An explainable active learning method driven by cell image class attributes in collaboration

Abstract

The application belongs to the technical field of explainable active learning, and particularly relates to an explainable active learning method driven by cell image class attribute cooperation. In the inspection process, an image is acquired from a blood cell smear by using a camera, a newly collected cell image dataset D1 is constructed, a small amount of samples are randomly extracted from the dataset D1 to construct an initial annotation set D2, the D2 is subjected to class annotation and model pre-training is completed based on the D2, the remaining unannotated data in the D1 is subjected to standardization and pretreatment, and then is divided into a plurality of batches according to the quantity to form a candidate unannotated pool D3, the data in the D3 is sequentially input into the model according to a preset batch, for each batch of input data, a class imbalance adaptive sample screening mechanism CASS is adopted, three indexes of uncertainty, representativeness and diversity are comprehensively utilized under the current weight configuration, attribute information is introduced, and class frequency and effective sample number are combined to perform weighted calculation, and candidate high-value samples of the batch are screened.

Description

Technical Field

[0001] This invention belongs to the field of interpretable active learning technology, and particularly relates to an interpretable active learning method driven by the collaborative interaction of cell image class attributes. Background Technology

[0002] While active learning technology has been used to reduce annotation costs in the intelligent application of cell morphology examination, it still faces core technological bottlenecks. Traditional methods often focus on a single screening indicator, failing to fully integrate cell morphology attributes and inadequately addressing common clinical class imbalances. This leads to the easy overlooking of rare pathological cell samples, resulting in weak model recognition capabilities for minority classes and difficulty meeting the accuracy requirements of clinical testing. Furthermore, existing sample screening strategies lack dynamic adaptability, failing to adjust and optimize the direction at different stages of model training, further limiting learning efficiency and model generalization performance.

[0003] Furthermore, most mainstream active learning methods operate in a black-box manner, lacking interpretability to support the sample selection logic. Clinical laboratory experts cannot know the specific criteria by which the model identifies high-value samples, making it difficult to effectively verify and trust the results. This problem severely restricts the practical application of the technology in medical settings. Simultaneously, the lack of mechanisms to maintain the semantic stability of attributes during model training makes them prone to attribute drift, leading to insufficient coherence between sample selection and model updates in subsequent learning processes, thus affecting the long-term optimization performance of the system. Summary of the Invention

[0004] The purpose of this invention is to address the aforementioned technical problems by providing an interpretable active learning method driven by the collaborative interaction of cell image class attributes.

[0005] In view of this, the present invention provides an interpretable active learning method driven by the collaborative interaction of cell image class attributes, comprising the following steps: Step 1: During the examination, images are acquired from blood cell smears using a camera to construct a newly acquired cell image dataset D1; Step 2: Randomly select a small number of samples from dataset D1 to construct an initial labeled set D2, label D2 with categories and complete model pre-training based on it; After standardization and preprocessing, the remaining unlabeled data in D1 are divided into several batches according to quantity to form a candidate unlabeled pool D3. Step 3: Input the data in D3 into the model in the preset batches. For each batch of input data, the Class Imbalanced Adaptive Sample Screening Mechanism (CASS) is used. Under the current weight configuration, the three indicators of uncertainty, representativeness and diversity are comprehensively used, and attribute information is introduced. The weighted calculation is performed by combining the class frequency and the number of effective samples to screen out the candidate high-value samples for that batch. Step 4: The AdaptiveMixer is used to schedule the weights of four indicators—uncertainty, representativeness, diversity, and attribute information—to achieve a gradual switching between the Early, Middle, and Late stages. The Early stage emphasizes representativeness and diversity to broaden sample coverage; the Middle stage maintains a relatively balanced weighting of the four indicators; and the Late stage increases the weights of uncertainty and attribute information. The weights output by the AdaptiveMixer are used to guide the sample selection strategy for the corresponding stage. Step 5: After each batch of samples is selected, the model is fine-tuned using the selected new samples in small steps, and the model is updated by combining exponential moving average (EMA) and lightweight regularization; during the evaluation phase, the classification accuracy (acc), balance accuracy (bal_acc), macro-F1 score (macro-F1), and expected calibration error (ECE) are monitored to ensure the model’s stable performance in class balance and calibration. Step Six: During the active learning process, the attribute information of each batch of selected sample sets is interpreted post-hocly using the AIAL method. Combined with cell morphology knowledge, the reasons why the samples are judged to be of high value are explained, and the interpretation results are visualized and output. Step 7: Use the system optimized and interpreted in the above steps to conduct a new round of blood cell morphology examination. Repeat steps 1-6 to achieve continuous evolution and dynamic optimization of the system.

[0006] Preferably, before model pre-training in step two, the backbone network and attribute network layers need to be constructed. Specifically, the backbone network includes a low-dimensional feature extractor L and a high-dimensional feature extractor H. The low-dimensional feature extractor L is composed of the first few residual blocks of the pre-trained model ResNeXt50, which is used to extract and output basic morphological features. The high-dimensional feature extractor H consists of several residual blocks of the pre-trained model ResNeXt50, and is used to learn and output higher-level discriminative features of cells. An attribute network layer ANL is introduced before the category classifier C to project the high-dimensional features output by the high-dimensional feature extractor H onto the attribute space to obtain an attribute vector. The cell morphology attribute dimensions corresponding to this attribute vector include cell size, nuclear-cytoplasmic ratio, and particle density. The Attribute Network layer (ANL) consists of two linear mapping layers. It is responsible for compressing and aligning the high-dimensional representation in the high-dimensional features to the attribute semantic space, so that the system can obtain interpretable attribute representations in addition to the classification output, forming a dual-view representation structure of classification prediction and attribute interpretation.

[0007] Preferably, the policyr and mixer need to be initialized before model pre-training, specifically including: The modules required to instantiate the four-way indicators include the uncertainty strategist U, the representativeness strategist R, the diversity strategist D, and the attribute value channel A; The uncertainty policymaker U uses entropy to measure the degree of ambiguity of a sample at the decision boundary; The representative strategy R adopts KNN-density, which measures the degree to which the sample covers the overall distribution by calculating the Top-k cosine similarity mean and normalizing it to [0,1]. The diversity strategist D employs the K-Center strategy to encourage the dispersed selection of feature space; The attribute value channel A is composed of a weighted combination of attribute uncertainty, attribute novelty, and attribute coverage gap. An adaptive mixer is constructed. During the initialization phase, the mixer is provided with three-stage weights according to a preset definition. The adaptive switching mechanism automatically decides the weight configuration based on the batch performance consistency index ST. After active learning and iterative sampling, the mixer loads the corresponding strategy configuration according to the current stage and dynamically switches between Early-Middle-Late.

[0008] Preferably, the adaptive switching mechanism in step four is implemented as follows: the performance consistency index ST is used to comprehensively measure the performance improvement, consistency and calibration status of the current batch relative to the previous batch. Its calculation basis includes the overall prediction accuracy, average recall of the category, macro average F1 score and the change in expected calibration error of the batch on the validation set. The phase switching rule is as follows: if the performance consistency index ST exceeds the switching threshold of the current phase for h consecutive times, then proceed to the next phase; If the performance consistency index ST drops significantly below the rollback threshold, or if a core performance index continues to degrade, then the system will roll back from the higher stage to the previous stage. The adaptive mixer outputs the weights of four indicators in each batch, which are directly used as input to the Class Imbalanced Adaptive Sample Selection Mechanism (CASS) for the corresponding stage, guiding the comprehensive evaluation, ranking, and selection of candidate samples.

[0009] Preferably, the specific implementation process of the Class Imbalanced Adaptive Sample Selection Mechanism (CASS) in step three includes: Category statistics are weighted by the number of valid samples; The class frequencies of labeled samples are statistically analyzed. A class sensitivity score is constructed based on these frequencies. This score is then normalized and truncated to obtain a final class weighting. This weighting is used to prioritize the screening of minority class samples. Class sensitivity relabeling: For each candidate sample, the expected rarity is calculated using the predicted distribution and class weighting, and this rarity is used as a sample rarity adjustment factor in the final sample selection; class-sensitive deduplication and class quota are jointly constrained. For candidate samples of the same category, a deduplication strategy is executed based on a similarity threshold set according to category sensitivity weighting. This deduplication strategy is not executed between different categories. If the number of samples after deduplication is insufficient to meet the selection budget B, samples from the remaining candidate pool are supplemented according to the comprehensive score, with priority given to filling back samples of minority classes. Set a lower and upper limit quota for the number of samples for each category. The quota is linearly allocated based on category sensitivity weighting and the sampling budget B for this batch. Comprehensive scoring: The four indicators of uncertainty, representativeness, diversity and attribute information are scaled in the same way. The comprehensive value score of the sample is calculated by combining the category-sensitive weighting and the sample-level expected rarity. High-value samples are selected while ensuring coverage of a minority class.

[0010] Preferably, in step five, the model training adopts an incremental loss function system driven by class attributes. This loss function system includes three parts: class-sensitive sparseness modulation classification loss, attribute consistency co-distillation loss, and lightweight regularization loss, specifically: Class-sensitive rarity modulation classification loss: By combining category-sensitive weighting with the expected rarity of samples to construct a comprehensive weight, the gradient update direction of the model is kept consistent with the sampling logic, thereby enhancing the robustness of the model to rare categories. Attribute Consistency Collaborative Distillation Loss: Utilizing the Attribute Network Layer (ANL) and attribute uncertainty, selective distillation is performed on the predictions of the old and new models at the attribute level. By constraining the differences in attribute probability distributions and the direction of attribute representations, attribute semantic drift is prevented, thus maintaining the interpretability and stability of the model in the attribute space. Lightweight regularization loss: Suppresses sharp parameter oscillations by calculating the L2 norm of the current model parameters and the snapshot of the model parameters at the end of the previous window.

[0011] Preferably, the specific implementation of AIAL post-hoc interpretation in step six includes: calculating three types of attribute interpretability indices: attribute uncertainty, which is used to measure the degree of uncertainty of the model's prediction of samples in the attribute prediction layer, and is achieved by taking the entropy average of the binary distribution of each attribute; Attribute novelty measures how much a sample deviates from its predicted class prototype in the attribute space, and is calculated using Mahalanobis distance. Attribute coverage gap is used to measure the ability of a sample to cover and supplement the existing labeled set in terms of attribute dimension. It is achieved by calculating the normalized distance of the sample attribute in the positive uncovered area. Based on the morphological attribute vectors extracted by the attribute network layer ANL, and combined with cell morphology knowledge, the selected samples are given a post-hoc interpretation at the attribute level, which explains the basis for the model's selection of samples. By combining attribute metrics and model attention, a cell-level feature heatmap is generated, which intuitively displays the key areas that the model focuses on, forming readable and verifiable interpretation results. The interpretation information is used as a feedback signal to assist model diagnosis and data selection for the next round of active learning.

[0012] Preferably, in step four, the dynamic mixer dynamically adjusts the weights of the four types of indicators according to the model training requirements at each stage. In the Early stage, the weight ratio of representativeness and diversity indicators is increased; in the Middle stage, the weight ratio of the four types of indicators tends to be balanced; and in the Late stage, the weight ratio of uncertainty and attribute information indicators is increased.

[0013] Preferably, in step five, the model fine-tuning adopts the method of joint training of high-value samples selected in the current window and historical memory samples, and only performs small-step fine-tuning. At the same time, the classification accuracy acc, balanced accuracy bal_acc, macro-average F1 value macro-F1 and expected calibration error ECE index are monitored. The evaluation signal is generated based on the changes of the index and injected back into the adaptive mixer to drive the weight update and stage switching of the next window.

[0014] Preferably, in step two, the standardization and preprocessing of the unlabeled data in D1 includes image noise reduction, size unification, and pixel value normalization operations. The batch division is performed according to a preset fixed quantity or a dynamically adjusted quantity standard.

[0015] The beneficial effects of this invention are: This method effectively addresses the shortcomings of traditional active learning in cell image analysis, namely, its inadequacy in handling class imbalance and lack of dynamic adaptability in screening strategies, through the synergistic effect of the Class Imbalanced Adaptive Sample Selection Mechanism (CASS) and the dynamic mixer AdaptiveMixer. CASS, by combining class frequency and effective sample number weighting, significantly improves the screening priority of rare pathological cell samples. Combined with a three-stage weight scheduling strategy—expanding sample coverage in the Early stage, balancing various indicators in the Middle stage, and focusing on boundary cases and key attributes in the Late stage—and an incremental loss function system driven by class attributes, this method not only ensures the model's ability to identify minority classes but also suppresses parameter oscillations and attribute semantic drift through attribute consistency distillation and lightweight regularization. This leads to continuous optimization of core indicators such as classification accuracy and balance accuracy, and a significant improvement in model generalization and robustness.

[0016] Meanwhile, this method constructs a dual-perspective representation structure of classification prediction and attribute interpretation through an attribute network layer (ANL). Combined with the AIAL post-interpretation method, it deeply integrates the criteria for determining the high value of samples with cell morphology knowledge and outputs the results in visual forms such as heatmaps. This completely breaks down the black-box barrier of traditional active learning, allowing clinical laboratory experts to intuitively verify the sample selection logic and significantly improving the clinical credibility of the technology. Through a continuously iterative active learning process, the system achieves a closed loop of data screening, model optimization, and interpretation feedback. It can dynamically adapt to the actual needs of blood cell morphology testing, reducing the cost of manual annotation while steadily improving the accuracy of testing, providing core support for the clinical implementation of intelligent cell morphology testing. Attached Figure Description

[0017] Figure 1 This is a diagram of the method structure; Figure 2 Construct a diagram of the backbone network and attribute network layers; Figure 3 This is a schematic diagram of the adaptive mixer stage switching process; Figure 4 A schematic diagram of the sample screening and deduplication process; Figure 5 A flowchart for attribute explanation and visualization output. Detailed Implementation

[0018] The technical solutions of the embodiments of this application will be clearly described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application are within the scope of protection of this application.

[0019] It should be noted that all directional and positional terms used in this invention, such as "up," "down," "left," "right," "front," "back," "vertical," "horizontal," "inner," "outer," "top," "lower," "lateral," "longitudinal," and "center," are only used to explain the relative positional relationships and connections between components in a specific state (as shown in the accompanying drawings). They are merely for the convenience of describing the invention and do not require the invention to be constructed and operated in a specific orientation; therefore, they should not be construed as limitations on the invention. Furthermore, descriptions involving "first," "second," etc., are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated.

[0020] In the description of this invention, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal communication between two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0021] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "illustrative embodiment," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0022] S1, During the testing process, images are acquired from blood cell smears using a camera to construct a newly acquired cell image dataset D1; S2, randomly select a small number of samples from the newly acquired cell image dataset D1 to construct an initial annotation set D2, label D2 with categories, and pre-train the model based on D2; after standardization and preprocessing, the remaining unlabeled data in D1 are divided into several batches according to quantity to form a candidate unlabeled pool D3. S3: The system uses D3 as the data stream source for subsequent processes, inputting it into the model sequentially according to pre-defined batches. For each batch of input data, a Class Imbalanced Adaptive Sample Selection Mechanism (CASS) is employed. Under the current weight configuration, it comprehensively utilizes three types of indicators—uncertainty, representativeness, and diversity—and incorporates attribute information. It combines category frequency and the number of valid samples for weighted calculation to obtain candidate high-value samples for that batch. S4. To differentiate the focus of CASS across different learning stages, the system uses an Adaptive Mixer to schedule the weights of the four categories of indicators, enabling a gradual switching between the Early, Middle, and Late stages. The Early stage emphasizes representativeness and diversity to broaden sample coverage; the Middle stage maintains a relatively balanced weight across the four categories; and the Late stage increases the weights of uncertainty and attributes to accommodate the needs of later model refinement. The weights output by the Adaptive Mixer will guide CASS's sample selection strategy at each stage. S5. After each batch of samples is selected, the model is micro-fine-tuned using the selected new samples, and the model is updated by combining exponential moving average (EMA) and lightweight regularization. During the evaluation phase, the classification accuracy (acc), balance accuracy (bal_acc), macro-F1 score (macro-F1), and expected calibration error (ECE) are monitored to ensure the model’s stable performance in class balance and calibration. S6. During the active learning process, the attribute information of each batch of selected sample set w is interpreted post-hocly using the AIAL method. Combined with cell morphology knowledge, the reasons why the sample is judged to be of high value are explained, and the interpretation results are visualized and output to enhance the trust of clinical laboratory experts in the system. S7. Using the system that has been continuously optimized and interpreted through the above steps, a new round of blood cell morphology examination is conducted. Steps S1–S6 are repeated to achieve continuous evolution and dynamic optimization of the system.

[0023] The specific method is as follows: CAS-AL method; Initialization Phase: A newly collected cell dataset D1 is constructed based on existing cell category data. A small number of samples randomly selected from D1 form an initial labeled set D2, and D2 is labeled with categories for model pre-training. The remaining data is input into the system for standardization and preprocessing, and divided into batches to obtain a candidate unlabeled pool D3. During this phase, the system completes the model structure construction and core module initialization, providing a foundation for subsequent active learning.

[0024] Construction of backbone network and attribute network layers; The network includes a low-dimensional feature extractor L, which consists of the first few residual blocks of the pre-trained model (ResNeXt50), used to extract and output basic morphological features. It also includes a high-dimensional feature extractor H, composed of several residual blocks from the pre-trained model, used to learn and output higher-level discriminative features of cells. Before the classifier C, an attribute network layer (ANL) is introduced to encode the high-dimensional features. Projecting onto the attribute space yields the attribute vector: ; These are dimensions of cell morphology attributes (such as cell size, nucleoplasmic ratio, granule density, etc.).

[0025] ANL consists of two layers of linear mappings. responsible for The high-dimensional representation in the model is compressed and aligned to the attribute semantic space. By introducing ANL, the system obtains interpretable attribute representations in addition to the classification output, thus realizing a dual-view representation structure of classification prediction and attribute interpretation.

[0026] (2) Initialization of the strategy unit and mixer; The system instantiates the modules required for the four metrics and builds an adaptive mixer, including: Uncertainty Strategist (U): Employs Entropy to measure the ambiguity of a sample's location on the decision boundary; Representative strategy (R): KNN-density (calculate the top-k cosine similarity mean and normalize to [0,1]) is used to measure the coverage of the overall distribution by the sample; Diversity Strategist (D): Employs the K-Center strategy to encourage the dispersed selection of feature space; Attribute Value Channel (A): Attribute Uncertainty Attribute novelty With attribute coverage gap It is composed of three weighted components (see “Phase 1: Indicator Calculation” for details).

[0027] During the initialization phase, the three-stage weights are provided to the mixer according to a predefined set; the adaptive switching mechanism (see Phase 3) then determines the weights based on batch-to-batch performance consistency metrics. Automatic decision-making. Once active learning begins iterative sampling, the mixer loads the corresponding strategy configuration based on the current stage and dynamically switches between Early, Middle, and Late.

[0028] Phase 1: Samples are taken from the candidate unlabeled pool D3 and input into the backbone network and attribute network layer. Morphological features are extracted through the low-dimensional feature extractor L and the high-dimensional feature extractor H. Four types of indicators, namely uncertainty, representativeness, diversity and attribute value, are calculated to achieve multi-dimensional evaluation of each sample, providing data support for subsequent sample value judgment and class imbalance adaptive screening.

[0029] Indicator Calculation: Uncertainty : Used to measure the model's prediction uncertainty for sample x, employing information entropy: ; in K represents the predicted probability of the model for category c, and K represents the total number of categories.

[0030] Representative The mean cosine similarity of neighborhood density is used to measure the extent to which a sample covers the overall distribution; the larger the value, the more common the region to which the sample belongs.

[0031] ; in The feature vector output by the feature extraction network. These are the k nearest neighbor samples to x.

[0032] diversity The K-Center strategy encourages the selection of samples that are far from the current center set. ; in This is the current set of selected center samples. The feature vector output by the feature extraction network; Attribute Value Used to measure the information value provided by a sample x in the morphological attribute space, due to the uncertainty of its attributes. Attribute novelty With attribute coverage gap It is composed of three weighted combinations: ; in These are the weighting coefficients. , , See Stage Six for the definition.

[0033] Phase Two: Based on the calculation of multiple indicators, a weighted mechanism of class frequency and effective sample number is introduced to perform class-sensitive relabeling and quota constraints on candidate samples. By combining uncertainty, representativeness, diversity and attribute value, a dynamic balance is achieved between prioritizing rare cells and sample distribution, ensuring that the model pays full attention to and utilizes minority class samples during the active learning process.

[0034] Category statistics are weighted by the number of valid samples; During the initialization phase, the class frequency of the labeled samples (including old class samples replayed from memory) is counted. ,in K represents the total number of categories. A category sensitivity score is constructed based on category frequency: ; in, For smoothing parameters.

[0035] Subsequently, the category sensitivity scores are normalized and truncated with upper and lower limits to obtain the final category weighted score: ; in, , representing the mean of sensitivity scores for all categories, where γ>0 indicates the reinforcement index for rare classes. This is the cutoff range for the weighting coefficients. Final weights. This will be used to increase the priority of minority groups in subsequent sample screening stages.

[0036] Category-sensitive recalibration; For each candidate sample x, use the predicted distribution Weighted by category Calculate the expected rarity: ; in This represents the total number of categories.

[0037] Category weighting of Phase 2 output Compared with sample-level expected rarity This will serve as the category constraint and sample rarity adjustment factor for the comprehensive scoring in stage four, and will participate in the final sample selection together with the four weights output from stage three.

[0038] Phase 3: The model integrates four indicators—uncertainty, representativeness, diversity, and attribute value—through an adaptive mixer and uses a phase switching threshold ST to achieve a smooth transition between the Early, Middle, and Late phases. This dynamically adjusts the policy weights, allowing the model to balance exploration and utilization at different phases.

[0039] Three-stage switching strategy: Early stage; Primarily designed for early-stage training when the model structure is not yet stable, the goal is to quickly establish a complete understanding of the data distribution. Enhancing representativeness (R) and diversity (D) ensures comprehensive sampling coverage and sufficient inter-class representation. Uncertainty (U) is treated as a secondary factor to avoid premature concentration on noisy or difficult examples. Attribute value (A) is used to select samples with typical morphological features, improving the quality of initial features.

[0040] Middle stage; Once the model has established a preliminary inter-class structure, it enters the Middle stage, which aims to maintain training stability and class balance. The weights of the four metrics tend to be balanced, taking into account both exploration (D, R) and utilization (U, A). Emphasis is placed on class balance to avoid weakening overall generalization ability due to oversampling of certain classes. The Middle stage is the "steady-state region" of the active learning system and has the most significant impact on training quality.

[0041] Late stage; For cases where the model has converged but still requires improved boundary discrimination capabilities, the following measures are taken: Increase the uncertainty U to pay more attention to difficult boundary examples and confused samples. Add attribute value A to highlight key attribute differences and enhance the identification of minority classes and fine-grained morphologies.

[0042] Adaptive switching mechanism: To provide a consistent measure of overall model performance variation across adjacent batches, a stage switching threshold (SwitchThreshold, ST) based on performance consistency is introduced to comprehensively evaluate batch performance. Compared to the previous batch The performance improvement, consistency, and calibration status. The switching threshold is defined as: ; in ; The four indicators are defined as follows; :batch Overall prediction accuracy on the validation set; :batch Category average recall is used to measure class balance. :batch The macro-average F1 score reflects the ability to distinguish between categories. :batch The expected calibration error reflects the degree of matching between model confidence and prediction accuracy.

[0043] The performance variation between adjacent batches is defined as: ; the remaining , , Defined in the same way.

[0044] The phase switching rules are as follows: like The handover threshold for the current stage (e.g., the handover threshold from Early to Middle) is exceeded h consecutive times. Middle→Late switching threshold If so, proceed to the next stage; like Significantly reduced to the fallback threshold If a core performance indicator continues to degrade, it is permissible to revert from a higher stage to a previous stage to avoid frequent jitter caused by noise.

[0045] In summary, the Adaptive Mixer outputs four weights in each batch. This will be directly used as input to the next stage of the CASS-type imbalanced adaptive sample screening mechanism to guide the comprehensive evaluation, ranking and selection of candidate samples, thereby forming a continuous closed loop of "weight scheduling - sample screening".

[0046] Phase Four: Based on the four weights output from Phase Three, the CASS class imbalanced adaptive screening mechanism is used to perform weighted comprehensive scoring on all candidate samples. After determining the weights of each path, a candidate sample pool is constructed and deduplicated, backfilled, and comprehensively scored. By normalizing and weighted summing the sample indicators, the final comprehensive value score is calculated. Under the premise of ensuring coverage of the minority class, the most informative high-value samples are selected to provide high-quality data input for incremental training of the model.

[0047] Joint constraints of class-sensitive deduplication and category quotas; Class-sensitive similarity threshold ; For candidate samples of the same class c, according to their class sensitivity weights Set similarity threshold ; This deduplication strategy is not applied between different categories to avoid accidentally deleting cross-category samples.

[0048] Backfilling: If the number of samples after deduplication is less than budget B, supplement them from the remaining samples in the candidate pool according to the comprehensive score, and prioritize backfilling of minority classes.

[0049] Category quota constraint range; To prevent majority class samples from consuming the entire budget while ensuring that minority class samples can be included in training, the system sets a minimum quota for the number of selectable samples for each class c. and upper limit quota .

[0050] Quotas are weighted by category. With the sampling budget for this batch Linear allocation: ; in Control the upper and lower limits of intensity. .

[0051] Overall score; For each metric, perform homoscaling (min-max) and combine it with the category-sensitive weighting of the stage two output. Compared with sample-level expected rarity The overall score is: ; ; in Four weights from the output of the adaptive mixer; in The aforementioned expected rarity is obtained by normalizing its value to the [0,1] interval. , used for weighting. , is a hyperparameter for controlling the intensity of the effects of class sensitivity and sample rarity.

[0052] Phase 5: The selected high-value samples and historical memory samples are combined for micro-fine-tune training. By combining EMA smoothing, lightweight regularization and attribute-aware loss function system, the model can be stably updated and knowledge is preserved. At the same time, during the validation phase, metrics such as acc, bal_acc, macro-F1 and ECE are monitored to evaluate model performance and calibration stability.

[0053] Model training and loss function design; This invention constructs a class-attribute collaboratively driven incremental loss function system. Within each window, this system only utilizes "high-value samples selected in the current window" and "historical memory samples" for small-step fine-tuning. It modulates the loss term through class-sensitive weights, expected sample rarity, and attribute interpretation weights, thereby highlighting the contributions of rare classes and key attributes while maintaining model stability.

[0054] The loss function during training consists of three parts: class-sensitive sparse modulation classification loss, attribute consistency co-distillation loss, and lightweight regularization loss. ; Class-Sensitive Rarity Modulation Classification Loss : To ensure that the training process is consistent with the class sensitivity evaluation and sample expected rarity in Phase 2, this invention constructs a class sensitivity rarity modulated classification loss.

[0055] Let the true label of the i-th sample be... The model predicts the probability of this category as: The category sensitivity weights obtained in Phase 2 are: The expected rarity of the sample is The overall weight is defined as follows: ; in is the total number of categories, and c represents the category index. Controlling the intensity of rarefaction modulation The classification loss is then: ; in As a focusing factor, This is the current number of fine-tuned samples.

[0056] This loss ensures that the gradient update direction of the model is completely consistent with the sampling logic of the preceding stage, thereby maintaining robustness to rare classes during continuous active learning.

[0057] Attribute Consistency Collaborative Distillation Loss ; To maintain the semantic stability of the attribute space and enhance the consistency between the old and new models in the attribute space, this invention utilizes the Attribute Network Layer (ANL) constructed in stages one through four and the attribute uncertainty defined in stage six. Selective distillation is performed on the attribute-level predictions of the old and new models.

[0058] Assume the old model in the sample The output attribute logit vector is The new model output is After smoothing with temperature parameter T, the attribute probability vector is obtained: ; Considering the differences in stability among different samples in the attribute space, this invention introduces adaptive weighting coefficients based on attribute uncertainty. This is used to impose stronger consistency constraints on samples with unstable attributes. Its definition is as follows: ; in, The uncertainty of the attributes of sample x. , These represent the mean and standard deviation of the attribute uncertainty within the current batch, respectively. To prevent division by zero of extremely small constants, This is the Sigmoid mapping function.

[0059] Therefore, the attribute consistency collaborative distillation loss is defined as: ; in, , The first term in the equation constrains the confidence of the new model at the attribute level through the difference in attribute probability distribution, while the second term constrains the direction of attribute representation through cosine similarity. This together prevents attribute semantic drift and maintains the interpretability and stability of the model in the attribute space.

[0060] ③ Lightweight regularization loss

[0061] To further suppress drastic parameter oscillations, this invention introduces a lightweight regularization term in addition to the two losses mentioned above: ; in, These are the current model parameters. This is a snapshot of the model parameters at the end of the previous window.

[0062] By minimizing the total loss L during the fine-tuning process of each window, the system achieves the combined effect of class balance constraints, attribute consistency, and parameter smoothing between consecutive windows. This enables the model to quickly absorb new knowledge while maintaining its memory of rare pathological cells and key morphological attributes when faced with a continuously growing stream of blood cell data, thus providing a more stable foundation for subsequent CASS sampling and AIAL interpretation.

[0063] (2) Evaluation and signal reinjection; In each window After fine-tuning, the system calculates four performance metrics for the window on the validation set: { , , , These respectively reflect overall accuracy, category balance, classification discrimination ability, and model calibration degree; Then, these indicators are compared with the previous window. By comparing the corresponding metrics, a performance increment vector is obtained. ; And combined with the change in calibration error Together they constitute the evaluation signal.

[0064] The evaluation signal is fed back to AdaptiveMixer, driving the next window according to the stage switching strategy. The weight updates and phase switching form a learning loop.

[0065] Phase Six: After active learning is completed, the system performs attribute-driven interpretability analysis (AIAL) on the selected sample set in each round. By calculating interpretability indicators at the attribute level and generating a visual heatmap, the system reveals the decision-making basis of the model and feeds the interpretation results back to the system to support the continuous optimization of subsequent active learning.

[0066] (1) Calculation of attribute interpretability index Three types of interpretability indicators were calculated for each selected sample to reveal its selection value: Attribute uncertainty (notation) ); Used to measure the uncertainty of a model's predictions of samples at the attribute prediction head.

[0067] Assume the model outputs the attribute prediction probability vector as follows: Where A is the total number of attributes, and , .

[0068] The uncertainty of an attribute is defined as entropy: ; By taking the entropy average of the binary distributions of each attribute, the model's ambiguity regarding the samples at the attribute level is comprehensively measured. The higher the uncertainty, the greater the potential for information gain from the samples.

[0069] Attribute novelty (denoted as...) ); It is used to measure how much a sample deviates from its predicted class prototype in the attribute space.

[0070] The attribute vector of the sample is The attribute prototype vector of its category is The covariance matrix is .

[0071] Attribute novelty is defined as the Mahalanobis distance form: .; It reflects the "outlierness" of a sample in terms of attribute distribution. When A larger value indicates that the sample differs significantly from the existing training samples in terms of morphological features, and has potential novel information.

[0072] ③ Attribute Coverage Gap (denoted as...) ) It is used to measure the ability of a sample to cover and supplement the existing labeled set in terms of attribute dimensions.

[0073] Let the mean and standard deviation of the current annotation set on the a-th attribute dimension be respectively and The corresponding attribute value of the sample is .

[0074] The coverage gap is defined as the normalized distance of a sample attribute in the "positive uncovered area": ; in This is an adjustable hyperparameter (usually set to 1.0). To prevent division by zero of small constants.

[0075] If a sample exceeds the upper bound of the labeled distribution in some attribute dimensions, it is considered an extension of the knowledge coverage. The larger the value, the more significantly the sample can expand the cognitive space of the model at the attribute level.

[0076] (2) Attribute explanation and visualization output; After active learning is completed, the system uses the morphological attribute vectors extracted by the attribute network layer ANL to provide post-hoc interpretation of the selected samples at the attribute level, helping users understand the basis for the model's selection of these samples (such as karyotype, karyocytic ratio, staining degree, granularity, etc.).

[0077] Simultaneously, by combining attribute metrics and model attention, a cell-level feature heatmap is generated, which intuitively displays the key areas of focus of the model, forming readable and verifiable interpretation results.

[0078] The aforementioned explanatory information will serve as feedback signals for the system, assisting in model diagnosis and data selection for the next round of active learning, thereby enhancing the interpretability and continuously optimizing the system.

[0079] The embodiments of this application have been described above with reference to the accompanying drawings. Unless otherwise specified, the embodiments and features in the embodiments of this application can be combined with each other. This application is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of this application without departing from the spirit and scope of the claims, and all of these forms are within the protection scope of this application.

Claims

1. An interpretable active learning method driven by the collaborative interaction of cell image class attributes, characterized in that: Includes the following steps: Step 1: During the examination, images are acquired from blood cell smears using a camera to construct a newly acquired cell image dataset D1; Step 2: Randomly select a small number of samples from dataset D1 to construct an initial labeled set D2, label D2 with categories and complete model pre-training based on it; After standardization and preprocessing, the remaining unlabeled data in D1 are divided into several batches according to quantity to form a candidate unlabeled pool D3. Step 3: Input the data in D3 into the model in the preset batches. For each batch of input data, the Class Imbalanced Adaptive Sample Screening Mechanism (CASS) is used. Under the current weight configuration, the three indicators of uncertainty, representativeness and diversity are comprehensively used, and attribute information is introduced. The weighted calculation is performed by combining the class frequency and the number of effective samples to screen out the candidate high-value samples for that batch. Step 4: The AdaptiveMixer is used to schedule the weights of four indicators—uncertainty, representativeness, diversity, and attribute information—to achieve a gradual switching between the Early, Middle, and Late stages. The Early stage emphasizes representativeness and diversity to broaden sample coverage; the Middle stage maintains a relatively balanced weighting of the four indicators; and the Late stage increases the weights of uncertainty and attribute information. The weights output by the AdaptiveMixer are used to guide the sample selection strategy for the corresponding stage. Step 5: After each batch of samples is selected, the model is fine-tuned using the selected new samples in small steps, and the model is updated by combining exponential moving average (EMA) and lightweight regularization; during the evaluation phase, the classification accuracy (acc), balance accuracy (bal_acc), macro-F1 score (macro-F1), and expected calibration error (ECE) are monitored to ensure the model’s stable performance in class balance and calibration. Step Six: During the active learning process, the attribute information of each batch of selected sample sets is interpreted post-hocly using the AIAL method. Combined with cell morphology knowledge, the reasons why the samples are judged to be of high value are explained, and the interpretation results are visualized and output. Step 7: Use the system optimized and interpreted in the above steps to enter a new round of blood cell morphology examination, and repeat steps 1-6 to achieve continuous evolution and dynamic optimization of the system; The implementation method of the adaptive switching mechanism in step four is as follows: The performance consistency index ST is used to comprehensively measure the performance improvement, consistency and calibration status of the current batch relative to the previous batch. Its calculation is based on the batch’s overall prediction accuracy, class average recall, macro average F1 score and the change in expected calibration error on the validation set. The phase switching rule is as follows: if the performance consistency index ST exceeds the switching threshold of the current phase for h consecutive times, then proceed to the next phase; If the performance consistency index ST drops significantly below the rollback threshold, or if a core performance index continues to degrade, then the system will roll back from the higher stage to the previous stage. The adaptive mixer outputs the weights of four indicators in each batch, which are directly used as input to the corresponding stage-level class imbalance adaptive sample screening mechanism (CASS) to guide the comprehensive evaluation, ranking and selection of candidate samples. The specific implementation process of the Class Imbalanced Adaptive Sample Selection Mechanism (CASS) in step three includes: Category statistics are weighted by the number of valid samples; The class frequencies of labeled samples are statistically analyzed. A class sensitivity score is constructed based on these frequencies. This score is then normalized and truncated to obtain a final class weighting. This weighting is used to prioritize the screening of minority class samples. Class sensitivity relabeling: For each candidate sample, the expected rarity is calculated using the predicted distribution and class weighting, and this rarity is used as a sample rarity adjustment factor in the final sample selection; class-sensitive deduplication and class quota are jointly constrained. For candidate samples of the same category, a deduplication strategy is executed based on a similarity threshold set according to category sensitivity weighting. This deduplication strategy is not executed between different categories. If the number of deduplicated samples is insufficient to meet the selection budget B, samples from the remaining candidate pool are supplemented according to the comprehensive score, with priority given to filling samples from minority classes. Set a lower and upper limit quota for the number of samples for each category. The quota is linearly allocated based on category sensitivity weighting and the sampling budget B for this batch. Comprehensive scoring: The four indicators of uncertainty, representativeness, diversity and attribute information are scaled together, and the comprehensive value score of the sample is calculated by combining category-sensitive weighting and sample-level expected rarity, so as to select high-value samples while ensuring coverage of minority classes. The specific implementation of AIAL post-hoc interpretation in step six includes: calculating three types of attribute interpretability indices: attribute uncertainty, which measures the degree of uncertainty in the model's prediction of samples at the attribute prediction layer, and is achieved by taking the entropy average of the binary distribution of each attribute; Attribute novelty measures how much a sample deviates from its predicted class prototype in the attribute space, and is calculated using Mahalanobis distance. Attribute coverage gap is used to measure the ability of a sample to cover and supplement the existing labeled set in terms of attribute dimension. It is achieved by calculating the normalized distance of the sample attribute in the positive uncovered area. Based on the morphological attribute vectors extracted by the attribute network layer ANL, and combined with cell morphology knowledge, the selected samples are given a post-hoc interpretation at the attribute level, which explains the basis for the model's selection of samples. By combining attribute metrics and model attention, a cell-level feature heatmap is generated, which intuitively displays the key areas that the model focuses on, forming readable and verifiable interpretation results. The interpretation information is used as a feedback signal to assist model diagnosis and data selection for the next round of active learning.

2. The interpretable active learning method driven by the collaborative interaction of cell image class attributes according to claim 1, characterized in that: In step two, before model pre-training, the backbone network and attribute network layers need to be constructed. Specifically, the backbone network includes a low-dimensional feature extractor L and a high-dimensional feature extractor H. The low-dimensional feature extractor L is composed of the first few residual blocks of the pre-trained model ResNeXt50, which is used to extract and output basic morphological features. The high-dimensional feature extractor H consists of several residual blocks of the pre-trained model ResNeXt50, and is used to learn and output higher-level discriminative features of cells. An attribute network layer ANL is introduced before the category classifier C to project the high-dimensional features output by the high-dimensional feature extractor H onto the attribute space to obtain an attribute vector. The cell morphology attribute dimensions corresponding to this attribute vector include cell size, nuclear-cytoplasmic ratio, and particle density. The Attribute Network layer (ANL) consists of two linear mapping layers. It is responsible for compressing and aligning the high-dimensional representation in the high-dimensional features to the attribute semantic space, so that the system can obtain interpretable attribute representations in addition to the classification output, forming a dual-view representation structure of classification prediction and attribute interpretation.

3. The interpretable active learning method driven by the collaborative interaction of cell image class attributes according to claim 2, characterized in that: Before model pre-training, the policyr and mixer need to be initialized, which includes: The modules required to instantiate the four-way indicator include the uncertainty strategist U, the representativeness strategist R, the diversity strategist D, and the attribute value channel A; The uncertainty policymaker U uses entropy to measure the degree of ambiguity of a sample at the decision boundary; The representative strategy R adopts KNN-density, which measures the degree to which the sample covers the overall distribution by calculating the Top-k cosine similarity mean and normalizing it to [0,1]. The diversity strategist D employs the K-Center strategy to encourage the dispersed selection of feature space; The attribute value channel A is composed of a weighted combination of attribute uncertainty, attribute novelty, and attribute coverage gap. An adaptive mixer is constructed. During the initialization phase, the mixer is provided with three-stage weights according to a preset definition. The adaptive switching mechanism automatically decides the weight configuration based on the batch performance consistency index ST. After active learning and iterative sampling, the mixer loads the corresponding strategy configuration according to the current stage and dynamically switches between Early-Middle-Late.

4. The interpretable active learning method driven by the collaborative interaction of cell image class attributes according to claim 3, characterized in that: In step five, the model training employs a class-attribute collaboratively driven incremental loss function system. This loss function system comprises three parts: class-sensitive sparseness modulation classification loss, attribute consistency collaborative distillation loss, and lightweight regularization loss, specifically: Class-sensitive rarity modulation classification loss: By combining category-sensitive weighting with the expected rarity of samples to construct a comprehensive weight, the gradient update direction of the model is kept consistent with the sampling logic, thereby enhancing the robustness of the model to rare categories. Attribute Consistency Collaborative Distillation Loss: Utilizing the Attribute Network Layer (ANL) and attribute uncertainty, selective distillation is performed on the predictions of the old and new models at the attribute level. By constraining the differences in attribute probability distributions and the direction of attribute representations, attribute semantic drift is prevented, thus maintaining the interpretability and stability of the model in the attribute space. Lightweight regularization loss: Suppresses sharp parameter oscillations by calculating the L2 norm of the current model parameters and the snapshot of the model parameters at the end of the previous window.

5. The interpretable active learning method driven by collaborative cell image class attributes according to claim 1, characterized in that: In step four, the dynamic mixer schedules the weights of the four types of indicators, dynamically adjusting them according to the model training requirements at each stage. In the Early stage, the weight ratio of representativeness and diversity indicators is increased; in the Middle stage, the weight ratio of the four types of indicators tends to be balanced; and in the Late stage, the weight ratio of uncertainty and attribute information indicators is increased.

6. The interpretable active learning method driven by the collaborative interaction of cell image class attributes according to claim 1, characterized in that: In step five, model fine-tuning adopts a joint training method using high-value samples selected in the current window and historical memory samples. Only small-step fine-tuning is performed, while monitoring classification accuracy (acc), balanced accuracy (bal_acc), macro-F1 score (macro-F1), and expected calibration error (ECE). Based on the changes in the indicators, an evaluation signal is generated and injected back into the adaptive mixer to drive the weight update and stage switching of the next window.

7. The interpretable active learning method driven by the collaborative interaction of cell image class attributes according to claim 1, characterized in that: Step two involves the standardization and preprocessing of the unlabeled data in D1, including image noise reduction, size unification, and pixel value normalization. Batch division is performed according to a preset fixed quantity or a dynamically adjusted quantity standard.

Images