Method for detection and classification of malignant lymph nodes in digestive endoscopic ultrasound
The method optimizes medical object detection models through dynamic data selection, adaptive augmentation, and curriculum learning, addressing data inefficiencies and class imbalance to enhance accuracy and adaptability in endoscopic ultrasound imaging.
Patent Information
- Application Number
- PCT/PT2025/050023
- Authority / Receiving Office
- WO · WO
- Patent Type
- Applications
- Current Assignee / Owner
- Priority Date
- 2025-05-09
- Filing Date
- 2025-09-19
- Publication Date
- 2026-11-12
AI Technical Summary
Existing object detection models in medical imaging, particularly endoscopic ultrasound, face challenges such as limited data availability, computational resource intensity, and suboptimal performance due to redundant data, class imbalance, and inefficient training strategies, necessitating improved data utilization and training efficiency.
A method incorporating Dynamic Data Selection, Adaptive Data Augmentation, Progressive Sampling Strategy, Error-Driven Resampling, and Class-Balanced Training, utilizing curriculum learning and informativeness scoring to prioritize and optimize training data, ensuring robustness and generalization.
Reduces computational costs and accelerates training times while enhancing detection accuracy and adaptability of medical object detection models, addressing class imbalance and improving model performance.
Smart Images

Figure PT2025050023_12112026_PF_FP_ABST
Abstract
Description
[0001] DESCRIPTION
[0002] METHOD FOR DETECTION AND CLASSIFICATION OF MALIGNANT LYMPH NODES IN DIGESTIVE ENDOSCOPIC ULTRASOUND
[0003] Background of the invention
[0004] The present invention relates generally to medical diagnostic methods and, more specifically, to a method of detecting and classifying lymph nodes in endoscopic ultrasound images using deep learning.
[0005] Distinguishing between malignant and benign lymph nodes (LNs) remains a significant challenge in clinical practice. Conventional imaging techniques, such as MRI and CT scan and various predictive factors, such as the LNs location, distribution, enhancement patterns, and size, have been employed to differentiate between malignant and benign conditions. However, imaging alone often proves insufficient for an accurate diagnosis due to the lack of specificity required for precise LN characterization. Accurate differentiation influences clinical decisions in oncological settings, ranging from neoadjuvant therapies instead of direct surgery to palliative care and serves as an essential prognostic factor in determining patient outcomes. Conversely, correctly identifying benign conditions can guide appropriate treatment, reduce unnecessary investigations, and alleviate patient anxiety. Endoscopic ultrasound (EUS) has emerged as a minimally invasive and highly effective technique for evaluating mediastinal and abdominal LNs. It provides real-time imaging and direct visualization, enabling the characterization of LNs with high precision and facilitating histopathological diagnosis.
[0006] The development of accurate and generalizable object detection models in EUS imaging heavily relies on the availability of vastamounts of labeled training data. Object detection inherently faces challenges such as viewpoint variation (where the same object can look drastically different from different angles), deformation (non-rigid objects can exhibit extreme pose variations), occlusion (objects can be partially or fully hidden), varying illumination conditions (affecting the appearance of objects at the pixel level), cluttered or textured backgrounds (making it difficult to distinguish objects), and intra-class variation (objects within the same category can have significant visual differences). Addressing these challenges often requires large and diverse datasets and sophisticated training strategies. The development of efficient training techniques, including those focusing on intelligent data utilization, remains a critical area of research in the field of object detection. This dependency becomes particularly pronounced in specialized and resource-constrained fields such as medical imaging, where the acquisition and annotation of data can be expensive, time-consuming, and require expert knowledge. While the performance of object detection models has seen significant advancements with the rise of deep learning techniques, the sheer volume of data often necessitates substantial computational resources and prolonged training times. The presence of redundant or irrelevant data within these large datasets can lead to suboptimal model performance despite the computational investment. Therefore, the exploration and implementation of efficient data utilization strategies are essential to optimize the training of object detection models, especially in domains like medical image analysis where data efficiency can directly translate to faster development cycles and broader accessibility of AI-powered EUS imaging diagnostic tools.
[0007] Patent document US 10430662 discloses a machine-learning (ML) system to detect objects in a sample of images or videos that are like previous indications of a location of an object of interest.
[0008] Further US patent US 2022 0147748 shows a method that calculates confidence scores for classification of a detected object in animage. The calculated confidence scores are used to perform partial, which do not account for the whole CNN network, backward and forward computations.
[0009] Patent document US 10504027 discloses a learning method to select useful training data using two CNN-based feature map generators that communicate with each other for an auto-screening while learning method.
[0010] Lastly, US 10540572 discloses a method for automatically detecting ROIs with a pretrained ML model and then using those ROIs for ML training. It does not score frames for ML training.
[0011] None of the disclosed patent applications scores frames for ML training. The proposed invention offers significant advantages over the state-of-the-art documents for training object detection models, particularly in the context of medical image analysis. By focusing on high-value data through dynamic data selection, the invention reduces computational costs and accelerates training times. The incorporation of advanced data augmentation strategies, such as reinforcement learning-driven policies, GAN-based synthetic data generation, and PPBA, enhances the model ’s robustness and generalization capabilities, leading to improved detection accuracy even with limited data. The use of focal loss and SMOTE within the class-balanced training framework ensures enhanced accuracy on underrepresented medical conditions, addressing the inherent class imbalance often found in medical datasets. Integrating curriculum learning based on informativeness scores facilitates faster convergence and better overall model performance. Furthermore, the potential incorporation of continual learning techniques allows the model to adapt to new medical data over time without forgetting previously learned patterns, ensuring its continuous improvement and relevance. The integration of active learning principles into the sample selection process can lead to a more efficient annotation workflow, reducing the burden on medical experts. Overall, this invention presents an advanced and scalable method for optimizing data utilization in medicalobject detection training, making AI-based diagnostic systems more efficient, accurate, and adaptable.
[0012] Brief summary of the invention
[0013] This invention hereby described provides a method for detecting and classifying LNs in EUS imaging. More specifically, the method addresses the critical need for improved data utilization in the context of training object detection models for medical applications. It focuses on the core problem of maximizing model performance and generalization while minimizing computational costs and training duration. The present invention introduces a comprehensive method to tackle this challenge through Dynamic Data Selection, Adaptive Data Augmentation, Progressive Sampling Strategy, Error-Driven Resampling, and Class-Balanced Training. Curriculum learning draws inspiration from how humans learn, suggesting that training machine learning models by gradually introducing examples from easy to hard can lead to improved generalization and faster convergence. The core idea is to structure the training process in a meaningful sequence, starting with simpler concepts and progressively increasing the complexity as the model learns and improves. Implementing curriculum learning requires defining a measure of difficulty for the training samples and designing a schedule that dictates how the difficulty of the examples presented to the model evolves over time. One approach involves using predefined difficulty metrics based on inherent properties of the data, such as object size, occlusion levels, or image complexity. Self-paced learning offers an alternative where the model itself plays a role in determining the order of learning, often by prioritizing examples it finds easier to learn initially. A balanced curriculum considers not only the difficulty but also the diversity of the samples being introduced, ensuring that the model is exposed to a wide range of variations even within the "easy" examples.What constitutes an "easy" or "hard" example can be highly context-dependent, influenced by factors like the specific object class, the presence of clutter, the degree of occlusion, and curriculum schedule. The effectiveness of curriculum learning can also be highly dependent on the specific object detection task and the characteristics of the dataset. Simply progressing from easy to hard examples might not be sufficient; ensuring diversity within each difficulty level is also crucial for the model ’ s ability to generalize. Employing effective data augmentation techniques, which will be discussed in more detail later, also plays a vital role in improving both the efficiency and robustness of training by artificially increasing the size and diversity of the training data.
[0014] Brief description of the drawings
[0015] FIG. 1 is a flow diagram illustrating the core iterative loop of the proposed object detection training method. The process begins with accessing image data from a data store 1000, creating subsets of this data 1001, and collecting specific data for object detection 1002. The selected data is then scored 2000. Using this scored data, the model training step 3000 is performed. Subsequently, the model ’ s performance is evaluated. If performance is deemed unacceptable, a resampling step 4001 is triggered for retraining. This cycle repeats until satisfactory performance is achieved, concluding the process.
[0016] FIG. 2 is a flow diagram representing a high-level block diagram of the computing system architecture suitable for implementing various embodiments of the present invention. The system is broadly divided into two main blocks: a Error-Driven Sampler Block 4000 and a Model Training Block 3000.
[0017] FIG. 3 illustrates the process within the Error-Driven Sampler block 4000 for determining informativeness scores and selecting ultrasound image samples. The process starts with acquiring asequence of ultrasound frames. For each frame, potential lesion Regions of Interest (ROIs) are identified, and relevant object detection features (such as pixel intensity distributions, texture patterns, edge information) are extracted from these regions. These features serve as input to the Informativeness Score Estimator (ISE, 4100), which employs a scoring algorithm based on predefined heuristics to calculate a score for each ROI, reflecting its potential contribution to training. The resulting set of scores associated with ROIs across the frames then feeds into the Sample Selector 4200. This module applies selection heuristics to choose samples based on the calculated informativeness scores. Optionally, Data Augmentation 4300 may subsequently be performed on the selected samples, potentially applying different transformations depending on their score. The output of the Error-Driven Sampler 4000 is the set of selected (and possibly augmented) samples.
[0018] FIG. 4a is flow diagrams representing the architecture and operation of the Informativeness Score Estimator (ISE) 4100 for ultrasound images. The ISE module 4100 takes as input the extracted features from identified regions of interest in the ultrasound images. Within the ISE 4100, these features are analyzed by a series of sub-modules, each responsible for calculating a specific heuristic score.
[0019] FIG. 4b illustrates a preferred embodiment of the Informativeness Score Estimator (ISE) 4100 through a detailed flow diagram. In this embodiment, the extracted features from ultrasound image regions of interest are processed through specific heuristic calculations. For instance, Entropy 4110 might be calculated based on the distribution of pixel intensities within the region. Diversity 4120 could be assessed by comparing the feature vectors of the region with those of other regions using clustering algorithms. Object Complexity 4130 might be determined by the size of the region and its degree of overlap with other detected structures. Variability 4140 could be quantified by analyzing thegradient magnitude within the region using filters like Sobel. The diagram shows how these individual scores are then combined using tunable weights, allowing for customization of the informativeness score based on the specific characteristics of the ultrasound imaging task and the types of lesions being detected.
[0020] FIG. 5 is a flow diagram illustrating the operation of the Sample Selector block 4200. The Sample Selector 4200 receives the informativeness scores calculated by the ISE module 4100 for all regions of interest in the training dataset. Based on these scores, the module ranks the regions, prioritizing those with higher informativeness. The diagram then depicts the application of a selection strategy, which could involve selecting a fixed number of top-ranked samples, selecting samples based on a statistical distribution of scores (e.g., from specific quartiles), or using additional criteria such as class representation or diversity metrics. The output of the Sample Selector 4200 is a subset of the most informative training samples, which are then passed to the Training Module 3000 for model learning.
[0021] FIG. 6a is an illustrative flow diagram representing image transformation for data augmentation 4300 that may be applied. These transformations can include geometric modifications such as rotation, scaling, translation, and flipping, as well as photometric adjustments like changes in brightness, contrast, hue, and saturation.
[0022] FIG. 6b is an illustrative flow diagram representing a data augmentation 4300 process that is specifically conditioned on the Informativeness Score. The process begins with an input image and its associated informativeness score. A decision-making step then determines the type and intensity of augmentation to be applied based on this score. For instance, images or regions with lower informativeness scores (Low Score Augmentation, 4310), potentially representing simpler or redundant examples, might undergo more aggressive augmentations to increase their complexity and variability. Conversely, samples with higher informativenessscores (High Score Augmentation, 4320), might receive lighter augmentations to preserve their valuable characteristics.
[0023] FIG. 7 shows the implementation of a specific embodiment of the disclosed invention to a dataset of ultrasound images containing approximately 21,000 malignant and 19,000 benign lesion instances. A YOLOvll model was trained using this cycle. Data was selected from the Data store 1000. Initially, frames corresponding to the top 20% informativeness scores (ISE) were prioritized for training. Subsequently, lower- scored samples were gradually introduced following a curriculum learning strategy. The train involved training for up to 20 epochs with a batch size of 32. Model performance was evaluated at the end of each epoch using mean Average Precision (mAP) on a validation set, with a target threshold of > 0.70 and using an early stopping criterion based on no improvement for 5 consecutive epochs. The resample was performed for frames where lesion detection confidence was below 0.6 or those containing misclassified lesions. Reintroducing these prioritized samples, as part of the Error-Driven Resampling module, improved the recall by 12% and boosted the overall validation mAP by 0.05 points compared to iterations omitting this targeted resampling. Our iterative approach, guided by informativeness scores and incorporating error-driven resampling as depicted in Figure 1, resulted in a final mAP of 0.75 on the held-out test set. Crucially, compared to a baseline approach training on all data randomly for the same maximum number of epochs, our method achieved convergence 30% faster and reached a 5% higher final mAP.
[0024] FIG. 8a shows an example of the trained model predicting a malignant object in an ultrasound image. The image also displays the expert-verified manually annotated ("ground truth") object location.
[0025] FIG. 8b shows an example of the trained model predicting a benign object in an ultrasound image. The image also displays the expert-verified manually annotated ("ground truth") object location.Detailed description
[0026] The present invention discloses a method to train object detectors of annotated image data 1002 from medical exams. The described herein discloses the implementation of training a ML model, or a combination thereof, to detect objects from a medical examination. As used herein, the term "image data" 1002 refers to all the data and metadata linked to an image — for example, the 3D array of pixels, bounding boxes coordinates, and bounding boxes classes. The term " Training data" refers to the source image data 1002 that has been used for model training. Expressions such as " An embodiment" or "another embodiment" of the present disclosure are not to be interpreted as excluding the existence of additional embodiments that also incorporate the recited features.
[0027] The main embodiment of the present invention discloses a method to train object detectors characterized in that it considers the following steps:
[0028] ISE module 4100 - evaluate and score training image data based on its estimated informativeness.
[0029] Sample selector module 4200
[0030] Training module 3000
[0031] Adaptive Data Augmentation Module 4300
[0032] Error-Driven Resampling Module 4000
[0033] The main embodiment of the present invention considers a module designed to evaluate and score training image data 1002 based on its estimated informativeness. A key component of this module is the Informativeness Score Estimator (ISE) 4100, which assigns a numerical score to each training instance, reflecting its potential contribution to the learning process of the object detection model. This score is then used by the Sample Selection module to prioritize the most valuable data for training.In a preferred embodiment, the Informativeness Score (IS) for a training sample x is calculated as a weighted sum of several heuristic scores: IS(x) = H(x) + D(x) + C(x) + V(x), where each component captures a different aspect of informativeness.
[0034] Entropy 4110 (H (x) ) measures the model ’ s confusion or uncertainty when presented with similar samples, with higher entropy indicating potentially more informative samples that lie near decision boundaries.
[0035] Diversity 4120 (D (x) ) ensures that the selected training data provides a diverse representation of object features, preventing the model from overfitting to a narrow subset of variations. In the same preferred embodiment, this is assessed by measuring the minimum cosine similarity between the feature representation of a sample and its nearest neighbors in the feature space identified through dimensionality reduction Principal Component Analysis (PCA). In another preferred embodiment of the invention, the feature space from clustering algorithms such as K-means or simply the latent space of a feature extractor model.
[0036] Object Complexity 4130 (C (x) ) prioritizes samples containing objects that are inherently more challenging to detect, such as small, occluded, or cluttered objects. In the preferred embodiment, this score is inversely proportional to the object size and the Intersection over Union (IoU) with other overlapping objects.
[0037] Variability 4140 (V (x) ) measures the contrast, sharpness, and illumination changes within an image, with higher variability potentially indicating a more informative sample that exposes the model to a wider range of visual conditions. In the preferred embodiment, it is assessed by calculating the variance of gradients using a Sobel filter. The weights associated with each of these heuristic scores allow for customization of the IS calculation based on the specific medical imaging task and the characteristics of the dataset. The flexibility to assign tunable weights to the initial heuristic scores (Entropy 4110, Diversity 4120, Complexity4130, Variability 4140) allows for the adaptation of the Informativeness Score to the specific requirements of different medical imaging tasks. For instance, in detecting subtle anomalies, the Variability score might be weighted more heavily, while for identifying rare conditions, the Complexity score might be prioritized.
[0038] Alternatively, the invention considers embodiments where the IS calculation methods is adapted to further enhance the robustness and adaptability of the informativeness assessment. For example, in yet another preferred embodiment, the initial heuristic-based score is fused with a model-based variability 4140 estimate. In the same preferred embodiment, a gradient-based influence score using a weighted average or more sophisticated fusion techniques is considered.
[0039] Yet in another preferred embodiment, the IS is adapted to enhance the performance of the M model implemented by tailoring the heuristics and their weights to the specific medical condition or object detection task. For instance, when detecting tumors, heuristics focusing on the sharpness of lesion boundaries or the contrast between the tumor and surrounding tissue are emphasized. For identifying specific types of cells, texture-based variability measures could be given higher weight. The invention further considers incorporating Metadata associated with the medical images to provide valuable contextual information for informativeness estimation. Parameters such as image resolution, acquisition modality (e.g., amplitude mode, brightness mode, motion mode, doppler mode), contrast agents used, and even anonymized and relevant patient demographics factored into the IS calculation. For example, higher resolution images or images acquired with specific protocols might be considered inherently more informative.
[0040] In yet another embodiment of the invention the IS is calculated based on pre-training heuristics or a pre-trained model. The object detection model is then trained for a number of epochs on a subsetof the data selected based on this initial IS. After this initial training phase, the IS is recalculated for the entire training dataset based on the model ’ s performance, predictions, and potentially the variability associated with those predictions. Such iterative refinement allows the informativeness assessment to become more accurate as the model learns. Optionally, the informativeness of a medical image for object detection can be assessed through model-based uncertainty estimation (MC Dropout for example). Alternatively, gradient similarity is another gradient-based approach that measures how aligned the gradient of a particular training sample is with the gradient of the overall training objective, often evaluated on a validation set. Samples whose gradients are highly similar to the target objective ’s gradient are likely to contribute more effectively to the learning process. Information theory metrics beyond entropy 4110 can also be employed. Mutual information measures the statistical dependence between two random variables, such as a data sample and its class label or the model ’ s prediction. Samples exhibiting high mutual information with the desired output are likely to be highly informative for the model to learn the correct associations. Jensen-Shannon divergence (JSD) provides a measure of similarity between two probability distributions. In the context of informativeness, JSD can be used to compare the probability distribution of the model ’ s predictions for a given sample with the average prediction distribution across the dataset or with the true label distribution. Samples that lead to a JSD indicating a significant difference might be more informative for learning diverse patterns or correcting the model ’ s biases.
[0041] The Sample Selection module receives the informativeness scores 2000 assigned by the ISE module 4100 and selects a subset of training samples 1001 to be used in the current training iteration. On the preferred embodiment, such selection is based on simply choosing the top-N ranked images.
[0042] Alternatively, the selection is a predetermined number of samples based on the statistical distribution of the IS values (e.g., fromspecific quartiles or the top-K IS images). The technical advantage of the sample selection module is the filtering out redundant or low-value images to optimize training efficiency while ensuring a diverse and balanced dataset by preventing the overrepresentation of common object classes.
[0043] In another preferred embodiment, the module further incorporates additional information available from the image data 1002, such as image resolution or provenance, into the ranking and selection process. A key technical advantage of this module is its ability to filter out redundant or low-value images, thereby optimizing training efficiency while ensuring a diverse and balanced dataset and preventing the overrepresentation of common object classes. Alternatively, sample selection uses Diversity -Aware Selection to maximize the diversity 4120 of the selected subset in the feature space. Clustering techniques, such as k-means clustering applied to the feature embeddings of the images, are be used to identify diverse and representative samples within the top-ranked pool. Subsequently, Stratified Sampling by Class ensures that a proportional number of high-IS samples are selected from each object class present in the dataset. This is particularly important in medical imaging where certain conditions or abnormalities might be underrepresented. By ensuring a balanced representation of classes in the selected subset, the model is less likely to be biased towards more frequent or inherently "informative" classes. In the same embodiment, the method considers a combination of IS Thresholding with Minimum Representation. In detail, all samples with an IS above a certain predefined threshold are initially considered for selection. However, to guarantee sufficient representation for all object classes, a minimum number of samples are selected for each class, even if some of those samples fall below the IS threshold. This ensures that the model receives adequate exposure to all categories of medical objects.
[0044] Finally, if an active learning strategy is being employed, the Sample Selection module is tightly integrated with the activelearning query strategy. The module selects unlabeled samples for annotation based on a combination of their informativeness score (as determined by the ISE) 4100 and their uncertainty according to the active learning algorithm. This allows for the annotation of the most valuable and uncertain samples, further enhancing the efficiency of data utilization.
[0045] The preferred embodiment of the invention considers a training module 3000, where Curriculum Learning Integration involves training the model using a curriculum guided by the Informativeness Scores 2000. The training preferably starts with the samples having the highest IS, which are presumably the "easiest" and most informative, and gradually introduce samples with lower IS as the training progresses. This easy-to-hard approach promotes a faster convergence of the training method and leads to better overall generalization. The Training module 3000 receives the preprocessed training data, which has been scored for informativeness, and the Sample Selection module provides the subset of data to be used for the current training iteration. Training begins with a small, diverse subset of high-informativeness samples, allowing for rapid initial convergence of the model. During training, the module monitors real-time model performance through loss tracking, validation metrics, and per-image detection prediction data. Samples that are consistently misclassified or for which the model produces low-confidence predictions are flagged for potential resampling in subsequent iterations. Optionally, data augmentation strategies 4300 are dynamically applied based on the complexity 4130 of the object and the model ’ s confidence scores. For instance, more aggressive augmentations (e.g., occlusions, distortions) can be used for harder-to-detect objects 4320, while lighter augmentations (e.g., brightness, rotation) can be applied to well-represented objects to prevent overfitting 4310. New data is introduced gradually based on their informativeness scores 2000 to improve learning stability, and harder examples are prioritized as training progresses to refine the model ’ s robustness. This module monitors the distribution of object classes within the trainingdataset and dynamically adjusts the sampling rates for different classes. Underrepresented classes are oversampled, meaning they are selected more frequently during training to ensure balanced learning across all categories. Conversely, overrepresented classes might be downsampled or their contribution to the loss function normalized to prevent bias towards the majority classes. This dynamic balancing ensures that the final model performs well across all object types, mitigating potential performance discrepancies caused by class imbalance. The training module 3000 receives preprocessed IS-estimated training data. A sample is selected after sending training data to the sample selection module. Training begins with a small, diverse subset of high-informativeness samples. Training is performed on this subset of samples. Monitoring real-time model performance through loss tracking, validation metrics, and per-image detection prediction data. Misclassified or low-confidence samples are resampled and reintroduced to reinforce learning. In yet another optional embodiment, the augmentation strategies are dynamically applied based on object complexity 4130 and confidence scores 2000 - for example, more aggressive augmentations (e.g., occlusions, distortions) are used for harder-to-detect objects; and lighter augmentations (e.g., brightness, rotation) are applied to well-represented objects to prevent overfitting.
[0046] New data is introduced gradually based on ISs to improve learning stability. Harder examples are prioritized as training progresses to refine model robustness. A subset of training data passes through the sample selection module using IS values. In an embodiment, the subset of training data comprises all training data. In another embodiment, the subset of training data is predefined by a subset size. In a preferred embodiment, the subset of training data are the top-K IS-ranked images.
[0047] The most generic embodiment of the invention considers the Adaptive Data Augmentation Module 4300. Such module selectively applies data transformations based on the complexity 4130 of the object class and the difficulty the model has in detecting it. ThisConfidence-Based Augmentation applies more aggressive augmentations to samples for which the model has low detection confidence, while lighter augmentations are used for high-confidence samples. This focuses the augmentation efforts on areas where the model is struggling to learn. In such embodiment, difficult-to-detect objects undergo more aggressive transformations such as rotation, scaling, and brightness variation to expose the model to a wider range of appearances and improve its resilience. An automated feedback loop monitors the model ’ s performance on different object classes and dynamically adjusts the augmentation rates per class. Augmentations that do not contribute to performance improvement, or even hinder it, are gradually phased out to optimize efficiency.
[0048] Alternatively, Reinforcement Learning for Optimal Policies can be employed to automate the process of determining the best augmentation strategies. An agent can be trained using reinforcement learning, where the state is defined by the current object class and detection difficulty, the actions are different augmentation operations and their parameters, and the reward is based on the model ’ s performance on a validation set. The agent learns an optimal policy that dictates which augmentations to apply in different scenarios to maximize performance. Medical Image Specific Augmentation involves adapting augmentation parameters based on the unique characteristics of the medical images being used. Further embodiments include Confidence-Based Augmentation, where more aggressive augmentations are applied to samples for which the model has low detection confidence, while lighter augmentations are used for high-confidence samples. This focuses the augmentation efforts on areas where the model is struggling to learn.
[0049] Still in the most generic embodiment of the present method, the Error-Driven Resampling module 4000 focuses on improving the model ’ s performance on challenging examples by reintroducing misclassified or uncertain predictions into the training dataset. Misclassified samples, where the model ’ s prediction does not matchthe ground truth, and samples with low confidence scores 2000 are flagged and given higher priority in subsequent training iterations. These samples are reintroduced with additional variations or modifications to reinforce learning and enhance the model ’ s robustness. A confidence threshold determines which samples should be resampled, preventing excessive reinforcement of patterns the model has already learned well.
[0050] Efficient data utilization is paramount for the effective training of object detection models, especially within the resource-sensitive domain of medical image analysis. The presented invention addresses this critical need through a comprehensive and integrated approach encompassing dynamic data selection, adaptive data augmentation 4300, progressive sampling, error-driven resampling 4000, and class-balanced training 3000. By intelligently prioritizing informative data, tailoring augmentations to the specific challenges of medical object detection and addressing issues like class imbalance and the need for continuous learning, this method offers a significant advancement in the field. The various additional embodiments and alternative strategies discussed further highlight the flexibility and potential of this invention to be adapted and optimized for a wide range of medical imaging applications. Ultimately, the method of the present invention promises to reduce computational costs, improve training efficiency, and enhance the performance of medical object detection models for more accessible and reliable AI-powered diagnostic tools.
[0051] The foregoing Detailed Description is to be understood as being in every respect illustrative and exemplary, but not restrictive, and the scope of the invention disclosed herein is not to be determined from the Detailed Description, but rather from the claims as interpreted according to the full breadth permitted by the patent laws. It is to be understood that the embodiments shown and described herein are only illustrative of the principles of the present invention and that various modifications may beimplemented by those skilled in the art within the scope of the
[0052] appended claims.
Claims
AMENDED CLAIMSreceived by the International Bureau on 27 March 2026 (27.03.2026)1. A computer-implemented method for training a model for lymph nodes detection in ultrasound imaging, comprising:- assigning an informativeness score to each training sample in the dataset of lymph node ultrasound images based on one or a combination of heuristics from entropy, diversity, object complexity, and variability calculated from said images;- prioritizing a subset of said training samples for training based on informativeness scores;- dynamically applying data augmentation to the prioritized training samples based on the complexity of lymph node classes depicted in the images and the difficulty in detection derived from the model' s scores;- progressively expanding the training dataset by introducing new lymph node ultrasound image samples based on their informativeness scores;- resampling uncertain predictions or misclassified lymph node image samples from the training dataset with increased augmentation adapted to improve learning;- dynamically adjusting the sampling rate or loss function weights for underrepresented lymph node classes to mitigate class imbalance during training.wherein the data augmentation strategy is one or a combination of using reinforcement learning and generative adversarial networks.
2. The method according to claim 1, wherein prioritizing the subset of samples comprises selecting samples having an informativeness score that exceeds a predetermined threshold.
3. The method according to claim 1 or 2, wherein the informativeness score calculation further considers the trained model ' s weights.
4. The method for training a model according to the previous claims, wherein the informativeness score is further adaptedincorporate specific features characteristic of lymph nodes extracted from the ultrasound images.
5. The method for training a model according to the previous claims, wherein the informativeness score is calculated based on the heuristics derived from a pre-trained model.
6. The method for training a model according to the previous claims, wherein the informativeness score is iteratively recalculated for the entire training dataset or a portion thereof based on the trained model' s performance throughout the training process.
7. The method for training a model according to the previous claims, wherein the informativeness score is one or a combination of techniques selected from the group consisting of model-based uncertainty estimation, gradient-based methods, information theory metrics, or Shapley values.
8. The method for training a model according to previous claims 1 or 2, wherein the sample selection is based on a predetermined number of samples based on the statistical distribution of the Informativeness Score values.
9. The method for training a model according to previous claims 1, 2 or 8 wherein the sample selection further incorporates additional information available from the image data, such as image resolution or provenance.
10. The method for training a model according to previous claims 1 or 2, 8 or 9 wherein the sample selection uses clustering technique in the feature embeddings of the images to identify diverse or representative samples within the top-ranked pool.
11. The method for training a model according to previous claims 1, 2 or 8 to 10 wherein the sample selection further considers stratified sampling by class adapted to ensure a balanced representation of classes in the selected subset.
12. The method for training a model according to previous claims 1, 2 or 8 to 11 wherein the sample selection is integrated with an active learning strategy adapted to detecting the most valuable and uncertain strategies.
13. The method for training a model according to previous claim 1, wherein the progressive expansion of the training dataset is adapted to prioritizing high Informativeness Score value samples and gradually introduce samples with lower Informativeness Score as the training progress.
14. The method for training a model according to previous claims 1, wherein progressively expansion of the training dataset is based on a learning curve analysis.
15. The method for training a model according to previous claim 1, wherein the resampling of misclassified instances is weighted based on the severity of the misclassification or involves the generation of synthetic hard examples.
16. The method for training a model according to previous claim 1, wherein class imbalance is addressed using focal loss or synthetic oversampling techniques.[0001]STATEMENT UNDER ARTICLE 19 (1)[0002]Indication of amendments and their basis[0003]A new set of claims 1-16 is submitted in replacement of the previously filed set of claims.[0004]Amended claim 1 contains subject-matter of claims 1 and 14 as originally filed.[0005]Previous claim 8 was excised as it was a repetition of claim 2.[0006]Previous claims 12 and 13 dependencies were corrected to include dependency on claim 2.[0007]The term “cost-sensitive learning” was excised from previous claim 18 to further clarify the scope of the claim.[0008]Lastly, the numbering of remaining dependent claims was corrected.[0009]It is respectfully submitted that the amended set of claims is in conformity with A123(2), EPC.
