Active learning method for ubiquitous organ ultrasonic analysis based on evolution diffusion sampling

Abstract

The invention discloses a generic organ ultrasonic analysis active learning method based on evolution diffusion sampling. The method comprises the following steps: constructing a heterogeneous data set by using multi-organ ultrasonic images in a DICOM format; the method comprises the following steps: pre-training a classifier-guided diffusion model by using an unlabeled set and a labeled set in a heterogeneous data set under the condition of an information label of an organ ultrasound image DICOM header file; sampling an unlabeled set by using a trained classifier-guided diffusion model, and introducing an information gain gradient guide mechanism and a local optimization evolutionary algorithm in the sampling process to generate a synthetic sample subset with a high information value; calculating the similarity between each synthetic sample and a real sample, and selecting the real sample with the most information content for expert labeling to form an active labeling set; the downstream diagnostic model is trained and fine-tuned on the combination of the label set and the active label set. According to the method, efficient sample selection can be realized in a scene in which annotation resources are limited, the annotation cost is remarkably reduced, and the model performance is improved.

Description

Technical Field

[0001] This invention belongs to the field of generative algorithms and active learning in deep learning technology, specifically involving an active learning method for pan-organ ultrasound analysis based on evolutionary diffusion sampling. Background Technology

[0002] Ultrasound (US) imaging, due to its non-invasiveness, safety, and wide applicability, has become a fundamental technology in modern medical imaging. This technology can examine multiple organ systems throughout the body, including but not limited to the thyroid, breast tissue, carotid arteries, liver, appendix, and ovarian structures. The broad organ applicability of ultrasound examination makes it a crucial diagnostic tool for clinicians to assess various anatomical structures and pathological conditions, playing an indispensable role from early detection to treatment planning. In recent years, deep learning has achieved significant results in medical ultrasound image analysis tasks. However, the effectiveness of these models highly depends on the acquisition of large-scale, high-quality labeled datasets. Since annotation requires manual work by experts, the construction of such datasets is particularly time-consuming and costly in the medical field.

[0003] To alleviate this burden, Active Learning (AL) has emerged as an intelligent sample selection strategy. This method iteratively selects the most informative samples for expert annotation, aiming to maximize model performance with minimal annotation workload. It maximizes the utility of limited annotation resources, providing a fundamental solution to the annotation bottleneck problem in medical image analysis. The Active Learning framework needs to extract meaningful representations from sparsely labeled datasets while developing strong generalization capabilities to address the inherent uncertainty in the distribution of unseen and unlabeled data.

[0004] However, the clinical workflow suffers from "image-label separation," resulting in highly redundant data across time and space, severe class imbalance, and extremely high data heterogeneity due to different anatomical structures, pathological types, equipment models, and acquisition protocols. These characteristics make traditional active learning methods ill-suited for the complex needs of multi-organ ultrasound diagnostic scenarios. Therefore, the core challenge of active learning in medical ultrasound image analysis lies in determining which samples to label to maximize model performance, given limited and unevenly distributed annotation resources.

[0005] Current research on active learning primarily focuses on sample selection strategies, which can be broadly categorized into two paradigms: those based on model uncertainty and those based on data representativeness. Methods based on model uncertainty quantify the model's prediction confidence for unlabeled samples using metrics such as entropy, marginal sampling, and variational ratio, prioritizing the labeling of samples with the least model uncertainty. Their core assumption is that uncertain samples contain more information and can more effectively improve the decision boundary. Methods based on data representativeness, on the other hand, focus on the distribution characteristics of unlabeled data. They identify samples that best represent the overall data distribution through core set selection, density-weighted sampling, or clustering-based methods, ensuring that the selected samples have broad coverage and diversity, and avoiding sampling bias.

[0006] However, existing active learning methods lack explicit optimization mechanisms for downstream diagnostic performance: methods based on model uncertainty rely solely on uncertainty, potentially leading to skewed training data distribution and neglecting sample learnability; methods based on data representativeness, while ensuring distribution coverage, tend to select redundant centroids while ignoring key outliers, resulting in poor performance in diagnostic tasks with subtle pathological differences. Although hybrid approaches attempting to balance uncertainty and representativeness exist, these existing methods are still constrained by a posterior active learning framework, where the selection model primarily operates under a semi-supervised learning (SSL) paradigm, and the performance of active learning methods cannot be validated until the downstream task is fully trained and evaluated.

[0007] Furthermore, existing active learning methods are primarily based on traditional convolutional neural networks (CNNs), which optimize decision boundaries through backpropagation. This approach can limit generalization ability, especially in imbalanced multi-class classification scenarios, where it struggles to accurately estimate class conditional probabilities, and estimation errors can negatively impact model performance. Studies also show that backpropagation can accumulate errors or cause gradients to vanish during iteration. This can lead active learning models to select suboptimal data points, achieving local rather than global optimization. Although modern optimization and regularization techniques have addressed some of these limitations, fundamental challenges remain in the complex learning environments of multi-organ ultrasound image classification tasks.

[0008] Furthermore, current active learning methods in the field of medical ultrasound imaging primarily target single-organ, single-center datasets, meaning they still operate in a task-specific manner, working in isolation for a single anatomical region or pathological type, rather than utilizing shared representations across structures. These scenarios implicitly rely on homogeneous embedding spaces, where features can be effectively acquired through contrastive learning methods. However, due to increased heterogeneity between organs, institutions, and image quality, existing active learning methods often perform poorly on more complex, pan-organ datasets.

[0009] Furthermore, previous methods typically relied on ranking selectors from a single data feature space (e.g., finite labeled data), which carries the risk of overfitting. While active learning techniques based on Generative Adversarial Networks (GANs) Bayesian methods mitigate this problem by generating synthetic samples and expanding the feature space, a key question arises: when the generative model is trained on the same finite dataset, how much real-world information value do artificially generated samples actually provide? Meanwhile, diffusion models of generative technologies demonstrate a powerful ability to capture complex medical image distributions and generate realistic anatomical samples, providing new generative prior opportunities for sample selection strategies in active learning. Diffusion models generate high-quality samples in the latent feature space through an iterative denoising process, and their sampling trajectories contain rich distribution information. Evolutionary algorithms, as a powerful class of global optimization methods, can avoid getting trapped in local optima through population evolution and diversity preservation mechanisms, performing excellently in high-dimensional complex optimization problems. These technological advancements provide a multi-dimensional theoretical and technical foundation for building a new generation of general active learning frameworks. In particular, combining the generative capabilities of diffusion models with the global optimization characteristics of evolutionary algorithms holds promise for overcoming the limitations of existing active learning methods in multi-organ ultrasound diagnosis. Summary of the Invention

[0010] To address the problems existing in the prior art, this invention provides an active learning method for pan-organ ultrasound analysis based on evolutionary diffusion sampling. By deeply integrating active learning with diffusion models and evolutionary algorithms, an active learning framework without backpropagation is constructed. This framework can globally optimize the labeled subset in a high-dimensional feature space, thereby achieving efficient sample selection in scenarios with limited labeling resources, significantly reducing labeling costs and improving model performance.

[0011] To solve the above-mentioned technical problems and achieve the above-mentioned technical effects, the present invention is implemented through the following technical solution: An active learning method for pan-organ ultrasound analysis based on evolutionary diffusion sampling includes the following steps: Step 1) Collect and preprocess DICOM format multi-organ ultrasound images covering multiple anatomical sites to construct a heterogeneous dataset containing an unlabeled set, a labeled set, and a test set; Step 2) A classifier-guided diffusion model is constructed using Denoising Diffusion Implicit Models (DDIM) and a guided classifier. The classifier-guided diffusion model is pre-trained using the information tags in the DICOM header file of the multi-organ ultrasound image as conditions, and the unlabeled set and the labeled set are used to master the complex feature distribution of multi-organ ultrasound images under different anatomical structures and pathological types. Step 3) The unlabeled set is sampled using the diffusion model guided by the classifier. During the sampling process, the information gain (IG) gradient guidance mechanism and the local optimization evolution algorithm (LOEA) are introduced to generate multiple synthetic samples with different guidance intensities, and form a subset of synthetic samples with high information value. Step 4) Calculate the cosine similarity between each synthetic sample in the synthetic sample subset and the corresponding real sample in the unlabeled set in the traditional convolutional neural network feature latent space, and select the real sample with the most information content for expert annotation to form an active annotation set; Steps 2-4 will be applied iteratively to each subset of data in the heterogeneous dataset; Step 5) The process of steps 2-4 will be applied iteratively to each subset of data in the heterogeneous dataset, and in each round of active learning iteration, the downstream diagnostic model will be trained and fine-tuned on the combination of the labeled set and the active labeled set.

[0012] Furthermore, the method for constructing the heterogeneous dataset is as follows: First, medical ultrasound image data covering six organs—appendix, breast, carotid artery, liver, ovary, and thyroid—were collected from multiple different medical centers. The multi-organ ultrasound images were in DICOM format, and the tag field of the DICOM header file contained information tags describing the relevant attributes of the corresponding multi-organ ultrasound images and their respective patients. The attribute information included the transmission format of the multi-organ ultrasound images, organ type, acquisition equipment manufacturer, and the patient's gender and age. Then, the multi-organ ultrasound images are preprocessed by image size standardization and normalization. Data from the same medical center are then grouped into corresponding data subsets and integrated to obtain the heterogeneous dataset. Finally, each data subset is randomly divided into an unlabeled set, a labeled set, and a test set according to a certain proportion. The number of attributes in the unlabeled set, the labeled set, and the test set is balanced according to the type of information tags in the DICOM header file of the multi-organ ultrasound images (including gender balance, age group balance, etc.).

[0013] Furthermore, the unlabeled set contains a large number of multi-organ ultrasound images with organ labels (e.g., appendix, breast, carotid artery, liver, ovary, thyroid) but no pathological labels, serving as the core data pool for the active learning framework, used for pre-training of the diffusion model and classifier, as well as for sample selection for the active learning framework.

[0014] Furthermore, the annotation set contains a small number of multi-organ ultrasound images that simultaneously carry organ labels (e.g., appendix, breast, carotid artery, liver, ovary, thyroid) and pathological labels (e.g., benign, malignant), for joint fine-tuning of the diffusion model and classifier, as well as training of downstream classification models. Furthermore, the test set also includes a small number of multi-organ ultrasound images (hereinafter referred to as test data) with both organ labels (e.g., appendix, breast, carotid artery, liver, ovary, thyroid) and pathological labels (e.g., benign, malignant) as an independent evaluation benchmark to evaluate the performance of downstream diagnostic models within the active learning framework.

[0015] Furthermore, the Denoising Diffusion Implicit Model (DDIM) adopts the U-Net architecture as the backbone structure of the denoising network. Its encoder consists of four downsampling modules with feature channels of 64, 128, 256 and 512 respectively. Its bottleneck layer uses a self-attention mechanism to capture global contextual information.

[0016] Furthermore, the guided classifier employs a downsampling backbone network based on the U-Net architecture, combined with an attention pooling mechanism, to enhance feature discrimination capabilities.

[0017] Furthermore, the pre-training process is divided into two stages: self-supervised pre-training and fully supervised fine-tuning, specifically as follows: 1) Self-supervised pre-training phase: Using the information tags in the DICOM header file of multi-organ ultrasound images as conditions, the denoising diffusion implicit model and the guided classifier are pre-trained on the unlabeled set using organ tags, respectively. This allows the denoising diffusion implicit model and the guided classifier to learn organ-level features (anatomical representations) in multi-organ ultrasound images in a self-supervised manner, thereby enabling them to extract organ features. At the same time, they can reflect the attribute information of the information tags in the DICOM header file of multi-organ ultrasound images (e.g., the difference between male and female organs), laying the foundation for the subsequent identification of pathological features. 2) Full-supervision fine-tuning stage: Using the information tags in the DICOM header file of multi-organ ultrasound images as conditions, the pre-trained guided classifier is integrated into the architecture of the pre-trained denoising diffusion implicit model to form a classifier-guided diffusion model. This model is then jointly fine-tuned on the labeled set using pathological labels. During the fine-tuning process, the initial stage focuses on the loss from organ reconstruction tasks, while the later stage focuses on the loss from pathological prediction tasks. This enables the classifier-guided diffusion model to capture subtle pathological features (such as benign / malignant tumors, normal / abnormal organs, etc.), thereby possessing the ability to identify pathological features and providing support for subsequent information gain-guided sampling. Through the pre-training described above, the classifier-guided diffusion model learns the potential feature spatial distribution of multi-organ ultrasound images, providing high-quality feature representations for subsequent sample sampling.

[0018] Furthermore, the pre-training method for the denoising diffusion implicit model is as follows: During the forward diffusion process, based on the information tags in the DICOM header files of multi-organ ultrasound images, the denoising diffusion implicit model first reads a large number of multi-organ ultrasound images with organ tags but no pathological tags from the unlabeled set, which are recorded as the original images. Then, Gaussian noise is gradually added to the original images according to a Markov chain, so that the original images gradually change from clear to blurry. ， This generates a series of noisy image sequences with different noise levels (gradually blurring from the clear original image to a purely noisy image); this process aims to simulate the degradation of data from clear to noisy, providing a training target for the model; In the reverse denoising process, using the information tags in the DICOM header file of the multi-organ ultrasound image as conditions, the denoising diffusion implicit model allows the denoising network to learn the mapping relationship between a series of noise image sequences with different noise levels generated by the forward diffusion process and the original image. The denoising network is trained to predict the noise added to the original image at each time step (i.e., given the current noise image and time step, predict the corresponding noise), and then starting from the pure noise image, through multiple iterations, gradually subtracts the noise, thereby restoring a series of noise images with different noise levels from blurry to clear original images. This process completes the training by minimizing the difference between the predicted noise and the real noise, enabling the model to learn the potential distribution of the data. During the training process, in the first 50% of the iterations, the number of organs is kept balanced. In the last 50% of the iterations, additional samples are added for organs with greater reconstruction loss. In addition, during the pre-training process of the denoising diffusion implicit model, two different organ ultrasound images can be weighted at the pixel level, and the information tags in the DICOM header files of these two organ ultrasound images can be used as reconstruction conditions, so that the denoising diffusion implicit model can understand different parts in the same organ ultrasound image during the reconstruction task.

[0019] Furthermore, during the pre-training forward diffusion process, the total number of diffusion steps (time steps) is defined as T=1000. Gaussian noise is gradually added to the original image according to the Markov chain, and the noise scheduling parameter β... t The value is increased linearly from 0.0001 to 0.02 to ensure that the final distribution of the data approximates the standard normal distribution.

[0020] Furthermore, the diffusion sampling process is as follows: The denoising diffusion implicit model is responsible for diffusion sampling of the unlabeled set. During the diffusion sampling process, the organ labels of the unlabeled set serve as prompts, and the information gain gradient is applied to guide the generation process of synthetic samples to avoid the cumulative error caused by backpropagation. At the same time, a local optimization evolution algorithm is introduced at each time step of the diffusion sampling process to achieve adaptive search of the high-dimensional feature space, thereby generating multiple synthetic samples with gradually increasing synthetic information content. After integration, a subset of synthetic samples with high information value is finally obtained.

[0021] The information gain is approximated by the entropy difference between adjacent time steps on the DDIM sampling trajectory of the classifier-guided denoising diffusion implicit model.

[0022] The information gain gradient is derived from the calculated information gain and is the rate of change of information gain with respect to the spatial variation of the input sample. It represents the direction and intensity of the change of information gain in the high-dimensional feature space and serves as the core signal guiding the sampling direction.

[0023] The synthetic samples are multi-organ synthetic ultrasound images with pseudo-organ labels (such as thyroid, breast, carotid artery, liver, appendix, ovary) and pseudo-pathological labels (such as benign / malignant, normal / abnormal, etc.), simulating the characteristics of medical ultrasound images in real clinical scenarios.

[0024] These synthetic samples can be used as a learning strategy for subsequent step-by-step training of downstream diagnostic models from simple to complex, or they can be retained and used directly to expand the annotation set.

[0025] Furthermore, the objective function for approximating the information gain is as follows: (8); In equation (8), t Indicates a time step. y i Indicates organ labeling, θ Indicates the parameters of the guiding classifier. x t Indicates time step t The intermediate state. x t-1 Indicates time step t-1 The intermediate state. Indicates intermediate state x t The corresponding entropy value, Indicates intermediate state x t-1 The corresponding entropy value.

[0026] Furthermore, the information gain gradient guidance mechanism is implemented by an implicit guidance field based on the gain information gradient. This implicit guidance field can bias the DDIM high-dimensional sampling space towards the gradient direction that generates the maximum gain information gradient. The implicit guidance field based on the gain information gradient contains a sampling gradient guidance function without backpropagation. (9); In equation (9), IG( x ) represents information gain, and λ represents the guiding strength parameter, λ∈1,3,5,7,9.

[0027] Furthermore, the specific operation of the Local Optimization Evolutionary Algorithm (LOEA) is as follows: (1) Population initialization: An initial population is randomly generated from the input set of unlabeled ultrasound samples. The unlabeled samples cover different organ structures and scanning angles to ensure the diversity of candidate samples in terms of anatomical morphology and imaging mode. The initial population size is set to 100 individuals, each individual representing a subset of candidate samples, which consists of ultrasound images of multiple organs. The number of generations is set to 50 to fully explore the distribution of discriminative features among different organs. (2) Fitness evaluation: The entropy value of each individual is calculated by guiding the classifier. The higher the entropy value, the stronger the uncertainty of the sample under different organ structures or pathological morphologies, and the stronger the challenge to model training. At the same time, the high-dimensional semantic features of each individual are extracted by the feature extractor and combined with the organ-level feature embedding space. The cosine similarity is calculated with the feature database of the labeled set. Individuals that are highly similar to existing samples are penalized to reduce redundancy. Finally, the entropy value and similarity penalty are combined to form the fitness function. (3) Selection operation: A tournament selection strategy is adopted, in which 5 individuals are randomly selected for comparison each time, and the individual with the highest fitness is retained each time. Through multiple rounds of selection operations, half of the individuals with the best fitness scores are retained as the parent population to maintain the discrimination diversity of the population in the multi-organ semantic space. (4) Crossover operation: Randomly select parent individuals pairs and use uniform crossover to recombine them at the sample subset level so that the offspring individuals can integrate sample features from different imaging modes. The crossover probability is set to 0.8. (5) Mutation operation: With a mutation probability of 0.1 and a standard deviation of 0.1, a controlled perturbation is added to the offspring individuals. By randomly replacing or perturbing some samples to select weights, the distribution perturbation caused by changes in probe angle and imaging quality in real clinical collection is simulated, thus avoiding the search process from getting trapped in local optima. (6) Elite retention strategy: The top 10% of individuals with the highest fitness in the previous generation are retained and directly enter the next generation to maintain the stable inheritance of high-value organ samples; when the improvement of the optimal fitness in 10 consecutive generations is less than 0.1%, the algorithm is judged to have converged and the evolution process is terminated in advance.

[0028] After 50 generations, the Local Optimization Evolutionary Algorithm (LOEA) returns the individual with the highest fitness score, which represents the optimal solution point in the parameter space that maximizes the fitness score.

[0029] Furthermore, the method for sample selection and labeling is as follows: First, calculate the cosine similarity between each synthetic sample and all real samples in the unlabeled set in the traditional convolutional neural network feature latent space. At the same time, consider the similarity of the information labels in the diagnostic report corresponding to the ultrasound image in a large text model. Then, perform a weighted average of the two similarities to obtain a comprehensive similarity. Then, for each synthetic sample, the real sample with the highest overall similarity is found as a candidate. All candidate real samples are sorted according to the overall similarity value of their corresponding synthetic samples. The top α% of real samples are selected and submitted to experts for independent annotation or summary annotation (the higher the real sample is, the more experts need to summarize the annotation at the same time), thus forming an active annotation set. Finally, the active annotation set is integrated into the existing annotation set to enrich the training data for subsequent optimization of the downstream diagnostic model.

[0030] Furthermore, since the synthetic samples with gradually increasing synthetic information content are used as the learning strategy for progressively training the downstream diagnostic model from simple to complex, each round of active learning iteration can begin by selecting relatively certain real samples for labeling and training the downstream diagnostic model. As the capabilities of the downstream diagnostic model improve, it gradually transitions to more challenging and uncertain real samples. This progressive learning strategy creates a more natural learning process for the downstream diagnostic model, significantly improving learning efficiency in scenarios with limited data.

[0031] The above-mentioned active learning method for pan-organ ultrasound analysis based on evolutionary diffusion sampling can be used to solve the sample selection problem in various multi-organ ultrasound image analysis tasks.

[0032] The beneficial effects of this invention are as follows: In detail, this invention achieves efficient sample selection in scenarios with limited annotation resources by combining the generative capabilities of the diffusion model and the global search advantages of the evolutionary algorithm. Based on a dual mechanism of information gain estimation and cosine similarity matching, this invention accurately maps the sample value assessment in the synthetic space to the real data space, significantly improving the accuracy and reliability of sample selection. This scheme defines information gain by calculating entropy differences on the DDIM sampling trajectory, achieving a technological innovation that quantifies sample information value without backpropagation. This provides an efficient solution for sample evaluation under weak supervision, significantly reducing computational overhead and enabling active learning of large-scale medical image data. The introduction of the evolutionary algorithm in this invention overcomes the limitation of traditional greedy strategies easily getting trapped in local optima. Through operations such as population initialization, fitness evaluation, selection, crossover, and mutation, it performs a global search in the high-dimensional feature space. Its population diversity mechanism ensures that the selected samples have broad distribution coverage in different regions of the feature space, effectively avoiding sample redundancy and bias problems. The progressive learning strategy introduced in this invention, which moves from deterministic to uncertain, generates sample subsets with increasing difficulty by dynamically adjusting the information gain strength. This allows downstream classifiers to follow a more cognitively consistent progressive learning path. In the early stages of model training, high-confidence samples are prioritized to establish a stable foundation. As the model's capabilities improve, it gradually transitions to more challenging and complex samples, creating a more natural and smooth learning process for downstream classifiers. This avoids the training instability caused by introducing complex samples all at once in traditional methods, and significantly improves learning efficiency and convergence stability in scenarios with limited labeled data.

[0033] Overall, the general active learning framework proposed in this invention achieves significant technological breakthroughs and clinical application value in multi-organ ultrasound image diagnosis tasks. By deeply integrating a diffusion model and an evolutionary algorithm, this invention overcomes the limitations of traditional active learning methods that target single anatomical sites or pathological types, constructing a unified multi-organ diagnostic architecture. It demonstrates stable performance advantages in six different diagnostic tasks: appendix diagnosis, breast tumor identification and diagnosis, carotid plaque assessment, fatty liver diagnosis, ovarian tumor assessment, and thyroid nodule diagnosis. Knowledge transfer between different anatomical structures is achieved through sharing the feature space of the diffusion model. Practical clinical application verification shows that comparative studies with radiologists of varying experience levels have reduced manual annotation time by 24-84%, significantly lowering the annotation cost of deploying clinical AI systems. This invention simultaneously supports two types of downstream classifier architectures: traditional convolutional neural networks and pedestal models, demonstrating its broad applicability and technological sustainability across different deep learning paradigms. The active learning framework of this invention has good versatility and scalability, and can be adapted to various model architectures and diagnostic tasks. It provides a practical technical path for the clinical translation and large-scale application of medical imaging artificial intelligence, and effectively alleviates the core bottlenecks of high expert annotation costs and difficult data acquisition.

[0034] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the invention and to implement it according to the contents of the specification, the preferred embodiments of the present invention are described in detail below with reference to the accompanying drawings. Specific embodiments of the present invention are given in detail below with reference to the accompanying drawings. Attached Figure Description

[0035] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this application, illustrate exemplary embodiments of the invention and, together with their description, serve to explain the invention and do not constitute an undue limitation thereof. In the drawings: Figure 1 This is a flowchart illustrating the steps of the active learning method for pan-organ ultrasound analysis based on evolutionary diffusion sampling, as described in this invention.

[0036] Figure 2 This is a diagram showing the overall network architecture and workflow of the active learning method for pan-organ ultrasound analysis based on evolutionary diffusion sampling, as described in this invention. Detailed Implementation

[0037] The preferred embodiments of the present invention will be described in detail below with reference to the accompanying drawings to provide a clearer understanding of the invention's purpose, features, and advantages. It should be understood that the embodiments shown in the drawings are not intended to limit the scope of the invention, but are merely illustrative of the essential spirit of the invention's technical solution.

[0038] In the following description, certain specific details are set forth for the purpose of illustrating various disclosed embodiments in order to provide a thorough understanding of the various disclosed embodiments. However, those skilled in the art will recognize that embodiments may be practiced without one or more of these specific details. In other instances, well-known apparatuses, structures, and techniques associated with this application may not have been shown or described in detail to avoid unnecessarily obscuring the description of the embodiments.

[0039] Unless the context requires otherwise, throughout the specification and claims, the word “comprising” and its variations, such as “including” and “having”, shall be understood to have an open, inclusive meaning, that is, to be interpreted as “including, but not limited to”.

[0040] Throughout this specification, references to "an embodiment" or "an embodiment" indicate that a particular feature, structure, or characteristic described in connection with the embodiment is included in at least one embodiment. Therefore, the appearance of "in an embodiment" or "an embodiment" in various places throughout the specification does not necessarily refer to the same embodiment. Furthermore, a particular feature, structure, or characteristic may be combined in any manner in one or more embodiments.

[0041] The singular forms “a” and “the” used in this specification and the appended claims include plural references unless otherwise expressly stated herein. It should be noted that the term “or” is generally used to mean “and / or” unless otherwise expressly stated herein.

[0042] Furthermore, the technical features involved in the different embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0043] To address the semi-supervised active learning problem in multi-organ ultrasound image classification, this invention redefines the traditional posterior probability-based sample selection problem as finding the optimal solution (e.g., the optimal labeled subset) in a high-dimensional feature space (such as the latent solution space). Based on this, this invention proposes a pan-organ ultrasound analysis active learning framework based on evolutionary diffusion sampling (EvoDiff), focusing on maximizing labeling efficiency in general medical image classification tasks. The proposed EvoDiff framework models sample selection as a global optimization problem while considering the maximum information gain (IG) during the sampling process, thereby more efficiently exploring the high-dimensional solution space. This effectively overcomes the limitation of existing active learning methods in improving the interpretability and generalization ability of models across different diagnostic scenarios.

[0044] See Figure 1 As shown, the method proposed in this invention, an active learning framework for pan-organ ultrasound analysis based on evolutionary diffusion sampling (EvoDiff), mainly includes the following steps: 1. Data preprocessing and dataset construction; 2. Diffusion model construction and pre-training; 3. Combining information gain gradient guidance and evolutionary algorithm diffusion sampling; 4. Sample selection and labeling based on cosine similarity matching; 5. Training of downstream diagnostic models.

[0045] See Figure 2 As shown below, the workflow of the EvoDiff framework of this invention will be described in detail.

[0046] Step 1) Data preprocessing and dataset construction: We collected and preprocessed DICOM format multi-organ ultrasound images covering multiple anatomical sites to construct a heterogeneous dataset containing unlabeled, labeled, and test sets.

[0047] The method for constructing the heterogeneous dataset is as follows: First, medical ultrasound image data covering six organs—appendix, breast, carotid artery, liver, ovary, and thyroid—were collected from multiple different medical centers. The multi-organ ultrasound images were in DICOM format, and the tag field of the DICOM header file contained information tags describing the corresponding multi-organ ultrasound images and the relevant attributes of the patients to which they belong. The attribute information included the transmission format of the multi-organ ultrasound images, organ type, acquisition equipment manufacturer, and the patient's gender and age. Then, all acquired medical ultrasound images are preprocessed, such as image size standardization and normalization. For example, medical ultrasound images from different anatomical sites are uniformly adjusted to 256×256 pixels, and the pixel value range is mapped to the [0,1] interval. The medical ultrasound image data from the same medical center are then grouped into corresponding data subsets, and then integrated to obtain the heterogeneous dataset. Each of the data subsets is represented as ,in D i This represents a subset of data from medical center i. Represents medical ultrasound images, The pathological labels indicating association (e.g., benign, nauseous) are classified according to organ type and diagnostic task; Finally, each data subset is randomly divided into an unlabeled set, a labeled set, and a test set according to a certain proportion. The number of attributes in the unlabeled set, the labeled set, and the test set is balanced according to the type of information tags in the DICOM header file of the multi-organ ultrasound images (including gender balance, age group balance, etc.).

[0048] The unlabeled set contains a large number of multi-organ ultrasound images with organ labels but no pathological labels (hereinafter referred to as unlabeled data), which serve as the core data pool for the active learning framework, used for pre-training of the diffusion model and classifier, as well as for sample selection for the active learning framework.

[0049] The annotation set contains a small number of multi-organ ultrasound images (hereinafter referred to as labeled data) with both organ labels (e.g., appendix, breast, carotid artery, liver, ovary, thyroid) and pathological labels (e.g., benign, malignant), which are used for joint fine-tuning of the diffusion model and classifier, as well as training of downstream classification models.

[0050] The test set contains a small number of multi-organ ultrasound images (hereinafter referred to as test data) with both organ labels (e.g., appendix, breast, carotid artery, liver, ovary, thyroid) and pathological labels (e.g., benign, malignant) as an independent evaluation benchmark to evaluate the performance of downstream diagnostic models within the active learning framework.

[0051] The unlabeled set, the labeled set, and the test set collaborate in a closed-loop process of "pre-training-fine-tuning-evaluation," achieving efficient optimization of the multi-organ ultrasound image classification task with limited labeling costs.

[0052] In a preferred embodiment, the unlabeled set, the labeled set, and the test set are divided in a ratio of 80:10:10. That is, The number of unlabeled multi-organ ultrasound images in the dataset accounts for 90% of the total data volume of the heterogeneous dataset, in order to simulate the scenario of limited labeling resources in clinical practice; The number of multi-organ ultrasound images in the labeled set accounts for 10% of the total data volume of the heterogeneous dataset; The number of multi-organ ultrasound images in the test set accounts for 10% of the total data volume of the heterogeneous dataset.

[0053] Step 2) Diffusion model construction and pre-training: A classifier-guided diffusion model is constructed using denoising diffusion implicit models (DDIM) and a guided classifier. The classifier-guided diffusion model is pre-trained using the information tags in the DICOM header file of multi-organ ultrasound images as conditions, and the unlabeled set and the labeled set are used to enable the classifier-guided diffusion model to master the complex feature distribution of multi-organ ultrasound images under different anatomical structures and pathological types.

[0054] The guided classifier employs a downsampling backbone network based on the U-Net architecture, combined with an attention pooling mechanism, to enhance feature discrimination capabilities.

[0055] To reduce sampling randomness and improve computational efficiency, this invention uses the Denoising Diffusion Implicit Model (DDIM) as the generation backbone network to replace the traditional Denoising Diffusion Probabilistic Model (DDPM).

[0056] The diffusion sampling principle of the Denoising Diffusion Implicit Model (DDIM) is as follows: During the forward diffusion process, the data is gradually distributed by fixing the Markov chain. q ( x 0) Add random Gaussian noise: (1); In equation (1), x t Indicates adding t Initial sample after each time step noise x 0, α t = 1- β t Represents the noise variance scheduling factor. β This represents the hyperparameter controlling the increase in noise variance, where at each step, the noise follows a constant. distributed.

[0057] Equation (1) shows that at each time step, its intermediate state is x t It is sampled from a Gaussian distribution, which represents the intermediate state of the previous time step. x t-1 The scaling was applied, and noise was added. Therefore, DDIM can be directly derived from the initial samples. x 0 at any time step t Calculate intermediate states x t Use the following closed expression: (2); In equation (2), Indicates noise.

[0058] During the reverse denoising process, the denoising network adds samples to the initial sample at each step according to the DDIM deterministic sampling trajectory prediction. x 0 The noise in the sample is then gradually reduced through iterative denoising to approximate the initial sample. x 0 .

[0059] DDIM achieves this by constructing a non-Markovian chain that breaks time-step dependencies and allows for navigation in intermediate states. x t Initial noise xT Compared with the initial sample x 0 Establishing a clear and deterministic relationship between them ensures that the sampling trajectory is deterministic, thereby ensuring that the subsequent active learning framework maintains consistency and stability in the sampling trajectory and generated samples guided by the information gain (IG) gradient in the latent space.

[0060] As a preferred embodiment, DDIM adopts the U-Net architecture as the backbone structure of the denoising network. Its encoder consists of four downsampling modules with feature channels of 64, 128, 256 and 512 respectively. Its bottleneck layer uses a self-attention mechanism to capture global contextual information.

[0061] The mathematical expression for DDIM is as follows: (3); (4); In equations (3) and (4), , Indicates the noise The learning approximation, the noise will affect the initial samples x 0 The original image is contaminated to an intermediate state. x t Noisy images can be parameterized using the U-Net architecture of a denoising network. σ t When =0, the sampling trajectory becomes deterministic.

[0062] To make full use of the massive amount of data in the unlabeled set and the small amount of data in the labeled set, the pre-training process is divided into two stages: self-supervised pre-training and fully supervised fine-tuning.

[0063] 1) Self-supervised pre-training phase: Using the information tags in the DICOM header file of multi-organ ultrasound images as conditions, the DDIM and the guided classifier are pre-trained on the unlabeled set using organ tags, respectively. This allows the DDIM and the guided classifier to learn organ-level features (anatomical representations) in multi-organ ultrasound images in a self-supervised manner, thereby enabling them to extract organ features. At the same time, they can reflect the attribute information of the information tags in the DICOM header file of multi-organ ultrasound images (e.g., the difference between male and female organs), laying the foundation for the subsequent identification of pathological features.

[0064] The guided classifier is optimized using the following loss function: (5), In equation (5), Represents the cross-entropy loss function. This represents a classifier.

[0065] The DDIM is optimized using the following loss function: (6); In equation (6), This represents the cross-entropy loss function.

[0066] 2) Full-supervision fine-tuning stage: Using the information tags in the DICOM header file of multi-organ ultrasound images as conditions, the pre-trained guided classifier is integrated into the pre-trained DDIM architecture to form a classifier-guided diffusion model. Joint fine-tuning is performed on the labeled set using pathological tags. In the fine-tuning process, the loss of the organ reconstruction task is mainly considered in the early stage, and the loss of the pathological prediction task is mainly considered in the later stage. This enables the classifier-guided diffusion model to capture subtle pathological features (such as benign / malignant tumors, normal / abnormal organs, etc.), thereby possessing the ability to identify pathological features and providing support for subsequent information gain-guided sampling. Through the pre-training described above, the classifier-guided diffusion model learns the potential feature spatial distribution of multi-organ ultrasound images, providing high-quality feature representations for subsequent sample sampling.

[0067] The specific pre-training method for DDIM is as follows: During the forward diffusion process, based on the information tags in the DICOM header file of multi-organ ultrasound images, DDIM first reads a large number of multi-organ ultrasound images with organ tags but no pathological tags from the unlabeled set, which are recorded as the original images. Then, Gaussian noise is gradually added to the original images according to a Markov chain, so that the original images gradually change from clear to blurry. ， This generates a series of noisy image sequences with different noise levels (gradually blurring from the clear original image to a purely noisy image); this process aims to simulate the degradation of data from clear to noisy, providing a training target for the model; In the reverse denoising process, using the information tags in the DICOM header file of the multi-organ ultrasound image as conditions, the DDIM allows the denoising network to learn the mapping relationship between a series of noisy image sequences with different noise levels generated by the forward diffusion process and the original image. The denoising network is trained to predict the noise added to the original image at each time step (i.e., given the current noisy image and time step, predict the corresponding noise). Starting from the pure noisy image, the noise is gradually subtracted through multiple iterations, thereby restoring a series of noisy images with different noise levels from blurry to clear original images. This process completes the training by minimizing the difference between predicted noise and real noise, enabling the model to learn the potential distribution of the data.

[0068] During the training process, the number of organs is kept balanced in the first 50% of the iterations, and additional samples are added for organs with greater reconstruction loss in the last 50% of the iterations.

[0069] In addition, during the pre-training process of the denoising diffusion implicit model, two different organ ultrasound images can be weighted at the pixel level, and the information tags in the DICOM header files of these two organ ultrasound images can be used as reconstruction conditions, so that the denoising diffusion implicit model can understand different parts in the same organ ultrasound image during the reconstruction task.

[0070] In a preferred embodiment, during the forward diffusion process of the DDIM pre-training, the total number of diffusion steps (time steps) is defined as T=1000, and Gaussian noise is gradually added to the original image according to the Markov chain. The noise scheduling parameter β... t The value is increased linearly from 0.0001 to 0.02 to ensure that the final distribution of the data is close to the standard normal.

[0071] In a preferred embodiment, the pre-training employs the Adam optimizer with an initial learning rate of 1×10⁻⁶. -4 The batch size is 4, and the training lasts for about 200 epochs until convergence.

[0072] Step 3) Combining information gain gradient guidance and evolutionary algorithm diffusion sampling: The unlabeled set is sampled using a diffusion model guided by the classifier. During the sampling process, an information gain (IG) gradient guidance mechanism and a local optimization evolution algorithm (LOEA) are introduced to generate multiple synthetic samples with different guidance intensities, forming a subset of synthetic samples with high information value.

[0073] Specifically, DDIM is responsible for performing diffusion sampling on the unlabeled set. During the diffusion sampling process, the organ labels of the unlabeled set serve as prompts, and the IG gradient is applied to guide the generation process of synthetic samples, thereby avoiding the cumulative error caused by backpropagation. At the same time, the LOEA algorithm is introduced at each time step of the diffusion sampling process to achieve adaptive search of the high-dimensional feature space, thereby generating multiple synthetic samples with gradually increasing synthetic information content, and finally obtaining a subset of synthetic samples with high information value after integration.

[0074] The IG is approximated by the entropy difference between adjacent time steps on the DDIM sampling trajectory of the classifier-guided denoising diffusion implicit model. Further derivation yields the IG gradient, which is the IG gradient with respect to the input samples. x The spatial rate of change represents the direction and intensity of change of IG in the high-dimensional feature space, serving as the core signal guiding the sampling direction.

[0075] The synthesized samples are multi-organ synthetic ultrasound images with pseudo-organ labels (such as thyroid, breast, carotid artery, liver, appendix, ovary) and pseudo-pathological labels (such as benign / malignant, normal / abnormal, etc.), simulating the characteristics of medical ultrasound images in real clinical scenarios. These synthetic samples are generated through IG gradient-guided sampling and the LOEA algorithm, possessing the most critical features for improving model performance (such as high uncertainty and low redundancy representativeness), reflecting subtle pathological differences in tissue texture and lesion morphology of different organs, and maintaining a high degree of distributional consistency with the unlabeled set in the feature space.

[0076] These synthetic samples, with their gradually increasing amount of synthetic information, can serve as a learning strategy for progressively training downstream diagnostic models from simple to complex, or they can be retained and directly used to expand the labeled set. The classifier-guided diffusion model, acting as a data augmentation module, continuously provides high-quality synthetic samples.

[0077] Since accurately calculating the Information Gain Gradient (IG) in high-dimensional space is computationally infeasible, this invention utilizes the characteristics of the DDIM sampling process to approximate the IG calculation through the entropy difference between adjacent time steps. In the deterministic sampling trajectory of DDIM, the entropy difference between adjacent time steps quantifies the amount of information provided by the sample during the diffusion sampling process, reflecting the potential value of the sample in reducing the uncertainty of model prediction. This serves as a fundamental indicator for subsequent calculation of the information gain gradient and the construction of the hidden guiding field. This approximate calculation method is theoretically supported by the non-Markov property of DDIM, which ensures a consistent relationship between intermediate states and variables during the denoising process. This allows us to capture samples with high information value without explicitly calculating the difficult-to-handle posterior distribution.

[0078] In each sampling step (50 time steps apart), the guided classifier calculates the entropy value of the current prediction. The entropy value is calculated based on the noise distribution predicted by DDIM, and is estimated using the feature activation values ​​normalized by softmax. The calculation formula is as follows: (7); In equation (7), Indicates frozen classifier f θ ( x , y i ) in input x Lower output category y i The softmax probability, Indicates the classifier output y In the input x The entropy value is given by C, where C represents the total number of categories.

[0079] The approximate objective function for IG is as follows: (8); In equation (8), t Indicates a time step. y i Indicates organ labeling, θ Indicates the parameters of the guiding classifier. x t Indicates time step t The intermediate state. x t-1 Indicates time step t-1 The intermediate state. Indicates intermediate state x t The corresponding entropy value, Indicates intermediate state x t-1 The corresponding entropy value.

[0080] The objective function described above quantifies the entropy change during the sampling process, reflecting the intermediate state at time step t. x t Its independence.

[0081] To incorporate the above mechanism into diffusion sampling, this invention defines a sampling gradient guidance function without backpropagation. : (9); In equation (9), IG( x ) represents information gain, and λ represents the guiding strength parameter (λ∈1,3,5,7,9).

[0082] Ultimately, this invention constructs an implicit guiding field based on the IG gradient, which biases the high-dimensional sampling space of DDIM towards the gradient direction that generates the maximum IG gradient. This innovative method accumulates information gain along the DDIM sampling trajectory to obtain synthetic samples with increased information gain, achieving sampling without backpropagation and thus efficiently traversing the algorithm's solution space. As a differentiable control mechanism, this implicit guiding field can systematically redirect the generative trajectory towards regions with higher information value, eliminating the computational overhead required for explicit gradient backpropagation across the entire model.

[0083] This invention's method adaptively guides sampling in the continuous latent space of a diffusion model, optimizing along the IG gradient direction. Unlike traditional adaptive learning methods that discretely search for high-uncertainty samples in a fixed dataset, this invention's method effectively constructs an implicit guiding field without relying on backpropagation, directing the generated results towards the direction of maximum IG. The final generated results will be biased towards those samples that maximize model performance improvement after labeling.

[0084] Evolutionary algorithms are optimization algorithms that simulate natural selection by iteratively optimizing solutions through operations such as selection, crossover, and mutation. By establishing a mathematical connection between diffusion models and evolutionary algorithms, and utilizing the noise addition and denoising properties of diffusion models, evolutionary algorithms can accurately optimize solutions within the parameter space.

[0085] This invention proposes a Local Optimization Evolutionary Algorithm with Class Constraints (LOEA), which generates high-entropy, label-consistent samples based on samples guided by global information gain, in order to find the optimal solution (such as the optimal subset of labeled samples) in a high-dimensional complex parameter space (such as the potential solution space).

[0086] The LOEA algorithm of this invention avoids convergence to local optima by adding controlled noise to the sampling process, ensuring that the selected samples can maximize the information gain of the downstream classifier.

[0087] As a preferred embodiment, the specific operation of the LOEA algorithm of the present invention is as follows: (1) Population initialization: An initial population is randomly generated from the input set of unlabeled ultrasound samples. The unlabeled samples cover different organ structures and scanning angles to ensure the diversity of candidate samples in terms of anatomical morphology and imaging mode. The initial population size is set to 100 individuals, each individual representing a subset of candidate samples, which consists of ultrasound images of multiple organs. The number of generations is set to 50 to fully explore the distribution of discriminative features among different organs. (2) Fitness evaluation: The entropy value of each individual is calculated by guiding the classifier. The higher the entropy value, the stronger the uncertainty of the sample under different organ structures or pathological morphologies, and the stronger the challenge to model training. At the same time, the high-dimensional semantic features of each individual are extracted by the feature extractor and combined with the organ-level feature embedding space. The cosine similarity is calculated with the feature database of the labeled set. Individuals that are highly similar to existing samples are penalized to reduce redundancy. Finally, the entropy value and similarity penalty are combined to form the fitness function. (3) Selection operation: A tournament selection strategy is adopted, in which 5 individuals are randomly selected for comparison each time, and the individual with the highest fitness is retained each time. Through multiple rounds of selection operations, half of the individuals with the best fitness scores are retained as the parent population to maintain the discrimination diversity of the population in the multi-organ semantic space. (4) Crossover operation: Randomly select parent individuals pairs and use uniform crossover to recombine them at the sample subset level so that the offspring individuals can integrate sample features from different imaging modes. The crossover probability is set to 0.8. (5) Mutation operation: With a mutation probability of 0.1 and a standard deviation of 0.1, a controlled perturbation is added to the offspring individuals. By randomly replacing or perturbing some samples to select weights, the distribution perturbation caused by changes in probe angle and imaging quality in real clinical collection is simulated, thus avoiding the search process from getting trapped in local optima. (6) Elite retention strategy: The top 10% of individuals with the highest fitness in the previous generation are retained and directly enter the next generation to maintain the stable inheritance of high-value organ samples; when the improvement of the optimal fitness in 10 consecutive generations is less than 0.1%, the algorithm is judged to have converged and the evolution process is terminated in advance.

[0088] After 50 generations, the Local Optimization Evolutionary Algorithm (LOEA) returns the individual with the highest fitness score, which represents the optimal solution point in the parameter space that maximizes the fitness score.

[0089] The LOEA of this invention can efficiently identify informational samples without gradient calculation through a classifier, thereby effectively navigating in a complex search space and discovering samples that challenge the classifier while maintaining differences from existing data.

[0090] Step 4) Calculate the cosine similarity between each synthetic sample in the synthetic sample subset and the corresponding real sample in the unlabeled set in the traditional convolutional neural network feature latent space, and select the real sample with the most information content for expert annotation to form an active annotation set.

[0091] In each round of active learning iteration, by increasing the intensity of the information gain-guided generation process, the classifier-guided diffusion model synthesizes multiple synthetic samples with different difficulty levels. The information content of these synthetic samples gradually increases, reflecting the confidence of the downstream diagnostic model in these data. To map the sample selection in the synthetic space back to the real data space, this invention employs a principle-based batch-level selection mechanism (batch size = 10). Specifically, the cosine similarity between each synthetic sample and all corresponding real samples in the unlabeled set in the latent space of the traditional convolutional neural network features is calculated.

[0092] The formula for calculating the cosine similarity is as follows: (10); In equation (10), .

[0093] Simultaneously, the similarity of information tags in the diagnostic report corresponding to the ultrasound image under a large text model is considered, and the two calculated similarities are weighted and averaged to obtain a comprehensive similarity.

[0094] Then, for each synthetic sample, the real sample with the highest overall similarity is selected as a candidate. All candidate real samples are ranked according to the overall selection of the corresponding synthetic sample. The top α% of real samples (α value is dynamically adjusted according to the annotation budget) are submitted to experts for independent annotation or summary annotation (the higher the real sample is, the more doctors need to perform expert summary annotation at the same time), thus forming an active annotation set. Finally, the active annotation set is integrated into the existing annotation set to enrich the training data for subsequent optimization of the downstream diagnostic model.

[0095] By strategically selecting samples to iteratively enhance the labeled set, this method significantly improves model performance while effectively reducing labeling costs, which is a key advantage in the field of medical imaging where expert labeling resources are particularly scarce.

[0096] Steps 2-4 will iteratively apply to each subset of data in the heterogeneous dataset.

[0097] Step 5) Downstream diagnostic model training: In each round of active learning iteration, the downstream diagnostic model is trained and fine-tuned on a combined dataset of the labeled set and the active labeled set.

[0098] Because the synthetic samples, with their gradually increasing information content, serve as the learning strategy for progressively training the downstream diagnostic model from simple to complex, each round of active learning iterations can begin by selecting relatively certain real samples for labeling and training the downstream diagnostic model. As the capabilities of the downstream diagnostic model improve, it gradually transitions to more challenging and uncertain real samples. This progressive learning strategy creates a more natural learning process for the downstream diagnostic model, significantly improving learning efficiency in scenarios with limited data.

[0099] The active learning framework of this invention supports two architectures: traditional convolutional neural networks and base models. After completing the training of the downstream diagnostic model in the current round, it is determined whether the preset number of iteration rounds or performance target has been reached. If all iterations have been completed, the final verification stage is entered, where the performance of the downstream diagnostic model within the active learning framework is evaluated on the test set and external verification dataset. Evaluation metrics include accuracy, sensitivity, and specificity.

[0100] Compared with traditional active learning methods, this invention has at least the following innovations: 1. Compared with existing active learning methods, this invention is the first active learning data filtering method based on sample information content that integrates generative models on a pan-organ ultrasound dataset.

[0101] 2. Compared with the prior art, the present invention approximates the calculation of information content with the difference of entropy in the sampling process of the diffusion model, thereby constructing the theoretically optimal solution for active learning sample selection, and ensuring the performance improvement of the optimal downstream diagnostic model with the minimum amount of labeling.

[0102] 3. Compared with the diffusion model technique in existing generative algorithms, this invention uses the gradient of information content as the direction of sample denoising and introduces an evolutionary algorithm to add noise during the diffusion process to avoid the denoising process falling into local optima.

[0103] 4. Compared with existing active learning methods, this invention defines the sample selection strategy for the first time as finding a subset of the optimal solution based on information content in a high-dimensional space, thereby optimizing the feature space confusion problem in the pan-organ ultrasound dataset.

[0104] This invention integrates diffusion models, information gain gradient guidance, evolutionary algorithm optimization, and similarity selection to construct an efficient backpropagation-free active learning framework. This framework achieves efficient and accurate sample selection, demonstrating superior performance in medical image classification tasks, significantly outperforming methods based on uncertainty and random sampling. Its core advantage lies in its ability to globally optimize labeled subsets in a high-dimensional feature space, thereby significantly reducing labeling costs and improving model performance.

[0105] The active learning framework of this invention has demonstrated its superior performance on a comprehensive multi-organ ultrasound dataset covering six organs from nine public datasets. Experimental results show that the active learning framework of this invention outperforms existing active learning methods in terms of accuracy, F1 score, precision, and area under the curve (AUC), and in particular, exhibits stronger information sample selection ability in complex and imbalanced feature spaces.

[0106] The aforementioned active learning method for pan-organ ultrasound analysis based on evolutionary diffusion sampling can be used to solve the sample selection problem in multi-organ ultrasound image analysis tasks.

[0107] Specifically, the method can be applied to the following medical image analysis tasks: (1) Ultrasound diagnosis of appendix-related diseases, including the identification and discrimination of inflammatory lesions such as appendicitis, by actively screening samples with blurred boundaries or structural abnormalities, to improve the model’s ability to distinguish between acute and atypical manifestations. (2) Breast tumor identification and diagnosis task, including classification and structural feature analysis of benign and malignant masses. By optimizing the selection of samples with different echo patterns and morphological differences, the model’s ability to express the diversity of breast lesions is improved. (3) Ultrasound analysis tasks related to cardiovascular event risk prediction, including modeling the structural and functional features of cardiac and related vascular ultrasound images, and actively learning to mine uncertain samples that contribute highly to prognostic judgment, thereby enhancing the model's ability to predict the risk of cardiovascular events. (4) Fatty liver diagnosis task, including the analysis and grading of liver echo characteristics under different degrees of fat deposition, and improving the robustness of the model to the grading of fatty degeneration by screening representative intermediate state samples. (5) Ovarian tumor assessment task, including the judgment of benign or malignant ovarian masses and structural complexity analysis. By introducing samples with complex morphology and echo patterns, the model’s ability to distinguish diverse pathological types is enhanced. (6) Thyroid nodule diagnosis task, including nodule benign and malignant identification and risk stratification analysis. By actively selecting samples with high uncertainty such as blurred boundaries and complex internal echoes, the stability and accuracy of the model in clinical screening scenarios are improved.

[0108] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A pan-organ ultrasound analysis active learning method based on evolutionary diffusion sampling, characterized in that, Includes the following steps: Step 1) Collect and preprocess DICOM format multi-organ ultrasound images covering multiple anatomical sites to construct a heterogeneous dataset containing an unlabeled set, a labeled set, and a test set; Step 2) A classifier-guided diffusion model is constructed using a denoised diffusion implicit model and a guided classifier. The classifier-guided diffusion model is pre-trained using the information tags of the DICOM header file of the multi-organ ultrasound image as a condition, and the unlabeled set and the labeled set are used to master the complex feature distribution of multi-organ ultrasound images under different anatomical structures and pathological types. Step 3) The unlabeled set is sampled using the diffusion model guided by the classifier. During the sampling process, an information gain gradient guidance mechanism and a local optimization evolution algorithm are introduced to generate multiple synthetic samples with different guidance intensities, and a subset of synthetic samples with high information value is formed. Step 4) Calculate the cosine similarity between each synthetic sample in the synthetic sample subset and the corresponding real sample in the unlabeled set in the traditional convolutional neural network feature latent space, and select the real sample with the most information content for expert annotation to form an active annotation set; Step 5) In each round of active learning iteration, the downstream diagnostic model is trained and fine-tuned on the combination of the labeled set and the active labeled set.

2. The organ-agnostic ultrasound analysis active learning method based on evolutionary diffusion sampling of claim 1, wherein, The method for constructing the heterogeneous dataset is as follows: First, medical ultrasound image data covering six organs—appendix, breast, carotid artery, liver, ovary, and thyroid—were collected from multiple different medical centers. The multi-organ ultrasound images were in DICOM format, and the tag field of the DICOM header file contained information tags describing the relevant attributes of the corresponding multi-organ ultrasound images and their respective patients. The attribute information included the transmission format of the multi-organ ultrasound images, organ type, acquisition equipment manufacturer, and the patient's gender and age. Then, the multi-organ ultrasound images are preprocessed by image size standardization and normalization. Data from the same medical center are then grouped into corresponding data subsets and integrated to obtain the heterogeneous dataset. Finally, each data subset is randomly divided into an unlabeled set, a labeled set, and a test set according to a certain proportion. Furthermore, the number of attributes in each of the unlabeled set, the labeled set, and the test set is balanced according to the types of information tags in the DICOM header file of the multi-organ ultrasound images. The unlabeled set contains a large number of multi-organ ultrasound images with organ labels but no pathological labels, which serve as the core data pool for the active learning framework, used for pre-training of the diffusion model and classifier, as well as for sample selection for the active learning framework. The annotation set contains a small number of multi-organ ultrasound images with both organ and pathology labels, which are used for joint fine-tuning of the diffusion model and classifier, as well as training of the downstream separation model. The test set contains a small number of multi-organ ultrasound images with both organ and pathology labels, serving as an independent evaluation benchmark to assess the performance of downstream diagnostic models within the active learning framework.

3. The organ-agnostic ultrasound analysis active learning method based on evolutionary diffusion sampling of claim 1, wherein, The pre-training process is divided into two stages: self-supervised pre-training and fully supervised fine-tuning. Specifically: 1) Self-supervised pre-training phase: Using the information tags in the DICOM header file of multi-organ ultrasound images as conditions, the denoising diffusion implicit model and the guided classifier are pre-trained on the unlabeled set using organ tags, respectively. This allows the denoising diffusion implicit model and the guided classifier to learn organ-level features in multi-organ ultrasound images in a self-supervised manner, thereby enabling them to extract organ features. At the same time, they can reflect the attribute information of the information tags in the DICOM header file of multi-organ ultrasound images, laying the foundation for subsequent identification of pathological features. 2) Full-supervision fine-tuning stage: Using the information tags in the DICOM header file of multi-organ ultrasound images as conditions, the pre-trained guided classifier is integrated into the architecture of the pre-trained denoising diffusion implicit model to form a classifier-guided diffusion model. Joint fine-tuning is performed on the labeled set using pathological labels. In the fine-tuning process, the loss of the organ reconstruction task is mainly focused in the early stage, and the loss of the pathological prediction task is mainly focused in the later stage. This enables the classifier-guided diffusion model to capture subtle pathological features, thereby possessing the ability to identify pathological features and providing support for subsequent information gain-guided sampling. Through the pre-training described above, the classifier-guided diffusion model learns the potential feature spatial distribution of multi-organ ultrasound images, providing high-quality feature representations for subsequent sample sampling.

4. The active learning method for pan-organ ultrasound analysis based on evolutionary diffusion sampling according to claim 3, characterized in that, The pre-training method for the denoising diffusion implicit model is as follows: During the forward diffusion process, using the information tags in the DICOM header file of multi-organ ultrasound images as conditions, the denoising diffusion implicit model first reads a large number of multi-organ ultrasound images with organ tags but no pathological tags from the unlabeled set, which are recorded as the original images. Then, the total number of diffusion steps is defined as 1000, and Gaussian noise is added to the original images step by step according to the Markov chain. The noise scheduling parameter is linearly increased from 0.0001 to 0.02, so that the original image gradually changes from clear to blurry. ， This generates a series of noisy image sequences with different noise levels; In the reverse denoising process, using the information tags in the DICOM header file of the multi-organ ultrasound image as conditions, the denoising diffusion implicit model allows the denoising network to learn the mapping relationship between a series of noise image sequences with different noise levels generated by the forward diffusion process and the original image. The denoising network is trained to predict the noise added to the original image at each time step in reverse, and then starts from the pure noise image again, gradually subtracting the noise through multiple iterations, thereby restoring a series of noise images with different noise levels from blurry to clear original image. During the training process, in the first 50% of the iterations, the number of organs is kept balanced. In the last 50% of the iterations, additional samples are added for organs with greater reconstruction loss. In addition, during the pre-training process of the denoising diffusion implicit model, two different organ ultrasound images can be weighted at the pixel level, and the information tags in the DICOM header files of these two organ ultrasound images can be used as reconstruction conditions, so that the denoising diffusion implicit model can understand different parts in the same organ ultrasound image during the reconstruction task.

5. The organ-agnostic ultrasound analysis active learning method based on evolutionary diffusion sampling of claim 1, wherein, The diffusion sampling process is as follows: The denoising diffusion implicit model is responsible for diffusion sampling of the unlabeled set. During diffusion sampling, the organ labels of the unlabeled set serve as cues, and information gain gradients are applied to guide the generation of synthetic samples to avoid the cumulative error caused by backpropagation. Simultaneously, a local optimization evolutionary algorithm is introduced at each time step of the diffusion sampling process to achieve adaptive search of the high-dimensional feature space, thereby generating multiple synthetic samples with gradually increasing synthetic information content. After integration, a subset of synthetic samples with high information value is finally obtained. The information gain is approximated by the entropy difference between adjacent time steps on the DDIM sampling trajectory of the classifier-guided denoising diffusion implicit model. The information gain gradient is derived from the calculated information gain and is the information gain relative to the input sample. x The spatial rate of change represents the direction and intensity of information gain change in the high-dimensional feature space, serving as the core signal guiding the sampling direction. The synthetic sample is a multi-organ synthetic ultrasound image with pseudo-organ labels and pseudo-pathological labels, which simulates the medical ultrasound image characteristics in real clinical scenarios. The synthesized samples can be used as a learning strategy for training downstream diagnostic models from simple to complex, or they can be retained and used directly to expand the annotation set.

6. The organ-agnostic ultrasound analysis active learning method based on evolutionary diffusion sampling of claim 1, wherein, The objective function for approximating the information gain is as follows: （8）； In equation (8), t Indicates a time step. y i Indicates organ labeling, θ Indicates the parameters of the guiding classifier. x t Indicates time step t The intermediate state. x t-1 Indicates time step t-1 The intermediate state. Indicates intermediate state x t The corresponding entropy value, Indicates intermediate state x t-1 The corresponding entropy value.

7. The active learning method for pan-organ ultrasound analysis based on evolutionary diffusion sampling according to claim 1, characterized in that, The information gain gradient guidance mechanism is implemented by an implicit guidance field based on the gain information gradient. This implicit guidance field biases the DDIM high-dimensional sampling space towards the gradient direction that generates the maximum gain information gradient. The implicit guidance field based on the gain information gradient contains a sampling gradient guidance function without backpropagation. （9）； In equation (9), IG( x ) represents information gain, and λ represents the guiding strength parameter, λ∈1,3,5,7,9.

8. The organ-agnostic ultrasound analysis active learning method based on evolutionary diffusion sampling of claim 1, wherein, The specific operation of the local optimization evolutionary algorithm is as follows: (1) Population initialization: An initial population is randomly generated from the input set of unlabeled ultrasound samples. The unlabeled samples cover different organ structures and scanning angles to ensure the diversity of candidate samples in terms of anatomical morphology and imaging mode. The initial population size is set to 100 individuals, each individual representing a subset of candidate samples, which consists of ultrasound images of multiple organs. The number of generations is set to 50 to fully explore the distribution of discriminative features among different organs. (2) Fitness evaluation: The entropy value of each individual is calculated by guiding the classifier. The higher the entropy value, the stronger the uncertainty of the sample under different organ structures or pathological morphologies, and the stronger the challenge to model training. At the same time, the high-dimensional semantic features of each individual are extracted by the feature extractor and combined with the organ-level feature embedding space. The cosine similarity is calculated with the feature database of the labeled set. Individuals that are highly similar to existing samples are penalized to reduce redundancy. Finally, the entropy value and similarity penalty are combined to form the fitness function. (3) Selection operation: A tournament selection strategy is adopted, in which 5 individuals are randomly selected for comparison each time, and the individual with the highest fitness is retained each time. Through multiple rounds of selection operations, half of the individuals with the best fitness scores are retained as the parent population to maintain the discrimination diversity of the population in the multi-organ semantic space. (4) Crossover operation: Randomly select parent individuals pairs and reorganize them at the sample subset level using a uniform crossover method so that the offspring individuals can integrate sample features from different imaging modes. The crossover probability is set to 0.

8. (5) Mutation operation: With a mutation probability of 0.1 and a standard deviation of 0.1, a controlled perturbation is added to the offspring individuals. By randomly replacing or perturbing some samples to select weights, the distribution perturbation caused by changes in probe angle and imaging quality in real clinical collection is simulated, thus avoiding the search process from getting trapped in local optima. (6) Elite retention strategy: The top 10% of individuals with the highest fitness in the previous generation are retained and directly enter the next generation to maintain the stable inheritance of high-value organ samples; when the improvement of the optimal fitness in 10 consecutive generations is less than 0.1%, the algorithm is considered to have converged and the evolution process is terminated in advance.

9. The organ-agnostic ultrasound analysis active learning method based on evolutionary diffusion sampling of claim 1, wherein, The method for sample selection and labeling is as follows: First, the cosine similarity between each synthetic sample and all real samples in the unlabeled set is calculated in the latent space of the traditional convolutional neural network feature. At the same time, the similarity of the information labels in the diagnostic report corresponding to the ultrasound image is considered in a large text model. The two similarities are then weighted and averaged to obtain a comprehensive similarity. Then, for each synthetic sample, the real sample with the highest overall similarity is found as a candidate, and all candidate real samples are sorted according to the overall similarity value of their corresponding synthetic samples; the top α% of real samples are selected and submitted to experts for independent annotation or summary annotation, thereby forming an active annotation set; Finally, the active annotation set is integrated into the existing annotation set to enrich the training data for subsequent optimization of the downstream diagnostic model.

10. The application of the active learning method for pan-organ ultrasound analysis based on evolutionary diffusion sampling as described in any one of claims 1-9 in sample screening for multi-organ ultrasound image analysis tasks.

Images