the final pixel-level segmentation mask, and LoRA fine-tunes it to learn to translate the integrated features into precise boundaries that conform to a specific standard contour.
[0080] By running end-to-end training and fine-tuning and LoRA fine-tuning for each delineation standard, multiple delineation standard-specific models are obtained. They share the same powerful image encoder, but each has a unique, lightweight LoRA adapter specifically designed to perform the specific delineation specification.
[0081] The effects of the above technical solution are as follows:
[0082] In the second stage, most of the parameters of the model are frozen (the image encoder usually has the largest number of parameters), and only a very small number of parameters are updated using LoRA, making the training of each new standard model extremely fast and significantly reducing the required computing resources (such as GPU memory). Saving a complete model only requires storing the large initial optimized model once, and afterwards, each standard dedicated model is essentially just a LoRA weight file of a few megabytes, which is crucial for deploying and managing dozens of models with different standards in a clinical environment. Freezing the image encoder ensures that the general visual representation ability learned by the model from a large amount of data will not be damaged or overwritten (i.e., catastrophic forgetting) due to learning new standards. LoRA focuses on adapting the decision-making part (decoder) of the model to new standards, enabling the dedicated model to utilize powerful general features and perform precise standard-specific delineation, usually with better and more stable results than directly training from scratch or fine-tuning all parameters of a small-data model.
[0083] In one possible implementation, adjusting the parameters of its decoding-related units using different delineation standard datasets respectively includes:
[0084] Sequentially using different delineation standard datasets to sequentially adjust the parameters of the decoding-related units in a continuous learning manner;
[0085] During the sequential adjustment process, an anti-forgetting strategy based on gradient projection is adopted to reduce the interference to the previously learned knowledge of the delineation standards.
[0086] During the sequential adjustment process, when training using the k-th delineation standard dataset, calculate the cosine similarity between the current model parameter update direction and the gradient directions of important parameters on the previously learned k - 1 delineation standards;
[0087] Based on the cosine similarity, project and correct the gradient vector of the current training to reduce the interference to the previously learned knowledge of the delineation standards.
[0088] In a specific application: Whenever the model finishes learning the i-th delineation standard (i < k), the system will analyze and save a set of gradient directions of important parameters. Usually, it is achieved by calculating the gradient of the task loss function with respect to the model parameters (such as the square or mean of the gradient), identifying the neural network paths that are most sensitive to completing the task. These gradient directions together form a protection space, symbolizing the acquired old knowledge.
[0089] [[ID=二十一]]When starting to train using the dataset of the k-th standard, the model will perform forward propagation to calculate the loss and backward propagation to obtain the original gradient vector Gk for the current task, that is, the proposed update direction of the model parameters. The system will calculate the cosine similarity between Gk and each old-task important gradient direction in the knowledge protection space.
[0090] Cosine similarity measures the similarity of two vectors along their directions. A positive value indicates that the directions are roughly the same, and updating Gk may be harmless or even beneficial to the old task. A negative value indicates that the directions are opposite, and updating Gk will directly destroy the key parameters required to complete the old task, leading to performance degradation (forgetting).
[0091] For all old task gradient directions that have a negative cosine similarity to Gk, the system performs a mathematical projection operation. The important gradient directions of the old tasks are considered as a wall. The model wants to update in the direction of Gk (potentially hitting the wall). The projection correction is equivalent to calculating a new direction parallel to all these walls (i.e., without conflict), and then letting the model update along this corrected direction. The component overlapping with all conflicting old task gradient directions is subtracted from the original gradient Gk. This results in an update direction that reduces the loss on the new task without increasing the loss on the old task. After the k-th criterion is learned, its important parameter gradient directions are also added to the knowledge protection space to prepare for learning the (k+1)-th and subsequent criters.
[0092] The effects of the above technical solution are as follows:
[0093] It fundamentally alleviates the catastrophic forgetting problem in continuous learning. Like an expert, the system can learn new guidelines while firmly remembering all previously learned guidelines, becoming a truly knowledge-rich AI assistant. It ensures the performance stability and reliability of each delineation standard-specific model, allowing doctors to trust that the system will not forget how to execute old standards after learning new ones. It supports low-cost, sustainable incremental updates. When new international delineation guidelines are released, there is no need to collect all old standard data for time-consuming and labor-intensive full retraining. New knowledge can be safely and efficiently injected directly based on the current model, greatly simplifying model maintenance processes. It perfectly matches the parameter efficiency of LoRA, ensuring that the same lightweight LoRA adapter parameters can carry knowledge from multiple optimization standards without interference, maximizing parameter utilization efficiency.
[0094] In one possible implementation, the loss function used during the adjustment process is a composite loss function, which includes:
[0095] The standard adaptive loss term is used to minimize the prediction error on the current delineation standard dataset;
[0096] In addition, a knowledge conservation loss term is used to ensure that the adjusted model maintains consistency with the unadjusted model in terms of output on common anatomical structures.
[0097] The effects of the above technical solution are as follows:
[0098] Unlike passive defenses based on gradient projection, the knowledge conservation loss term provides an active anchoring mechanism that explicitly fixes the model’s general representational capabilities at the state before adjustment. This more directly and effectively prevents the model from catastrophic forgetting when adapting to new standards, making the model more robust on all learned standards.
[0099] The knowledge conservation loss term acts as a powerful regularizer, preventing the model from overfitting to the current small-scale new standard dataset; it encourages the model to learn an incremental update that is compatible with existing knowledge, resulting in better generalization ability when faced with noise or rare cases.
[0100] In one possible implementation, training the base segmentation model using different delineation standard datasets to optimize all its parameters and obtain an initial optimized model includes:
[0101] Based on the clarity of the target region outline in the image or the contrast with the surrounding tissue, a difficulty score is calculated for each training sample; then, the training samples are sorted from easy to difficult according to the difficulty score from low to high, and training is performed in this order.
[0102] In one possible implementation, the targets are sorted according to the sharpness of their outlines in the image, including:
[0103] Calculate the boundary tortuosity of the gold standard target region mask in each training sample; wherein the boundary tortuosity is quantified by the ratio of the contour perimeter to the circumference of a circle with equal area or the Fourier descriptor entropy of the boundary pixels.
[0104] The training samples are sorted in ascending order of the boundary tortuosity values.
[0105] In one possible implementation, the sorting based on the contrast between the target region and surrounding tissue in the image includes:
[0106] A ring is extended outward from the gold standard target mask to form the background area;
[0107] Calculate a measure of difference between the target region and the background region, such as the average Hausdorff distance or the average pixel value difference.
[0108] The training samples are sorted from highest to lowest based on the differences mentioned above.
[0109] In one possible implementation, the method includes:
[0110] During training, the model's current performance is periodically evaluated on the validation set;
[0111] Based on the model's segmentation performance on samples of different difficulty levels on the validation set, the mixing ratio of difficult and easy samples in subsequent training data is dynamically adjusted; once the model's performance on easy samples is stable, the proportion of difficult samples is automatically increased.
[0112] The working principle of the above technical solution is as follows:
[0113] It directly addresses the core challenge of delineation tasks—boundary uncertainty. By quantifying boundary curvature, the system prioritizes learning simple target areas with smooth boundaries and regular shapes, and then gradually challenges complex target areas with odd shapes and many boundary burrs, significantly improving the model's ability to generalize to complex contours and making it less prone to errors when encountering rare shapes.
[0114] Boundary curvature is an indicator closely related to the subjective difficulty of doctors when drawing. Using this indicator for ranking makes the AI's learning process closer to the real experience of clinical experts.
[0115] Specifically targeting the common challenge of low contrast in medical image segmentation, this method teaches the model to first establish a clear concept of what constitutes the target area and what constitutes the background in high-contrast images, and then gradually challenge images with blurred boundaries. This allows the model to find effective discriminative features even when the signal-to-noise ratio is low.
[0116] This training method effectively prevents the model from overfitting to obvious features that can be segmented by simple thresholds, forcing it to learn deeper and more fundamental feature representations, thus exhibiting stronger robustness in complex situations such as tumor adhesion to surrounding tissues.
[0117] It breaks away from fixed, pre-set schedules and can adjust the teaching plan in real time based on the model's own learning status. Once the model has learned sufficiently on simple samples, it automatically increases the proportion of difficult samples to prevent wasted training time and accelerate model convergence.
[0118] When the model is still unstable at the current difficulty level, the system will maintain or reduce the difficulty to ensure that the model fully masters the knowledge of the current stage before entering the next stage, effectively preventing training crashes and overfitting.
[0119] In one possible implementation, the segmentation of the target medical image based on multiple adaptive models includes:
[0120] The target medical image is subjected to feature extraction based on a shared image encoder;
[0121] The extracted image features are input into decoders of multiple different delineation standard dedicated models for processing to generate the multiple candidate target region segmentation results.
[0122] The working principle of the above technical solution is as follows:
[0123] When a target medical image is input, the system does not feed it into each individual delineation standard-specific model. Instead, it invokes a shared image encoder. This encoder is shared by all delineation standard-specific models, and its parameters have been fixed during previous multi-standard adaptation training. This image encoder performs a forward pass through the input image exactly once, generating a highly abstract feature map rich in semantic information. This feature map serves as the common basis for all subsequent decoding operations.
[0124] The shared feature map generated in the previous step is simultaneously copied and distributed to multiple different decoder parts of the delineation standard-specific model (i.e., memory attention units and mask decoders).
[0125] Each dedicated model's decoder has been trained to be an expert in a particular delineation criterion, receiving the same shared feature maps, but interpreting these features differently based on its unique internal parameters (especially adapter parameters learned through techniques such as LoRA):
[0126] The decoder for Model A (such as the RTOG standard) generates the target boundary according to the specifications of the RTOG guidelines.
[0127] The decoder for Model B (e.g., the ESTRO standard) will output results according to the ESTRO standard.
[0128] These decoding processes are executed in parallel without waiting for each other, maximizing the use of computing resources (such as the parallel computing capabilities of GPUs).
[0129] At almost the same time, multiple decoders each output a complete segmentation mask. These masks correspond to the segmentation results of multiple candidate target regions according to various delineation criteria, and are presented to the doctor for selection and further processing.
[0130] The effects of the above technical solution are as follows:
[0131] By reducing its computational load from N (the standard number) times to 1 time, the inference speed of the entire system is improved by orders of magnitude. On devices such as GPUs, it is not necessary to load N complete models simultaneously; only a shared encoder and N lightweight decoders need to be loaded. Since the number of decoder parameters is much smaller than that of the encoder, runtime memory usage is greatly reduced. All candidate results originate from the exact same image features, eliminating contrast bias caused by minor random differences in feature extraction, and providing doctors with an absolutely fair and consistent benchmark for multi-standard comparisons. The differences observed by doctors are purely due to different decoding logic (i.e., different delineation standards), rather than instability of the model itself.
[0132] In one possible implementation, the segmentation of the target medical image based on multiple adaptive models includes:
[0133] Acquire the target medical image data and parse its format;
[0134] The target medical image is processed using multiple segmentation models configured with different delineation standards to generate the segmentation results of the multiple candidate target regions.
[0135] In practical applications, the system first acquires the target medical image data (such as CT sequences in DICOM format), performs necessary format parsing and preprocessing (such as pixel value normalization and resampling to standard resolution), and converts it into a tensor format that meets the model's input requirements. The system has multiple pre-trained segmentation models, each specifically configured and optimized for a particular delineation standard (such as RTOG or ESTRO). Subsequently, the system simultaneously inputs the same preprocessed target medical image data into these different segmentation models.
[0136] Each model performs forward propagation (inference) independently:
[0137] The encoder inside the model extracts image features.
[0138] The decoder decodes features into segmentation masks based on delineation logic learned under specific criteria.
[0139] Because each model carries different expertise (delineation criteria), they will output different segmentation results for the same image.
[0140] After all models have been processed, the system collects their respective outputs to form a set (e.g., 3 to 5) of candidate target region segmentation results corresponding to different delineation criteria, and presents them to the user in a unified manner.
[0141] Doctors can see the differences in target area under different guidelines within seconds without having to manually switch software or repeat operations, providing a key basis for developing the best treatment plan.
[0142] Since each model is trained independently and completely for a specific standard, its output is the purest representation under that standard, avoiding mutual interference between knowledge from different standards and ensuring the authority and accuracy of each candidate result.
[0143] In one possible implementation, during segmentation, different candidate target region segmentation results are generated by providing different cue information to the cue-based segmentation model; wherein the cue information includes at least one of points, bounding boxes, or text descriptions.
[0144] Users set prompts on the interface (such as click points, bounding boxes, text, or coarse masks), and the model quickly provides multiple candidate target region segmentation results, along with corresponding standard descriptions.
[0145] The effects of the above technical solution are as follows:
[0146] By changing the prompt information (points, boxes, text), multiple candidate target regions can be dynamically generated using only a basic segmentation model. This fundamentally replaces the traditional approach that requires pre-setting and loading multiple dedicated models, resulting in a sharp decrease in storage space and computing resource consumption, greatly simplifying the system architecture, and reducing deployment and maintenance costs.
[0147] Doctors are no longer passively accepting a limited number of pre-set standard results. Instead, they can immediately obtain new segmentation candidates by adjusting prompts in real time (such as moving the position of the point, changing the size of the box, and modifying the text description). This realizes an interactive and exploratory delineation mode, enabling doctors to actively and precisely guide the AI output until it meets their clinical needs.
[0148] Different prompts are the most direct reflection of doctors' professional knowledge and operational intentions. This mechanism transforms doctors' spatial prior knowledge (through points and boxes) and semantic instructions (through text) into guidance that the model can understand, ensuring that the generated candidate results are deeply integrated with doctors' clinical judgment from the beginning, significantly improving the controllability, accuracy and clinical acceptability of the results.
[0149] It covers multiple prompting methods, including points, bounding boxes, and text descriptions, providing doctors with a unified and flexible interactive interface. It can adapt to different doctors' operating preferences and handle full-scene drawing tasks, from coarse-grained positioning (bounding boxes) to fine-grained correction (points) and advanced semantic commands (text), greatly improving the system's practicality and ease of use.
[0150] In one possible implementation, the online learning of the target segmentation model to update its parameters includes:
[0151] Use the segmentation results obtained after user selection and / or correction operations as new training samples;
[0152] Based on the newly added training samples, the parameters of the target segmentation model are updated in real time or quasi-real time using incremental learning, so that the model can adapt to the drawing habits of specific institutions or users.
[0153] In one possible implementation, the online learning of the target segmentation model to update its parameters includes:
[0154] Maintain a training sample cache queue for online learning;
[0155] The segmentation results confirmed by user interaction and the corresponding medical images are used as sample pairs and added to the queue;
[0156] When the number of samples in the queue reaches a preset threshold, or when a user triggers a command, the parameters of the target segmentation model are updated once using all the samples in the queue.
[0157] The effects of the above technical solution are as follows:
[0158] By setting a preset threshold, the computational overhead and the risk of model performance instability (oscillation) caused by updating the model immediately after each sample are avoided. User-triggered commands empower doctors to actively instruct the model to learn after making important corrections, ensuring the timely injection of key clinical experience and meeting the high demands of clinical practice for certainty and controllability. Accumulating a small batch of samples before updating provides a more stable and representative gradient direction compared to single-sample updates, effectively preventing overfitting on individual samples and making the learning process smoother and more robust.
[0159] In one possible implementation, the method further includes:
[0160] When displaying the segmentation results of the multiple candidate target regions, the delineation criteria on which each result is based are automatically labeled;
[0161] And / or, calculate the difference between each result and the corresponding standard template;
[0162] And / or, the segmentation results of the multiple candidate target regions are visualized and overlaid.
[0163] The effects of the above technical solution are as follows:
[0164] It directly eliminates the risk of doctors confusing results from different standards. By clearly labeling standard names such as RTOG and ESTRO, it provides a clear standard basis for treatment plans, which not only makes it easier for doctors to identify, but also meets the strict requirements of medical quality control and auditing for the traceability of the plan.
[0165] Transforming subjective visual comparisons into objective quantitative data (such as Dice coefficient and Hausdorff distance) can directly and accurately tell doctors how much deviation there is between the current automatically drawn results and the gold standard template. This provides doctors with a strong scientific reference for judging the quality and acceptability of the candidate results and reduces subjective assumptions.
[0166] By overlaying the delineation results of different standards onto the same image using semi-transparent overlays of different colors, doctors can clearly see the areas of consistency and discrepancy between the different standards. This intuitive presentation method greatly reduces the cognitive load of visually fusing and spatially comparing multiple results, allowing doctors to quickly focus their attention on key anatomical areas where there is controversy, thereby making more efficient and accurate decisions.
[0167] By combining labeling, quantification, and overlay display, a complete closed-loop workflow for comparison, analysis, and decision-making is formed. Doctors no longer need to switch back and forth between different views or tabs, nor do they need to manually measure and calculate. They can obtain comprehensive information within a single interface, raising the efficiency of the complex task of multi-standard comparison to an unprecedented level.
[0168] In one possible implementation, the method further includes:
[0169] It stores multiple versions of data during the target delineation process and supports switching and comparing multiple versions based on delineation standards, timestamps, or user identifiers.
[0170] By fully recording the entire process of AI initial candidates, user selection, interactive corrections to the final draft, an immutable electronic audit trail is formed. In clinical quality review, scientific research data traceability, or medical disputes, it can provide a legally valid and complete chain of decision-making evidence, meeting the highest standards of medical compliance requirements.
[0171] When multiple doctors collaborate on the same case, each doctor's revisions and decision-making process are clearly visible, facilitating communication, experience sharing, and a unified final solution. Doctors can revert to any historical point in time. If the current revision is unsatisfactory or an error occurs, a one-click revert to any previous stable version is possible, greatly enhancing error tolerance and operational safety.
[0172] The stored dynamic version data serves as extremely vivid teaching material. Trainees can watch, like watching a video replay, how experts select from multiple candidate results generated by AI, and how they achieve the perfect result through several key interactive corrections.
[0173] By switching between and comparing versions based on different standards and at different time points, doctors can gain a deeper understanding of the origins of AI-generated results and the differences caused by different decisions. This transparency and comparability greatly enhances doctors' trust in AI-assisted results, enabling them to move from passively accepting them to actively understanding and utilizing them.
[0174] In one possible implementation, the method includes the user's correction operation on the candidate target region segmentation result, which includes at least one of adding regions, deleting regions, or modifying the contour boundaries of the target region.
[0175] By providing three core operations—adding regions (to handle missing strokes), deleting regions (to handle extra strokes), and modifying outline boundaries (to handle inaccurate boundaries)—the system covers all possible correction scenarios, giving doctors the ability to precisely control the AI-generated results at the pixel level. This allows them to perform full-granular corrections on the target area, from overall to local, from coarse to fine adjustments, until the results fully meet their clinical requirements. The intuitive graphical operations (such as adding strokes, erasing and deleting, and dragging boundaries) greatly simplify the cumbersome process of traditional layer-by-layer manual drawing.
[0176] In one possible implementation, after the online learning step, the following is also included:
[0177] The final target segmentation results are converted into a standard medical image format and output for use by external systems.
[0178] The external system includes a radiotherapy planning system (TPS) or a picture archiving and communication system (PACS).
[0179] The above method eliminates the cumbersome steps of manually exporting, importing, and converting data between different systems, achieving one-click data transfer and greatly improving overall diagnostic and treatment efficiency. By adhering to standard medical image formats (such as DICOM) for output, it ensures that the outline results are not distorted or lost during transmission to TPS and PACS, providing a reliable data foundation for subsequent accurate dosage calculations, long-term data archiving, and sharing, and eliminating medical risks caused by data format errors from the workflow perspective. This function allows the automatic outline system of this invention to be directly embedded into the hospital's existing information ecosystem, seamlessly connecting with the professional systems (TPS / PACS) most frequently used by doctors.
[0180] Example 2: This embodiment provides a delineation model construction system based on online learning, the system comprising:
[0181] The model generation module is configured to adjust the pre-trained base segmentation model on different delineation standard datasets to generate multiple adaptive models, each of which corresponds to a delineation standard.
[0182] The segmentation processing module is configured to segment the target medical image based on multiple adaptive models, generating multiple candidate target region segmentation results corresponding to multiple delineation criteria.
[0183] The online learning module is configured to determine a target segmentation model based on the user's selection and / or correction operations on the candidate target region segmentation results, and to perform online learning on the target segmentation model to update its parameters; wherein the target segmentation model is one selected from the plurality of adaptive models, or a newly created independent model with the same structure as the base segmentation model.
[0184] In one possible implementation, the underlying segmentation model is a cue-based segmentation model, which includes an image encoder, a cue encoder, and a mask decoder.
[0185] In one possible implementation, the underlying segmentation model is the MedSAM2 model or a variant of the MedSAM2 model.
[0186] In one possible implementation, the underlying segmentation model is configured to receive and process multiple modal images from CT, MRI, and PET, and to integrate multimodal information through its internal feature fusion mechanism.
[0187] In one possible implementation, the model generation module is further configured as follows:
[0188] The basic segmentation model was trained using different delineation standard datasets to optimize all its parameters and obtain an initial optimized model.
[0189] The parameters of the image encoder of the initial optimization model are fixed, and the parameters of its decoding-related units are adjusted using different delineation standard datasets to obtain multiple delineation standard-specific models; the decoding-related units include memory attention units and mask decoders.
[0190] In one possible implementation, adjusting the parameters of the decoding-related units using different delineation standard datasets includes:
[0191] Different delineation standard datasets are used sequentially to adjust the parameters of the decoding-related units in a continuous learning manner;
[0192] During the sequence adjustment process, a gradient projection-based anti-forgetting strategy is adopted to reduce interference with previously learned standard knowledge.
[0193] In one possible implementation, the loss function used during the adjustment process is a composite loss function, which includes:
[0194] The standard adaptive loss term is used to minimize the prediction error on the current delineation standard dataset;
[0195] In addition, a knowledge conservation loss term is used to ensure that the adjusted model maintains consistency with the unadjusted model in terms of output on common anatomical structures.
[0196] In one possible implementation, the model generation module is further configured as follows:
[0197] Based on the clarity of the target region outline in the image or the contrast with the surrounding tissue, a difficulty score is calculated for each training sample; then, the training samples are sorted from easy to difficult according to the difficulty score from low to high, and training is performed in this order.
[0198] In one possible implementation, the targets are sorted according to the sharpness of their outlines in the image, including:
[0199] Calculate the boundary tortuosity of the gold standard target region mask in each training sample; wherein the boundary tortuosity is quantified by the ratio of the contour perimeter to the circumference of a circle with equal area or the Fourier descriptor entropy of the boundary pixels.
[0200] The training samples are sorted in ascending order of the boundary tortuosity values.
[0201] In one possible implementation, the sorting based on the contrast between the target region and surrounding tissue in the image includes:
[0202] A ring is extended outward from the gold standard target mask to form the background area;
[0203] Calculate a measure of difference between the target region and the background region, such as the average Hausdorff distance or the average pixel value difference.
[0204] The training samples are sorted from highest to lowest based on the differences mentioned above.
[0205] In one possible implementation, the model generation module is further configured as follows:
[0206] During training, the model's current performance is periodically evaluated on the validation set;
[0207] Based on the model's segmentation performance on samples of different difficulty levels on the validation set, the mixing ratio of difficult and easy samples in subsequent training data is dynamically adjusted; once the model's performance on easy samples is stable, the proportion of difficult samples is automatically increased.
[0208] In one possible implementation, the segmentation processing module is further configured as follows:
[0209] The target medical image is subjected to feature extraction based on a shared image encoder;
[0210] The extracted image features are input into decoders of multiple different delineation standard dedicated models for processing to generate the multiple candidate target region segmentation results.
[0211] In one possible implementation, the segmentation processing module is further configured to:
[0212] Acquire the target medical image data and parse its format;
[0213] The target medical image is processed using multiple segmentation models configured with different delineation standards to generate the segmentation results of the multiple candidate target regions.
[0214] In one possible implementation, the segmentation processing module is further configured to:
[0215] During segmentation, different candidate target region segmentation results are generated by providing different cue information to the cue-based segmentation model; wherein the cue information includes at least one of points, bounding boxes, or text descriptions.
[0216] In one possible implementation, the online learning module is further configured for:
[0217] Use the segmentation results obtained after user selection and / or correction operations as new training samples;
[0218] Based on the newly added training samples, the parameters of the target segmentation model are updated in real time or quasi-real time using incremental learning, so that the model can adapt to the drawing habits of specific institutions or users.
[0219] In one possible implementation, the online learning module is further configured for:
[0220] Maintain a training sample cache queue for online learning;
[0221] The segmentation results confirmed by user interaction and the corresponding medical images are used as sample pairs and added to the queue;
[0222] When the number of samples in the queue reaches a preset threshold, or when a user triggers a command, the parameters of the target segmentation model are updated once using all the samples in the queue.
[0223] In one possible implementation, the system further includes:
[0224] The results display module is configured as follows:
[0225] When displaying the segmentation results of the multiple candidate target regions, the delineation criteria on which each result is based are automatically labeled;
[0226] And / or, calculate the difference between each result and the corresponding standard template;
[0227] And / or, the segmentation results of the multiple candidate target regions are visualized and overlaid.
[0228] In one possible implementation, the system further includes a version management module configured to:
[0229] It stores multiple versions of data during the target delineation process and supports switching and comparing multiple versions based on delineation standards, timestamps, or user identifiers.
[0230] In one possible implementation, the system further includes an interactive correction module configured to receive correction operations from a user on the candidate target region segmentation results, the correction operations including at least one of adding regions, deleting regions, or modifying the contour boundaries of the target region contour.
[0231] In one possible implementation, the system further includes:
[0232] The data interface module is configured to encapsulate and output the final target segmentation results according to a preset medical image communication standard, so that an external radiotherapy planning system (TPS) or image archiving and communication system (PACS) can receive and process the results.
[0233] The working principle and effect of the above technical solution are the same as the online learning-based delineation model construction method in Example 1, and will not be repeated here.
[0234] An embodiment of the present invention provides a radiotherapy system, including a processing unit and a memory. The memory stores a computer program, which, when executed by the processing unit, implements the steps of the method described in embodiment 1 of the present invention.
[0235] This invention also provides an electronic device, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the steps of the method described in Embodiment 1 of this invention or the functions of the system described in Embodiment 2 of this invention.
[0236] This invention also provides a computer-readable storage medium for storing a computer program. When the computer program is executed, it implements the steps of the method in this invention. The specific implementation method is consistent with the implementation method and the technical effect achieved in the above method embodiments, and some contents will not be repeated.
[0237] In this invention, a readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. The program product can take the form of any combination of one or more readable media. A readable medium can be a readable signal medium or a readable storage medium. A readable storage medium can be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of readable storage media (a non-exhaustive list) include: an electrical connection having one or more wires, a portable disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof.
[0238] Computer-readable storage media may include data signals propagated in baseband or as part of a carrier wave, carrying readable program code. Such propagated data signals may take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. The readable storage medium may also be any readable medium capable of sending, propagating, or transmitting a program for use by or in conjunction with an instruction execution system, apparatus, or device. The program code contained on the readable storage medium may be transmitted using any suitable medium, including but not limited to wireless, wired, optical fiber, RF, or any suitable combination thereof. Program code for performing operations of the present invention may be written in any combination of one or more programming languages, including object-oriented programming languages such as Java and C++, as well as conventional procedural programming languages such as C or similar programming languages. The program code may be executed entirely on a user computing device, partially on an associated device, as a standalone software package, partially on a user computing device and partially on a remote computing device, or entirely on a remote computing device or server. In cases involving remote computing devices, the remote computing devices can be connected to user computing devices via any type of network, including local area networks (LANs) or wide area networks (WANs), or they can be connected to external computing devices (e.g., via the Internet using an Internet service provider).
[0239] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the invention without departing from the principles and spirit of the invention, and all such changes should fall within the protection scope of the claims of the present invention.
Claims
1. A method for constructing a delineation model based on online learning, characterized in that, The method includes: The pre-trained basic segmentation model is adjusted on different delineation standard datasets to generate multiple adaptive models, each corresponding to a delineation standard. The basic segmentation model includes an image encoder, a memory attention unit, and a mask decoder. When generating the adaptive model, all parameters of the basic segmentation model are first optimized to obtain an initial optimized model. Then, the image encoder parameters of the initial optimized model are fixed, and only the parameters of the memory attention unit and the mask decoder are adjusted using the low-rank adaptation (LoRA) method to obtain multiple delineation standard-specific models for the adaptive models. Based on multiple adaptive models, the target medical image is segmented to generate multiple candidate target region segmentation results corresponding to various delineation standards. Based on the user's selection and / or correction of the candidate target region segmentation results, a target segmentation model is determined. The user-corrected results are stored as incremental samples in an online learning cache queue. When the queue reaches a preset threshold or a user trigger command is received, the target segmentation model is updated with parameters through incremental online learning. The target segmentation model is used to receive the incremental online learning updates, and the target segmentation model is one selected from the plurality of adaptive models, or a newly created independent model with the same structure as the basic segmentation model. The segmentation process of the target medical image based on multiple adaptive models includes: The target medical image is subjected to feature extraction based on a shared image encoder; The extracted image features are input into decoders of multiple different delineation standard dedicated models for processing to generate the multiple candidate target region segmentation results.
2. The method for constructing a delineation model according to claim 1, characterized in that, The basic segmentation model is a cue-based segmentation model, which includes an image encoder, a cue encoder, and a mask decoder.
3. The method for constructing a delineation model according to claim 1, characterized in that, The underlying segmentation model is the MedSAM2 model or a variant of the MedSAM2 model.
4. The method for constructing a delineation model according to claim 1, characterized in that, The loss function used to adjust the basic segmentation model is a composite loss function, which includes: The standard adaptive loss term is used to minimize the prediction error on the current delineation standard dataset; In addition, a knowledge conservation loss term is used to ensure that the adjusted model maintains consistency with the unadjusted model in terms of output on common anatomical structures.
5. The method for constructing a delineation model according to claim 1, characterized in that, The initial optimized model is obtained by optimizing all parameters of the basic segmentation model, including: Based on the clarity of the target region outline in the image or the contrast with the surrounding tissue, calculate the difficulty score of each training sample, and sort the training samples from easy to difficult according to the difficulty score from low to high, and train them in this order. The sharpness of the contour is quantified by calculating the boundary curvature of the gold standard target area mask, and the boundary curvature is quantified by the ratio of the contour perimeter to the circumference of a circle with equal area or the Fourier descriptor entropy of the boundary pixels. The contrast with the surrounding tissue is quantified by forming a background region by extending a ring outward from the gold standard target region mask, and calculating the average Hausdorff distance or average pixel value difference between the target region and the background region.
6. The method for constructing a delineation model according to claim 1, characterized in that, The segmentation process of the target medical image based on multiple adaptive models includes: Acquire the target medical image data and parse its format; The target medical image is processed using multiple segmentation models configured with different delineation standards to generate the segmentation results of the multiple candidate target regions.
7. The method for constructing a delineation model according to claim 1, characterized in that, During segmentation, different candidate target region segmentation results are generated by providing different cue information to the cue-based segmentation model; wherein the cue information includes at least one of points, bounding boxes, or text descriptions.
8. The method for constructing a delineation model according to claim 1, characterized in that, Also includes: When displaying the segmentation results of the multiple candidate target regions, the delineation criteria on which each result is based are automatically labeled; And / or, calculate the difference between each result and the corresponding standard template; And / or, the segmentation results of the multiple candidate target regions are visualized and overlaid.
9. The method for constructing a delineation model according to claim 1, characterized in that, Also includes: It stores multiple versions of data during the target delineation process and supports switching and comparing multiple versions based on delineation standards, timestamps, or user identifiers.
10. The method for constructing a delineation model according to claim 1, characterized in that, Also includes: User correction operations for candidate target region segmentation results include at least one of adding regions, deleting regions, or modifying the contour boundaries of the target region.
11. A delineation model construction system based on online learning, characterized in that, The system includes: The model generation module is configured to adjust the pre-trained base segmentation model on different delineation standard datasets to generate multiple adaptive models, each of which corresponds to a delineation standard. The base segmentation model includes an image encoder, a memory attention unit, and a mask decoder. When generating the adaptive model, all parameters of the base segmentation model are first optimized to obtain an initial optimized model. Then, the image encoder parameters of the initial optimized model are fixed, and only the parameters of the memory attention unit and the mask decoder are adjusted using the low-rank adaptation (LoRA) method to obtain multiple delineation standard-specific models that serve as the adaptive models. The segmentation processing module is configured to segment the target medical image based on multiple adaptive models, generating multiple candidate target region segmentation results corresponding to multiple delineation criteria. The online learning module is configured to determine a target segmentation model based on the user's selection and / or correction of the candidate target region segmentation results, store the user-corrected results as incremental samples in an online learning cache queue, and perform incremental online learning to update the parameters of the target segmentation model when the queue reaches a preset threshold or when a user triggers an instruction; wherein, the target segmentation model is used to receive the incremental online learning update, and the target segmentation model is one selected from the plurality of adaptive models, or a newly created independent model with the same structure as the basic segmentation model; The segmentation processing module is configured as follows: The target medical image is subjected to feature extraction based on a shared image encoder; The extracted image features are input into decoders of multiple different delineation standard dedicated models for processing to generate the multiple candidate target region segmentation results.
12. A radiotherapy system, characterized in that, It includes a processing unit and a memory, the memory storing a computer program that, when executed by the processing unit, implements the steps of the method as claimed in any one of claims 1-10.
13. An electronic device, characterized in that, The electronic device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the steps of the method according to any one of claims 1-10.
14. A computer-readable storage medium, characterized in that, The storage medium stores computer instructions, and when the computer reads the computer instructions, the computer implements the steps of the method according to any one of claims 1-10.