Gynecological malignant tumor intelligent diagnosis model generation method and system
By calculating sample density distribution, performing confidence assessment and feature decoupling and recombination in the AI diagnostic model for gynecological malignancies, and embedding active learning, the problem of local optima in model training was solved, improving the ability to identify early lesions and the robustness of the model.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- ZHUHAI HOSPITAL OF INTEGRATED TRADITIONAL CHINESE & WESTERN MEDICINE (ZHUHAI SECOND PEOPLES HOSPITAL)
- Filing Date
- 2025-12-26
- Publication Date
- 2026-05-01
- Estimated Expiration
- Not applicable · inactive patent
Smart Images
Figure CN121964161A_ABST
Abstract
Description
Methods and Systems for Generating Intelligent Diagnostic Models for Gynecological Malignant Tumors Technical Field
[0001] This invention relates to the field of neural network model training technology, and in particular to a method and system for generating intelligent diagnostic models for gynecological malignant tumors. Background Technology
[0002] In the field of AI-assisted diagnosis of gynecological malignancies, although deep learning technology has made significant progress in medical image analysis, existing models still face fundamental challenges in terms of clinical operability.
[0003] Specifically, early lesions of gynecological malignancies often exhibit a sparse distribution pattern with non-uniform aggregation in a multidimensional feature space. Conventional methods often mechanically equate these sparse regions with noise or low-value data, failing to recognize that such samples may precisely contain key discriminative information for precancerous lesions or early cancerous changes.
[0004] A deeper technical challenge lies in the fact that the difficulty in distinguishing these samples is not an inherent attribute, but rather exhibits a bidirectional transfer characteristic accompanying the dynamic evolution of the model's representational capabilities. Some difficult cases misclassified by the model in the early stages of training may become accurately distinguishable samples after the model learns specific pathological feature expression patterns; conversely, some typical cases that are easily identified in the early stages may reveal new identification difficulties during the process of deep feature mining. Therefore, existing technologies lack a mechanism for perceiving and modeling this dynamic evolution, causing model training to get stuck in local optima of static learning and unable to continuously break through the bottleneck of identifying cases with ambiguous boundaries. Summary of the Invention
[0005] This invention provides a method and system for generating intelligent diagnostic models for gynecological malignant tumors, in order to solve the problem that the model training is stuck in local optima of static learning and cannot continuously break through the bottleneck of identifying cases with fuzzy boundaries.
[0006] To address the aforementioned problems, the present invention adopts the following technical solution: The present invention provides a method for generating an intelligent diagnostic model for gynecological malignant tumors, comprising: acquiring training samples of gynecological malignant tumors; calculating the local density distribution of the training samples in the feature space, selecting samples with a density lower than a preset density as first training samples based on the local density distribution, evaluating the confidence level of the first training samples, and selecting the first training samples with a confidence level lower than a confidence level threshold as second training samples; generating a sample difficulty sequence based on the dynamic change characteristics of the second training samples, wherein the dynamic change characteristics characterize the regularity of the continuous evolution of sample difficulty exhibited by the second training samples in the feature space; training a neural network model based on the sample difficulty sequence, and performing feature decoupling and recombination during the training process, performing self-supervised training optimization of the neural network model based on the recombined features, and embedding uncertainty-guided active learning of the model during the optimization process, iteratively optimizing the model parameters based on active learning, and generating an intelligent diagnostic model for gynecological malignant tumors.
[0007] Preferably, the gynecological malignant tumor training samples include medical imaging data, pathological slide data, and clinical text data.
[0008] Preferably, generating a sample difficulty sequence based on the dynamic change characteristics of the second training samples includes: initially inputting the second training samples into a neural network model for multiple rounds of iterative training; recording the predicted loss value of each second training sample in each iteration and calculating the difference between the loss values in adjacent rounds to obtain the loss change rate; using the loss change rate as a quantitative indicator of the dynamic change characteristics of the samples; and sorting all the second training samples according to the order of the loss change rate from smallest to largest to generate a sample difficulty sequence.
[0009] Preferably, training the neural network model according to the sample difficulty sequence includes: dividing the second training sample into multiple third training samples with increasing difficulty according to the sample difficulty sequence; training the neural network model starting from the lowest difficulty third training sample; after the neural network model converges to training with the current difficulty third training sample, continuing to train the neural network model with the next lower difficulty third training sample until the highest difficulty third training sample is completed and the neural network model converges.
[0010] Furthermore, after training the neural network model based on the sample difficulty sequence, the method further includes: using the feature vector of the second training sample as graph nodes, calculating the cosine similarity between nodes; when the cosine similarity between two nodes is greater than the similarity threshold, constructing an edge connection between the two nodes; recalculating the cosine similarity between all nodes every preset number of iterations during the training process, and updating the edge connection relationship.
[0011] Preferably, the feature decoupling and recombination during training includes: decomposing the features extracted by the neural network model into lesion features, anatomical structure features and noise features during training; enhancing the lesion features and anatomical structure features, suppressing the noise features, and recombinating the enhanced lesion features and anatomical structure features.
[0012] Preferably, the step of performing self-supervised training optimization on the neural network model based on the recombined features includes: constructing a main diagnostic task and a self-supervised task relative to the main diagnostic task based on the recombined features, calculating the loss of the main diagnostic task and the loss of the self-supervised task respectively; constructing an overall optimization objective based on the loss of the main diagnostic task and the loss of the self-supervised task, and training and optimizing the neural network model according to the overall optimization objective.
[0013] Preferably, the active learning of the uncertainty-guided model embedded in the optimization process includes: filtering out unlabeled training samples from the second training samples to obtain a fourth training sample; predicting uncertainty for the fourth training sample, and selecting the fourth training sample with uncertainty higher than a preset value as a fifth training sample; labeling the fifth training sample and adding it to the second training sample, and retraining the neural network model based on the expanded second training sample until the model converges.
[0014] Preferably, the intelligent diagnostic model for gynecological malignant tumors is used to predict the probability that an input sample belongs to a malignant tumor and the probability of each type of malignant tumor.
[0015] This invention also provides a system for generating an intelligent diagnostic model for gynecological malignant tumors, comprising: an acquisition module for acquiring training samples of gynecological malignant tumors; a calculation module for calculating the local density distribution of the training samples in a feature space, selecting samples with a density lower than a preset density as first training samples based on the local density distribution, evaluating the confidence level of the first training samples, and selecting the first training samples with a confidence level lower than a confidence threshold as second training samples; a generation module for generating a sample difficulty sequence based on the dynamic change characteristics of the second training samples, wherein the dynamic change characteristics characterize the regularity of the continuous evolution of sample difficulty exhibited by the second training samples in the feature space; and a training module for training a neural network model based on the sample difficulty sequence, performing feature decoupling and recombination during the training process, performing self-supervised training optimization on the neural network model based on the recombined features, embedding uncertainty-guided active learning of the model during the optimization process, iteratively optimizing the model parameters based on active learning, and generating an intelligent diagnostic model for gynecological malignant tumors.
[0016] Compared with existing technologies, the technical solution of this invention has at least the following advantages: The intelligent diagnostic model generation method and system for gynecological malignant tumors provided by this invention calculates the local density distribution of training samples in the feature space and filters out low-density samples, achieving accurate identification of potentially high-value samples in sparsely distributed areas of early gynecological malignant tumor lesions. This avoids the loss of effective information caused by misclassifying boundary cases as noise in traditional methods, significantly improving the utilization rate of clinical information in training data. Secondly, by evaluating the confidence level of low-density samples and filtering out second training samples with low confidence, a refined distinction is achieved between real difficult cases and samples with quality defects. This effectively filters out interference factors such as labeling errors, while retaining ambiguous cases that play a crucial role in improving the model's discriminative ability, thus enhancing the challenge of the training samples.
[0017] Furthermore, by generating a sample difficulty sequence based on the dynamic changes of the second training sample, and training the neural network model according to the sample difficulty sequence, the regularity of the evolution of sample discrimination difficulty with the model's cognitive ability is realized. This enables the model to follow a dynamic learning path from basic to complex for progressive training, rather than relying on static data distribution, thus breaking through the performance ceiling of traditional methods in learning difficult cases.
[0018] Furthermore, by decoupling and recombining features during training, and then performing self-supervised training optimization based on the recombined features, background noise interference was effectively suppressed, enabling the model to focus on core discriminative features. Simultaneously, by embedding uncertainty to guide the model's active learning and iteratively optimizing model parameters, the quantitative feedback of the model's own predictive uncertainty was embedded within the training loop. This allows the model to proactively identify knowledge gaps and trigger targeted optimizations, significantly improving the efficiency of computational resource utilization and the model's generalization robustness to atypical cases. Attached Figure Description
[0019] Figure 1 is a flowchart of one embodiment of the intelligent diagnostic model generation method for gynecological malignant tumors according to the present invention; Figure 2 is a flowchart of another embodiment of the intelligent diagnostic model generation method for gynecological malignant tumors according to the present invention; Figure 3 is a structural block diagram of one embodiment of the intelligent diagnostic model generation system for gynecological malignant tumors according to the present invention. Detailed Implementation
[0020] To enable those skilled in the art to better understand the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings.
[0021] Please refer to Figure 1. This invention provides a method for generating an intelligent diagnostic model for gynecological malignant tumors, comprising the following steps: S11, obtaining training samples for gynecological malignant tumors; S12, calculating the local density distribution of the training samples in the feature space, selecting samples with a density lower than a preset density as first training samples based on the local density distribution, evaluating the confidence level of the first training samples, and using the first training samples with a confidence level lower than a confidence threshold as second training samples; S13, generating a sample difficulty sequence based on the dynamic change characteristics of the second training samples, wherein the dynamic change characteristics characterize the regularity of the continuous evolution of sample difficulty exhibited by the second training samples in the feature space; S14, training a neural network model based on the sample difficulty sequence, and performing feature decoupling and recombination during the training process, performing self-supervised training optimization on the neural network model based on the recombined features, embedding uncertainty-guided active learning of the model during the optimization process, iteratively optimizing the model parameters based on active learning, and generating an intelligent diagnostic model for gynecological malignant tumors.
[0022] This embodiment can extract gynecological malignant tumor case data from hospital imaging archives, including but not limited to contrast-enhanced ultrasound, whole-section pathological scan images, and corresponding clinical indicators. The samples must cover different stages, different pathological subtypes, and a benign control group to ensure the clinical authenticity of the data distribution. Preferably, the gynecological malignant tumor training samples include medical imaging data, pathological slide data, and clinical text data.
[0023] Then, the original training samples are mapped to the high-dimensional feature space extracted by the deep neural network, and the k-nearest neighbor algorithm is used to calculate the local density of each sample point, forming a local density distribution. The preset density can be set to 30%-50% of the global average density of the feature space. Sparse regions with densities lower than this preset density are selected as the first training samples. These samples typically correspond to clinically challenging cases with atypical tumor morphology, early small lesions, or severe imaging artifacts, rather than simple noise. Next, the confidence level of the first training samples is evaluated. Samples with confidence levels below a confidence threshold (e.g., 0.5) are further selected as the second training samples to effectively exclude invalid low-density samples due to poor sample quality, retaining boundary cases that present real discrimination challenges.
[0024] During model training, the positional offset of the second training sample in the feature space, the rate of change of prediction loss, and the uncertainty decay gradient are recalculated at fixed intervals to construct a three-dimensional dynamic feature vector. Through time series analysis, the model learns the continuous evolution pattern of sample difficulty as the training process progresses, ultimately generating a sequence of sample difficulty ordered in ascending order of difficulty. For example, an early-stage ovarian cancer sample initially has a difficulty rating of level 9 (out of 10) because the lesion diameter is less than 5mm. As the model learns the microcalcification feature, its difficulty rating is dynamically lowered to level 6, and its weight in subsequent training is adjusted accordingly.
[0025] In the initial stage of model training, low-difficulty samples from the difficulty sequence are used for warm-up. Subsequently, higher-difficulty samples are gradually introduced according to their difficulty level to avoid premature exposure to extremely difficult samples, which could cause gradient oscillations in the model. In addition, a dynamic adjustment strategy is adopted for the learning rate, allocating higher learning weights to higher-difficulty samples.
[0026] Simultaneously, a decoupling module is inserted into the neural network model to decompose the original features into general features related to organ anatomy (such as uterine morphology) and specific features related to tumor malignancy (such as spiculated features at lesion edges). The correlation between the two types of features is constrained by an orthogonal loss function, and then an attention mechanism is used to weightedly reorganize the specific features, thereby enhancing the expression of information related to the tumor microenvironment.
[0027] Auxiliary tasks are constructed using recombined specific features, such as positive sample pair matching or mask image modeling based on contrastive learning. Feature representation capabilities are continuously optimized on unlabeled data to prevent overfitting on difficult samples. In each iteration, the prediction uncertainty of the current model for all unlabeled samples is calculated. The top 30 samples with the highest uncertainty are prioritized and submitted to clinical experts for annotation. The second training sample pool is dynamically updated to embed active learning into the training loop, enabling proactive detection and targeted breakthroughs in model performance boundaries. The model parameters are continuously iterated and optimized, ultimately generating an intelligent diagnostic model for gynecological malignancies.
[0028] Preferably, the intelligent diagnostic model for gynecological malignant tumors is used to predict the probability that an input sample belongs to a malignant tumor and the probability of each type of malignant tumor.
[0029] This invention provides a method for generating an intelligent diagnostic model for gynecological malignant tumors. By calculating the local density distribution of training samples in the feature space and filtering out low-density samples, it achieves accurate identification of potentially high-value samples in sparsely distributed areas of early gynecological malignant tumor lesions. This avoids the loss of effective information caused by misclassifying boundary cases as noise in traditional methods, significantly improving the utilization rate of clinical information in training data. Secondly, by evaluating the confidence of low-density samples and filtering out second training samples with low confidence, it achieves a refined distinction between real difficult cases and samples with quality defects, effectively filtering out interference factors such as labeling errors, while retaining ambiguous cases that play a crucial role in improving the model's discriminative ability, thus increasing the challenge of the training samples.
[0030] Furthermore, by generating a sample difficulty sequence based on the dynamic changes of the second training sample, and training the neural network model according to the sample difficulty sequence, the regularity of the evolution of sample discrimination difficulty with the model's cognitive ability is realized. This enables the model to follow a dynamic learning path from basic to complex for progressive training, rather than relying on static data distribution, thus breaking through the performance ceiling of traditional methods in learning difficult cases.
[0031] Furthermore, by decoupling and recombining features during training, and then performing self-supervised training optimization based on the recombined features, background noise interference was effectively suppressed, enabling the model to focus on core discriminative features. Simultaneously, by embedding uncertainty to guide the model's active learning and iteratively optimizing model parameters, the quantitative feedback of the model's own predictive uncertainty was embedded within the training loop. This allows the model to proactively identify knowledge gaps and trigger targeted optimizations, significantly improving the efficiency of computational resource utilization and the model's generalization robustness to atypical cases.
[0032] In one embodiment, referring to Figure 2, the step of generating a sample difficulty sequence based on the dynamic change characteristics of the second training sample includes: S131, inputting the second training sample into the neural network model for multiple rounds of iterative training, recording the predicted loss value of each second training sample in each iteration, and calculating the difference in loss values between adjacent rounds to obtain the loss change rate; S132, using the loss change rate as a quantitative indicator of the dynamic change characteristics of the sample, sorting all the second training samples according to the order of loss change rate from smallest to largest to generate a sample difficulty sequence.
[0033] First, all the second training samples are input into the neural network model to be trained, and iterative training is performed in multiple consecutive rounds. Iterative training refers to the learning process in which the neural network model completely traverses the training samples in multiple rounds. Each round of forward propagation and backward parameter update of all training samples is called one iteration.
[0034] After each iteration, the predicted loss value for each sample is calculated. Then, the loss values from two adjacent iterations are differentially calculated, i.e., the rate of decrease in the loss value is determined; this difference is the loss change rate. A larger loss change rate indicates a lower match between the sample's inherent discrimination pattern and the model's current knowledge base, classifying it as a high-difficulty sample; conversely, a smaller rate indicates that the sample's features have been well understood.
[0035] Then, using the loss change rate calculated above as the sole quantification indicator, all the second training samples are sorted in ascending order (from smallest to largest). The sorting result directly constitutes the sample difficulty sequence, in which samples located further down the sequence will receive higher sampling weights or earlier intervention opportunities in the next round of training, and the sample difficulty sequence will be automatically updated after each round of complete training, forming a dynamic and adaptive learning path.
[0036] By recording the loss values across multiple iterations and calculating the difference between adjacent iterations, the loss change rate is obtained, enabling fine-grained quantitative capture of the dynamic evolution of sample difficulty and reflecting changes in the model's cognitive state during learning. Secondly, by using the loss change rate as a quantitative indicator for sorting, all second training samples are ranked according to their loss change rate from smallest to largest, generating a sample difficulty sequence. This achieves dynamic matching between training resource allocation and the model's real-time cognitive ability, allowing sample scheduling to respond in real-time to changes in the model's knowledge boundaries. This ensures the model doesn't waste time on already mastered samples and prevents premature introduction of extremely difficult samples that the current knowledge system cannot accommodate, thus preventing training collapse. Consequently, without increasing the amount of labeled data, the detection efficiency and training stability of the intelligent diagnostic model for gynecological malignant tumors in difficult pathological subtypes and early, subtle lesions are significantly improved.
[0037] In one embodiment, training the neural network model according to the sample difficulty sequence includes: dividing the second training sample into multiple third training samples with increasing difficulty according to the sample difficulty sequence; training the neural network model starting from the third training sample with the lowest difficulty; and after the neural network model converges in the training of the third training sample with the current difficulty, continuing to train the neural network model using the third training sample with the next lower difficulty until the training of the third training sample with the highest difficulty is completed and the neural network model converges.
[0038] This embodiment can divide the second training samples into multiple mutually exclusive subsets, i.e., the third training samples, based on the ranking results of the sample difficulty sequence. Each subset corresponds to a specific difficulty level. For example, the sample difficulty sequence can be divided into three subsets according to the range of loss change rate: "hard level" (loss change rate greater than 0.015), "medium level" (loss change rate between 0.010 and 0.015), and "easy level" (loss change rate less than 0.010). This ensures that the sample difficulty encountered by the model in any training stage is evenly distributed, avoiding excessive jumps in difficulty that could cause oscillations in the optimization path.
[0039] At the start of training, only the lowest difficulty third training sample is used for forward propagation and backpropagation parameter updates of the neural network model. During training, the trend of performance metrics (such as AUC value) on the validation set is monitored. When the fluctuation of the metric is less than the preset tolerance (such as 0.001) in 5-10 consecutive iterations, it is determined that the current difficulty level has converged. At this point, the current model parameters are locked as the initial weights for the next stage, and higher-level third training samples are gradually mixed into the current training batch at a preset ratio (such as 30%) to achieve a smooth transition in difficulty levels. This process continues iterating until the highest difficulty third training sample is fully incorporated and the model converges on the global validation set, indicating that the model has systematically mastered the complete discrimination ability from basic to extreme levels.
[0040] By dividing the second training samples into multiple third training samples of increasing difficulty, fine-grained control of the learning strategy is achieved, enabling the model to gradually improve with more subtle difficulty gradients and avoiding gradient vanishing or optimization path oscillations caused by abrupt changes in difficulty. Secondly, after the neural network model converges with the third training samples of the current difficulty, it continues to train with the third training samples of the next difficulty until the highest difficulty third training samples are used and the neural network model converges. This achieves controllable expansion of the model's capability boundaries and effectively prevents feature confusion and catastrophic forgetting caused by the premature introduction of difficult samples.
[0041] Furthermore, after training the neural network model based on the sample difficulty sequence, the method further includes: using the feature vector of the second training sample as graph nodes, calculating the cosine similarity between nodes; when the cosine similarity between two nodes is greater than the similarity threshold, constructing an edge connection between the two nodes; recalculating the cosine similarity between all nodes every preset number of iterations during the training process, and updating the edge connection relationship.
[0042] This embodiment can extract the feature vector (typically 512 or 1024 dimensions) of each second training sample, use this feature vector as a node in a graph neural network, and calculate the cosine similarity between every two nodes. The node's attributes not only include the original feature values, but also carry clinical label information such as the pathological stage and histological type of the sample, so as to reconstruct the discrete training samples into a topologically connected graph structure.
[0043] When the cosine similarity between two nodes is greater than a similarity threshold, an undirected edge is established between the two nodes. The weight of the edge can be set to the cosine similarity itself, thus constructing a similarity network for difficult samples. This clusters samples with similar discrimination patterns, providing the model with structured constraint information between samples.
[0044] After each preset number of iterations (e.g., 10 rounds), since the model parameters have been updated and the feature representations have changed, it is necessary to re-extract the current feature vectors of all the second training samples and recalculate the cosine similarity between nodes. Based on the new cosine similarity, dynamic edge addition and deletion operations are performed: new edges with cosine similarity exceeding the threshold but not previously connected are added, and old edges that are below the threshold are deleted. This allows the graph structure to reflect in real time the model's cognitive deepening process of discriminating difficult sample patterns. For example, samples that were misclassified as similar in the early stages may be disconnected because the model has learned specific features, and vice versa.
[0045] In one embodiment, the feature decoupling and recombination during training includes: decomposing the features extracted by the neural network model into lesion features, anatomical structure features, and noise features during training; enhancing the lesion features and anatomical structure features, suppressing the noise features, and recombinating the enhanced lesion features and anatomical structure features.
[0046] This embodiment decomposes the features extracted by the neural network model into lesion features, anatomical structure features, and noise features during training. Lesion features are multiplied by a large weighting coefficient (e.g., 1.5-2.0), and a gradient amplification strategy is employed during backpropagation to give them higher priority in parameter updates. Simultaneously, stability constraints are applied to anatomical structure features to suppress drastic fluctuations and ensure the model's consistent understanding of normal anatomical structures. Noise features are subject to weight attenuation or feature masking, and their contribution is directly set to zero in subsequent calculations, thereby completely removing interfering information.
[0047] The enhanced lesion features and anatomical structure features are recombined, and the fusion weights of the two types of features are dynamically calculated. For example, when the lesion features show high confidence, the weight of the anatomical structure features is reduced to highlight the lesion itself; when the lesion features are ambiguous, the weight of the anatomical structure features is increased to utilize context to assist localization. The final output is the recombined features, achieving a balance between specificity and robustness.
[0048] In one embodiment, the step of performing self-supervised training optimization on the neural network model based on the recombined features includes: constructing a main diagnostic task and a self-supervised task relative to the main diagnostic task based on the recombined features, calculating the loss of the main diagnostic task and the loss of the self-supervised task respectively; constructing an overall optimization objective based on the loss of the main diagnostic task and the loss of the self-supervised task, and training and optimizing the neural network model according to the overall optimization objective.
[0049] In this embodiment, auxiliary tasks that are relative to or complementary to the main diagnostic task are designed as self-supervised tasks. These tasks utilize only the structural information of the reconstructed features themselves and require no additional annotation. The relativity can be reflected in the fact that the task objective and the main diagnostic task form a constraint-versus-response relationship or complementary information.
[0050] The primary diagnostic task refers to the core task of supervised learning using labeled data (pathological results) samples, with the goal of accurately identifying gynecological malignancies. Its loss function reflects the deviation between the model's predictions and the standard, serving as a direct indicator of clinical efficacy. The self-supervised task refers to an auxiliary learning task designed using the inherent structure of the data without requiring manual labeling.
[0051] The two tasks propagate independently and calculate their respective losses. An overall optimization objective is constructed based on the losses of the primary diagnostic task and the self-supervised task. The neural network model is then trained and optimized according to this overall optimization objective, achieving dynamic balance and efficient collaboration in multi-task learning. The loss of the primary diagnostic task directly reflects the model's error rate in clinical discrimination, while the loss of the self-supervised task reflects the quality of feature representation (such as whether denoising is thorough and whether information is complete).
[0052] This embodiment can construct an auxiliary task that is opposite to the main task, forcing the model to learn more robust feature representations, thereby preventing the model from overfitting on difficult samples.
[0053] Preferably, the active learning of the uncertainty-guided model embedded in the optimization process includes: filtering out unlabeled training samples from the second training samples to obtain a fourth training sample; predicting uncertainty for the fourth training sample, and selecting the fourth training sample with uncertainty higher than a preset value as a fifth training sample; labeling the fifth training sample and adding it to the second training sample, and retraining the neural network model based on the expanded second training sample until the model converges.
[0054] In the second training sample, typically only a portion of the samples are labeled, while the remaining samples remain unlabeled due to the high cost of labeling. This embodiment identifies and separates these unlabeled samples to form the fourth training sample. This sample represents potentially high-value data resources within the model's current cognitive boundary and is the target of the active learning mechanism. The selection process relies on the labeled status of the samples, requiring no prediction operations, ensuring that the construction of the fourth training sample pool is efficient and unbiased.
[0055] The fourth training sample is input into the currently trained neural network model, and the uncertainty of each sample is quantified using Bayesian deep learning. Specifically, the Monte Carlo method can be used to perform multiple random forward propagations, and the variance or entropy of the predicted probability distribution can be calculated as a measure of uncertainty. The preset uncertainty threshold is usually set to an entropy greater than 0.6 or a variance greater than 0.2. Samples exceeding this threshold indicate that the model has a high degree of disagreement regarding their malignancy classification, i.e., the model's "most uncertain" or "most perplexing" cases. These samples are considered to contain new discrimination patterns or extreme variations not yet covered by the current model's knowledge system, and their annotation provides the greatest information gain for improving model performance.
[0056] The fifth selected training sample was submitted to senior pathologists or radiologists for manual annotation, which included diagnostic information such as lesion nature, stage, and pathological subtype. The annotated samples were immediately incorporated into the second training sample set, allowing the training set size to dynamically grow and its quality to continuously improve. Subsequently, the model was retrained based on the expanded training set until convergence. This systematically solved the performance bottleneck problem of intelligent diagnostic models for gynecological malignancies under limited annotation budgets, enabling the model to proactively identify and overcome cognitive limitations in difficult pathological subtypes, early minimal lesions, and rare variations. Furthermore, without increasing the total amount of annotated data, it significantly improved the model's discrimination accuracy and robustness in real-world clinical scenarios.
[0057] Furthermore, to enable those skilled in the art to better understand the present invention, the model structure, training process, and related parameters are described in detail below: The neural network model used in the present invention can be a hierarchical feature decoupling network based on the 3D ResNet-50 architecture. The specific module configuration and connection relationship are as follows: Input preprocessing module: used to receive gynecological malignant tumor training samples with dimensions of "N×D×H×W×C", where N is the batch size, D=64 (number of image layers), H=W=224 (single layer resolution), and C=3 (multiple sequence channels).
[0058] Feature extraction backbone network module: It has a built-in residual convolutional block layer, which uses 5 3D residual blocks. Each residual block has a mapping branch to add the input and convolution output element by element to prevent gradient vanishing. Each residual block contains: 3D convolutional layer: 3×3×3 convolutional kernel, stride 1, padding 1, output channel number 64, 128, 256, 512, 1024 respectively; normalization layer: configured with activation function; max pooling layer: pooling kernel 2×2×2, stride 2.
[0059] The feature decoupling module includes: a global average pooling layer that compresses a 1024×4×4×4 feature map into a 1024-dimensional vector; a lesion feature branch that outputs a lesion feature vector; an anatomical structure feature branch that outputs an anatomical feature vector; and a noise feature branch that outputs a noise feature vector. The module also includes a lesion branch connection classification loss, an anatomical branch connection reconstruction loss, and a noise branch connection adversarial loss. The gradients of these three branches are updated synchronously during backpropagation.
[0060] Feature Recombination Module: Performs global average pooling on the lesion feature vector and generates attention weights. These weights are multiplied element-wise with the lesion feature vector to obtain a weighted lesion feature vector. Fusion Layer: Concatenates the weighted lesion feature vector and the anatomical feature vector along the channel dimension to output a 1024-dimensional recombined feature vector.
[0061] The input preprocessing module, feature extraction backbone network module, feature decoupling module and feature reorganization module are connected sequentially. The gradients of the main diagnostic task loss and the self-supervised task loss are superimposed in the fusion layer and flow to the three decoupling paths respectively. The noise branch achieves adversarial optimization through the gradient inversion layer.
[0062] The specific training process and parameter configuration are as follows: Step 1: Training sample acquisition and preprocessing Data source: Export pelvic MRI or CT images from the hospital PACS system to form gynecological malignant tumor training samples.
[0063] Labeling guidelines: Labeling should be done by two gynecological radiologists with more than 10 years of experience in a double-blind manner, and the content should include the polygonal outline of the lesion ROI, pathological subtype or depth of invasion, etc.
[0064] Data partitioning: The training set: validation set: test set = 7:1:2 was randomly partitioned according to patient ID, and images of the same patient were not distributed across sets.
[0065] Step 2: Dual selection mechanism for feature extractor initialization: pre-train 3D ResNet-50 weights, freeze the first 3 convolutional blocks, and fine-tune layers 4-5; Local density calculation: use k-nearest neighbor kernel density to estimate the local density distribution of training samples in the feature space, k=20, the distance metric is Euclidean distance, and the density threshold is global mean × 0.4; Confidence evaluation: Monte Carlo simulation T=20, samples with variance threshold > 0.18 or entropy threshold > 0.65 are selected as the second training samples.
[0066] Step 3: Calculate the loss change rate for generating the dynamic difficulty sequence: Every 5 rounds, record the cross-entropy loss of the second training sample and the difference in loss values between adjacent rounds to obtain the loss change rate; Sort and update: Sort the samples in ascending order of loss change rate, retain the top 80% of the samples to form the dynamic difficulty sequence, and recalculate after each iteration.
[0067] Step 4: Training scheduling difficulty classification: Divided into 3 levels according to the rate of change of loss—hard level (rate of change of loss > 0.015), medium level (rate of change of loss 0.010-0.015), and easy level (rate of change of loss < 0.010); Training model parameters: Stage 1 (easy level): 30 rounds, batch size 16, initial learning rate 0.001 (Adam optimizer, β1=0.9, β2=0.999); Stage 2 (medium level): 50 rounds, batch size 32, learning rate 0.0005 (cosine annealing scheduling); Stage 3 (hard level): 100 rounds, batch size 16, learning rate 0.0001 (early stopping strategy, patience=10); Convergence criterion: The prediction accuracy of the validation set fluctuates less than 0.001 for 10 consecutive rounds.
[0068] Feature decoupling and reconstruction parameter loss weights: main diagnosis loss λ1=1.0, lesion branch loss γ1=2.0, anatomical branch contrast loss γ2=0.07; noise branch adversarial loss γ3=0.1; reconstruction loss λ2=0.3.
[0069] Active learning iterative optimization of model parameter uncertainty threshold: The initial entropy threshold is 0.6, which decreases by 0.02 after each iteration; Query batch: The 15 samples with the highest uncertainty are selected as the fifth training samples in each round; Iteration termination: When the average uncertainty of the fifth training sample is less than 0.3 or the active learning rounds reach 5, the intelligent diagnostic model for gynecological malignant tumors is finally generated.
[0070] Referring to Figure 3, an embodiment of the present invention also provides a system for generating an intelligent diagnostic model for gynecological malignant tumors, comprising: an acquisition module 31 for acquiring training samples of gynecological malignant tumors; a calculation module 32 for calculating the local density distribution of the training samples of gynecological malignant tumors in the feature space, selecting samples with a density lower than a preset density as first training samples based on the local density distribution, evaluating the confidence of the first training samples, and selecting the first training samples with a confidence level lower than a confidence threshold as second training samples; a generation module 33 for generating a sample difficulty sequence based on the dynamic change characteristics of the second training samples, wherein the dynamic change characteristics characterize the regularity of the continuous evolution of sample difficulty exhibited by the second training samples in the feature space; and a training module 34 for training a neural network model based on the sample difficulty sequence, performing feature decoupling and recombination during the training process, performing self-supervised training optimization on the neural network model based on the recombined features, embedding uncertainty-guided active learning of the model during the optimization process, iteratively optimizing the model parameters based on active learning, and generating an intelligent diagnostic model for gynecological malignant tumors.
[0071] This invention provides an intelligent diagnostic model generation system for gynecological malignant tumors. By calculating the local density distribution of training samples in the feature space and filtering out low-density samples, it achieves accurate identification of potentially high-value samples in sparsely distributed areas of early gynecological malignant tumor lesions. This avoids the loss of effective information caused by misclassifying boundary cases as noise in traditional methods, significantly improving the utilization rate of clinical information in training data. Secondly, by evaluating the confidence of low-density samples and filtering out second training samples with low confidence, it achieves a refined distinction between real difficult cases and samples with quality defects, effectively filtering out interference factors such as labeling errors, while retaining ambiguous cases that play a crucial role in improving the model's discriminative ability, thus increasing the challenge of the training samples.
[0072] Furthermore, by generating a sample difficulty sequence based on the dynamic changes of the second training sample, and training the neural network model according to the sample difficulty sequence, the regularity of the evolution of sample discrimination difficulty with the model's cognitive ability is realized. This enables the model to follow a dynamic learning path from basic to complex for progressive training, rather than relying on static data distribution, thus breaking through the performance ceiling of traditional methods in learning difficult cases.
[0073] Furthermore, by decoupling and recombining features during training, and then performing self-supervised training optimization based on the recombined features, background noise interference was effectively suppressed, enabling the model to focus on core discriminative features. Simultaneously, by embedding uncertainty to guide the model's active learning and iteratively optimizing model parameters, the quantitative feedback of the model's own predictive uncertainty was embedded within the training loop. This allows the model to proactively identify knowledge gaps and trigger targeted optimizations, significantly improving the efficiency of computational resource utilization and the model's generalization robustness to atypical cases.
[0074] Regarding the system in the above embodiments, the specific ways in which each module performs operations have been described in detail in the embodiments related to the method, and will not be elaborated here.
[0075] In one embodiment, the present invention also provides a storage medium storing computer-readable instructions, which, when executed by one or more processors, cause the one or more processors to perform the aforementioned method for generating an intelligent diagnostic model for gynecological malignant tumors. The storage medium may be a ROM, random access memory (RAM), CD-ROM, magnetic tape, floppy disk, or optical data storage device, etc.
[0076] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. This computer program can be stored in a storage medium, and when executed, it can include the processes of the embodiments of the methods described above. The aforementioned storage medium can be a non-volatile storage medium such as a magnetic disk, optical disk, or read-only memory (ROM), or random access memory (RAM).
[0077] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.
[0078] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of this patent should be determined by the appended claims.
Claims
1. A method for generating an intelligent diagnostic model for gynecological malignant tumors, characterized in that, include: A training sample of gynecological malignant tumors is obtained; the local density distribution of the training sample in the feature space is calculated, and samples with a density lower than a preset density are selected as first training samples based on the local density distribution. The confidence of the first training sample is evaluated, and the first training sample with a confidence level lower than a confidence threshold is selected as the second training sample. A sample difficulty sequence is generated based on the dynamic change characteristics of the second training sample, wherein the dynamic change characteristics characterize the regularity of the continuous evolution of sample difficulty in the feature space. A neural network model is trained based on the sample difficulty sequence, and feature decoupling and recombination are performed during the training process. The neural network model is then optimized through self-supervised training based on the recombined features, and uncertainty-guided active learning is embedded in the optimization process. The model parameters are iteratively optimized based on the active learning to generate an intelligent diagnostic model for gynecological malignant tumors.
2. The method according to claim 1, characterized in that, The training samples for gynecological malignant tumors include medical imaging data, pathological slide data, and clinical text data.
3. The method according to claim 1, characterized in that, The step of generating a sample difficulty sequence based on the dynamic change characteristics of the second training sample includes: initially inputting the second training sample into a neural network model for multiple rounds of iterative training; recording the predicted loss value of each second training sample in each iteration and calculating the difference between the loss values in adjacent rounds to obtain the loss change rate; using the loss change rate as a quantitative indicator of the dynamic change characteristics of the sample; and sorting all the second training samples according to the order of the loss change rate from smallest to largest to generate a sample difficulty sequence.
4. The method according to claim 1, characterized in that, The step of training the neural network model according to the sample difficulty sequence includes: dividing the second training sample into multiple third training samples with increasing difficulty according to the sample difficulty sequence; training the neural network model starting from the third training sample with the lowest difficulty; and continuing to train the neural network model using the third training sample with the current difficulty after the training of the neural network model has converged, until the training of the third training sample with the highest difficulty has been completed and the neural network model has converged.
5. The method according to claim 1, characterized in that, After training the neural network model based on the sample difficulty sequence, the method further includes: using the feature vector of the second training sample as graph nodes, calculating the cosine similarity between nodes; when the cosine similarity between two nodes is greater than the similarity threshold, constructing an edge connection between the two nodes; during the training process, recalculating the cosine similarity between all nodes every preset number of iterations, and updating the edge connection relationship.
6. The method according to claim 1, characterized in that, The feature decoupling and recombination during training includes: decomposing the features extracted by the neural network model into lesion features, anatomical structure features and noise features during training; enhancing the lesion features and anatomical structure features, suppressing the noise features, and recombinating the enhanced lesion features and anatomical structure features.
7. The method according to claim 1, characterized in that, The step of performing self-supervised training optimization on the neural network model based on the recombined features includes: constructing a main diagnostic task and a self-supervised task relative to the main diagnostic task based on the recombined features, calculating the loss of the main diagnostic task and the loss of the self-supervised task respectively; constructing an overall optimization objective based on the loss of the main diagnostic task and the loss of the self-supervised task, and training and optimizing the neural network model according to the overall optimization objective.
8. The method according to claim 1, characterized in that, The active learning of the uncertainty-guided model embedded in the optimization process includes: filtering out unlabeled training samples from the second training samples to obtain a fourth training sample; predicting uncertainty for the fourth training sample, and selecting the fourth training sample with uncertainty higher than a preset value as a fifth training sample; labeling the fifth training sample and adding it to the second training sample, and retraining the neural network model based on the expanded second training sample until the model converges.
9. The method according to claim 1, characterized in that, The intelligent diagnostic model for gynecological malignant tumors is used to predict the probability that an input sample belongs to a malignant tumor and the probability of each type of malignant tumor.
10. A system for generating intelligent diagnostic models for gynecological malignant tumors, characterized in that, include: The module is used to acquire training samples of gynecological malignant tumors; the module is used to calculate the local density distribution of the training samples of gynecological malignant tumors in the feature space, select samples with a density lower than a preset density as first training samples based on the local density distribution, evaluate the confidence of the first training samples, and select the first training samples with a confidence level lower than a confidence threshold as second training samples. The generation module is used to generate a sample difficulty sequence based on the dynamic change characteristics of the second training sample, wherein the dynamic change characteristics characterize the regularity of the continuous evolution of sample difficulty exhibited by the second training sample in the feature space. The training module is used to train the neural network model according to the sample difficulty sequence, and to perform feature decoupling and recombination during the training process. Based on the recombined features, the neural network model is optimized through self-supervised training. During the optimization process, uncertainty-guided active learning of the model is embedded. Based on the active learning, the model parameters are iteratively optimized to generate an intelligent diagnostic model for gynecological malignant tumors.