Nodule category prediction model training and ultrasound image nodule identification method and device

Abstract

The application relates to the technical field of ultrasonic image recognition, and discloses a nodule category prediction model training method and an ultrasonic image nodule recognition method and device. The method comprises the following steps: firstly, obtaining a training ultrasonic data set to perform preliminary training, obtaining a global threshold value, a local threshold value and a conservative threshold value, and then obtaining an initial nodule category prediction model. An ultrasonic image data set is obtained; the model is trained in a first stage by using the data to obtain a first nodule category prediction model; through the preliminary training in the first stage, the first nodule category prediction model can obtain preliminary classification capability; the first nodule category prediction model is further trained to obtain a trained nodule category prediction model; through the preliminary training to obtain the global threshold value, the local threshold value and the conservative threshold value, the confidence level of the initial nodule category prediction model is improved, the number of different nodule categories is increased, and the false positive rate of the nodule is reduced; finally, the nodule category prediction model is trained, so that the accuracy of nodule category prediction is improved.

Description

Technical Field

[0001] This invention relates to the field of ultrasound image processing technology, and in particular to a method and apparatus for training a nodule category prediction model and identifying nodules in ultrasound images. Background Technology

[0002] Current ultrasound examinations require doctors to classify nodules in a standardized manner, and different classifications directly determine the patient's subsequent management plan. For example, nodules are classified as follows: Class 1 is normal, Class 2 is benign, Class 3 is possibly benign (short-term follow-up recommended), and Class 4 is suspected malignant (biopsy recommended).

[0003] However, in actual clinical practice, radiologists may give different classification results for different categories of nodules in the same ultrasound image; this difference is particularly prominent when distinguishing between category 3 and category 4, and the boundary between category 3 and category 4 determines whether the patient should continue to be followed up for observation or undergo biopsy, which directly affects clinical decision-making.

[0004] Existing models have low accuracy in predicting nodules of the four classes, with most falling into three or fewer classes, resulting in class imbalance and a tendency for the model to favor the majority class. Furthermore, obtaining high-quality labeled data is costly, and the labeling itself suffers from inter-observer variability, leading to a scarcity of labeled data.

[0005] Therefore, a new training method for the nodule category prediction model is needed to improve the accuracy of nodule prediction. Summary of the Invention

[0006] This invention provides a method and apparatus for training a nodule category prediction model and identifying nodules in ultrasound images, which can improve the accuracy of nodule category prediction models in predicting nodules.

[0007] The first aspect of this invention discloses a method for training a nodule category prediction model, the method comprising: Obtain the training ultrasound dataset; The training ultrasound dataset is input into a preset training nodule category prediction model to obtain a global threshold, a local threshold, and a conservative threshold. The global threshold is used to reflect the overall confidence level of the model. The local threshold is used to adjust the number of nodule categories identified. The conservative threshold is used to reduce the false positive rate of the target nodule. The global threshold, the local threshold, and the conservative threshold are fed back into the trained nodule category prediction model to obtain the initial nodule category prediction model; Obtain a first ultrasound image dataset and a second ultrasound image dataset; the first ultrasound image dataset consists of labeled ultrasound images, and the second ultrasound image dataset consists of unlabeled ultrasound images. The initial nodule category prediction model is trained using the first ultrasound image dataset to obtain the first nodule category prediction model; The first nodule category prediction model is trained using the second ultrasound image dataset to obtain a trained nodule category prediction model; the nodule category prediction model is used to identify the category of nodules in ultrasound images.

[0008] As an optional implementation, in the first aspect of the present invention, the step of inputting the training ultrasound dataset into a preset training nodule category prediction model to obtain a global threshold includes: The training ultrasound dataset is input into a preset training nodule category prediction model to predict the nodule category, thereby obtaining the nodule category prediction probability value corresponding to each ultrasound image in the training ultrasound dataset. Based on the predicted probability values ​​of all the nodules of the aforementioned categories, the maximum predicted probability value of the nodule of the category corresponding to each ultrasound image is obtained; A global threshold is obtained based on the predicted probability values ​​of all the largest category nodules.

[0009] As an optional implementation, in the first aspect of the present invention, the step of inputting the training ultrasound dataset into a preset training nodule category prediction model to obtain a local threshold includes: Perform the following operations for each type of nodule: When it is determined that the predicted probability value of the nodule category is greater than the preset first target nodule probability threshold, a first target nodule is obtained; based on all the first target nodules, the number of first target nodules is obtained. When it is determined that the number of the first target nodules is less than the preset threshold for the number of the first target nodules, the probability threshold for the first target nodules is adjusted to obtain the adjusted probability threshold. A local threshold is obtained based on the adjusted probability threshold corresponding to each type of nodule.

[0010] As an optional implementation, in the first aspect of the present invention, the step of inputting the training ultrasound dataset into a preset training nodule category prediction model to obtain a conservative threshold includes: Based on the predicted probability values ​​of all the nodules of the aforementioned categories, the predicted probability value of the second target nodule is obtained; When it is determined that the predicted probability value of the second target nodule is less than the preset conservative nodule probability threshold, the second target nodule corresponding to the predicted probability value of the second target nodule is marked as a safe nodule. When it is determined that the predicted probability value of the second target nodule is greater than or equal to the preset conservative nodule probability threshold, the second target nodule corresponding to the predicted probability value of the second target nodule is marked as the first biopsy nodule. The nodule corresponding to the predicted probability value of the largest category nodule is marked as a contrast nodule; The second biopsy nodule is obtained by matching the comparison nodule and the safe nodule. Based on the second biopsy nodule and the first biopsy nodule, the proportion of the second biopsy nodule is obtained; when it is determined that the proportion of the second biopsy nodule is less than a preset proportion threshold, the conservative nodule probability threshold is adjusted to obtain a conservative threshold.

[0011] As an optional implementation, in a first aspect of the present invention, training the initial nodule category prediction model using the first ultrasound image dataset to obtain a first nodule category prediction model includes: The first ultrasound image dataset is augmented to obtain an augmented ultrasound image dataset. The initial nodule category prediction model is trained using the enhanced ultrasound image dataset and a preset first cross-entropy loss function to obtain a first nodule category prediction model; wherein, the first cross-entropy loss function is used to constrain the training of the initial nodule category prediction model during the training process.

[0012] As an optional implementation, in the first aspect of the present invention, training the first nodule category prediction model using the second ultrasound image dataset to obtain a trained nodule category prediction model includes: The second ultrasound image dataset is subjected to a first image enhancement process to obtain a first enhanced image; The second ultrasound image dataset is subjected to a second image enhancement process to obtain a second enhanced image; The first enhanced image is input into the first nodule category prediction model to obtain nodule prediction labels; when it is determined that the data of the nodule prediction label is greater than the global threshold and the local threshold, the nodule prediction label is used to participate in the training of the first nodule category prediction model. Based on the nodule prediction label, the first nodule category prediction model is trained using the second enhanced image and a preset second cross-entropy loss function to obtain a trained nodule category prediction model. The second cross-entropy loss function is used to constrain the training of the first nodule category prediction model during the training process.

[0013] A second aspect of this invention discloses a method for identifying nodules in ultrasound images, the method comprising: Acquire the ultrasound image to be predicted; The ultrasound image to be predicted is input into the nodule category prediction model to obtain the output result of the nodule category prediction model; based on the output result, the category of the nodule in the ultrasound image to be predicted is obtained; wherein, the nodule category prediction model is trained by any of the nodule category prediction model training methods of the first aspect of the present invention.

[0014] A third aspect of the present invention discloses a training apparatus for a nodule category prediction model, the apparatus comprising: A pre-training module is used to acquire a training ultrasound dataset; the training ultrasound dataset is input into a pre-defined training nodule category prediction model to obtain a global threshold, a local threshold, and a conservative threshold; the global threshold is used to reflect the overall confidence level of the model; the local threshold is used to adjust the number of nodule categories identified; the conservative threshold is used to reduce the false positive rate of the target nodule; the global threshold, the local threshold, and the conservative threshold are fed back into the training nodule category prediction model to obtain an initial nodule category prediction model; The training data acquisition module is used to acquire a first ultrasound image dataset and a second ultrasound image dataset; the first ultrasound image dataset consists of labeled ultrasound images, and the second ultrasound image dataset consists of unlabeled ultrasound images. The formal training module is used to train the initial nodule category prediction model using the first ultrasound image dataset to obtain a first nodule category prediction model; and to train the first nodule category prediction model using the second ultrasound image dataset to obtain a trained nodule category prediction model; the nodule category prediction model is used to identify the category of nodules in ultrasound images.

[0015] As an optional implementation, in a third aspect of the present invention, the specific operation method by which the pre-training module inputs the training ultrasound dataset into a preset training nodule category prediction model to obtain a global threshold includes: The training ultrasound dataset is input into a preset training nodule category prediction model to predict the nodule category, thereby obtaining the nodule category prediction probability value corresponding to each ultrasound image in the training ultrasound dataset. Based on the predicted probability values ​​of all the nodules of the aforementioned categories, the maximum predicted probability value of the nodule of the category corresponding to each ultrasound image is obtained; A global threshold is obtained based on the predicted probability values ​​of all the largest category nodules.

[0016] As an optional implementation, in a third aspect of the present invention, the specific operation method by which the pre-training module inputs the training ultrasound dataset into a preset training nodule category prediction model to obtain a local threshold includes: Perform the following operations for each type of nodule: When it is determined that the predicted probability value of the nodule category is greater than the preset first target nodule probability threshold, a first target nodule is obtained; based on all the first target nodules, the number of first target nodules is obtained. When it is determined that the number of the first target nodules is less than the preset threshold for the number of the first target nodules, the probability threshold for the first target nodules is adjusted to obtain the adjusted probability threshold. A local threshold is obtained based on the adjusted probability threshold corresponding to each type of nodule.

[0017] As an optional implementation, in a third aspect of the present invention, the pre-training module inputs the training ultrasound dataset into a preset training nodule category prediction model to obtain a conservative threshold. The specific operation includes: Based on the predicted probability values ​​of all the nodules of the aforementioned categories, the predicted probability value of the second target nodule is obtained; When it is determined that the predicted probability value of the second target nodule is less than the preset conservative nodule probability threshold, the second target nodule corresponding to the predicted probability value of the second target nodule is marked as a safe nodule. When it is determined that the predicted probability value of the second target nodule is greater than or equal to the preset conservative nodule probability threshold, the second target nodule corresponding to the predicted probability value of the second target nodule is marked as the first biopsy nodule. The nodule corresponding to the predicted probability value of the largest category nodule is marked as a contrast nodule; The second biopsy nodule is obtained by matching the comparison nodule and the safe nodule. Based on the second biopsy nodule and the first biopsy nodule, the proportion of the second biopsy nodule is obtained; when it is determined that the proportion of the second biopsy nodule is less than a preset proportion threshold, the conservative nodule probability threshold is adjusted to obtain a conservative threshold.

[0018] As an optional implementation, in a third aspect of the present invention, the formal training module uses the first ultrasound image dataset to train the initial nodule category prediction model to obtain a specific operation method for the first nodule category prediction model, including: The first ultrasound image dataset is augmented to obtain an augmented ultrasound image dataset. The initial nodule category prediction model is trained using the enhanced ultrasound image dataset and a preset first cross-entropy loss function to obtain a first nodule category prediction model; wherein, the first cross-entropy loss function is used to constrain the training of the initial nodule category prediction model during the training process.

[0019] As an optional implementation, in a third aspect of the present invention, the specific operation of the formal training module using the second ultrasound image dataset to train the first nodule category prediction model to obtain the trained nodule category prediction model includes: The second ultrasound image dataset is subjected to a first image enhancement process to obtain a first enhanced image; The second ultrasound image dataset is subjected to a second image enhancement process to obtain a second enhanced image; The first enhanced image is input into the first nodule category prediction model to obtain nodule prediction labels; when it is determined that the data of the nodule prediction label is greater than the global threshold and the local threshold, the nodule prediction label is used to participate in the training of the first nodule category prediction model. Based on the nodule prediction label, the first nodule category prediction model is trained using the second enhanced image and a preset second cross-entropy loss function to obtain a trained nodule category prediction model. The second cross-entropy loss function is used to constrain the training of the first nodule category prediction model during the training process.

[0020] A fourth aspect of the present invention discloses an ultrasound image nodule identification device, the device comprising: Image acquisition module, used to acquire the ultrasound image to be predicted; The nodule identification module is used to input the ultrasound image to be predicted into the nodule category prediction model to obtain the output result of the nodule category prediction model; and to obtain the category of the nodule in the ultrasound image to be predicted based on the output result; wherein the nodule category prediction model is trained by any of the nodule category prediction model training methods of the first aspect of the present invention.

[0021] A fifth aspect of the present invention discloses an apparatus comprising a memory and a processor, the apparatus comprising: Memory containing executable program code; A processor coupled to the memory; The processor calls the executable program code stored in the memory to execute the nodule category prediction model training method disclosed in the first aspect of the present invention or the ultrasound image nodule recognition method disclosed in the second aspect of the present invention.

[0022] Compared with the prior art, the present invention has the following beneficial effects: This invention first acquires a training ultrasound dataset and trains a pre-defined training nodule category prediction model to obtain global thresholds, local thresholds, and conservative thresholds. These thresholds are then fed back into the training nodule category prediction model to obtain an initial nodule category prediction model. Next, an ultrasound image dataset is acquired. The pre-defined initial nodule category prediction model is first trained using a labeled first ultrasound image dataset to obtain a first nodule category prediction model. This initial training allows the first nodule category prediction model to acquire preliminary classification capabilities. Then, the first nodule category prediction model is trained using an unlabeled second ultrasound image dataset to obtain a trained nodule category prediction model. By first obtaining global, local, and conservative thresholds through training, the confidence level of the initial nodule category prediction model is improved, the number of different nodule categories is increased, and the false positive rate of nodules is reduced. Finally, the nodule category prediction model is trained, thereby improving the accuracy of nodule category prediction. Attached Figure Description

[0023] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0024] Figure 1 This is a flowchart illustrating a training method for a nodule category prediction model disclosed in an embodiment of the present invention; Figure 2 This is a schematic flowchart of an ultrasound image nodule identification method disclosed in an embodiment of the present invention; Figure 3 This is a schematic diagram of the structure of a nodule category prediction model training device disclosed in an embodiment of the present invention; Figure 4 This is a schematic diagram of the structure of an ultrasound image nodule recognition device disclosed in an embodiment of the present invention; Figure 5 This is a schematic diagram of the structure of a device including a memory and a processor disclosed in an embodiment of the present invention; Figure 6 This is a schematic diagram of a two-stage semi-supervised training disclosed in an embodiment of the present invention; Figure 7 This is a schematic diagram illustrating the differences in classification and boundary case variations between a conservative and an open model as disclosed in an embodiment of the present invention. Figure 8 This is a schematic diagram of an ultrasound image of a nodule thermogram disclosed in an embodiment of the present invention. Detailed Implementation

[0025] To enable those skilled in the art to better understand the present invention, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0026] The terms "first," "second," etc., used in the specification, claims, and accompanying drawings of this invention are used to distinguish different objects, not to describe a specific order. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, apparatus, product, or system that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to these processes, methods, products, or systems.

[0027] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of the invention. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a mutually exclusive, independent, or alternative embodiment. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0028] This invention provides a method and apparatus for training a nodule category prediction model and identifying nodules in ultrasound images, which improves the accuracy of nodule category prediction models in predicting nodules. These will be described in detail below.

[0029] Example 1 Please see Figure 1 , Figure 1 This is a flowchart illustrating a training method for a nodule category prediction model disclosed in an embodiment of the present invention. Figure 1 The described nodule category prediction model training method can be applied to a nodule category prediction model training device, which can be integrated into a cloud server or a local server; this embodiment of the invention is not limited thereto. Figure 1 As shown, the training method for this nodule category prediction model may include the following operations: Step 101: Obtain the training ultrasound dataset; input the training ultrasound dataset into the preset training nodule category prediction model to obtain the global threshold, local threshold, and conservative threshold; feed the global threshold, local threshold, and conservative threshold back into the training nodule category prediction model to obtain the initial nodule category prediction model.

[0030] In this embodiment of the invention, the training ultrasound dataset can be ultrasound images with labeled nodule categories, such as breast nodules, thyroid nodules, etc.; the training ultrasound dataset can also be images without labeled nodule categories, or images with nodule categories used for comparison; the training ultrasound dataset can also be ultrasound images with both labeled nodule categories and ultrasound images without labeled nodule categories.

[0031] In this embodiment of the invention, the basic architecture for training the nodule category prediction model can be as follows: A DnCNN (Denoising Convolutional Neural Network) is used as the front-end module to first extract and denoise the image, enhancing diagnostic-related features. The output of the DnCNN is fused with the original input through residual connections, thus preserving the original image information while introducing enhanced features after denoising. The fused features are then fed into a ResNet (Residual Network) for deep learning, and finally, the predicted probabilities of the four BI-RADS (nodule) categories are output through softmax (normalized activation function).

[0032] In this embodiment of the invention, the global threshold can be used to reflect the overall confidence level of the model; the local threshold can be used to adjust the number of nodule categories identified; and the conservative threshold can be used to reduce the false positive rate of the target nodule. In this embodiment of the invention, the global threshold can be calculated by using an adaptive threshold filtering mechanism to filter out low-quality pseudo-labels, and by averaging the exponent of the maximum probability predicted by nodules within the same batch to reflect the overall confidence level of the current model.

[0033] In this embodiment of the invention, the local threshold can be set separately for each nodule category and adjusted based on the difference between the predicted distribution and the empirical distribution of that category. For categories with fewer model predictions (such as nodule category 4), the threshold will be adaptively reduced to retain more pseudo-labels of that category, thereby balancing the number of nodules in each category and alleviating the problem of category imbalance.

[0034] In this embodiment of the invention, the conservative threshold can be set as follows: first, a 90% probability threshold is set for nodule categories 4. Boundary cases that do not reach the threshold are automatically classified as category 3. This mode is defined as the first mode. Then, other modules directly output the category with the highest probability. This mode is defined as the second mode. The two modes operate in parallel. If the second mode directly outputs category 4 as the category with the highest probability, and this category 4 nodule has a probability of less than 90% in the first mode and is classified as category 3, then this category 3 nodule is automatically converted to category 4 nodule. At this time, the number of nodule categories output by the first mode and the number of nodule categories output by the second mode are compared to adjust the probability in the first mode, making the nodule prediction more accurate.

[0035] In this embodiment of the invention, global thresholds, local thresholds, and conservative thresholds are pre-trained; these thresholds are then fed back into the training nodule category prediction model in a timely manner, thereby obtaining the initial nodule category prediction model used in this application.

[0036] Furthermore, in this embodiment of the invention, the global threshold, local threshold, and conservative threshold are pre-set in the initial nodule category prediction model. At the beginning of model building, the relevant code architecture is set in the initial nodule category prediction model, and the model can be directly trained subsequently, generating relevant thresholds while training.

[0037] Step 102: Obtain the first ultrasound image dataset and the second ultrasound image dataset.

[0038] In this embodiment of the invention, the first ultrasound image dataset may be ultrasound images with data annotation, and the second ultrasound image dataset may be ultrasound images without data annotation.

[0039] Optionally, data annotation can be performed by manually annotating the nodules in the ultrasound images.

[0040] Step 103: Train the initial nodule category prediction model using the first ultrasound image dataset to obtain the first nodule category prediction model; train the first nodule category prediction model using the second ultrasound image dataset to obtain the trained nodule category prediction model.

[0041] In this embodiment of the invention, the nodule category prediction model can be used to identify the category of nodules in ultrasound images.

[0042] In this embodiment of the invention, the first nodule category prediction model may be trained using a first ultrasound image dataset to enable the model to acquire preliminary nodule classification capabilities.

[0043] As can be seen, in this embodiment of the invention, a training ultrasound dataset is first acquired, and a preset training nodule category prediction model is trained to obtain a global threshold, a local threshold, and a conservative threshold. These thresholds are then fed back into the training nodule category prediction model to obtain an initial nodule category prediction model. Next, an ultrasound image dataset is acquired. The preset initial nodule category prediction model is first trained using a labeled first ultrasound image dataset to obtain a first nodule category prediction model. Through this initial training, the first nodule category prediction model can acquire preliminary classification capabilities. Then, the first nodule category prediction model is trained using an unlabeled second ultrasound image dataset to obtain a trained nodule category prediction model. By first obtaining the global threshold, local threshold, and conservative threshold through training, the confidence level of the initial nodule category prediction model is improved, the number of different nodule categories is increased, and the false positive rate of nodules is reduced. Finally, the nodule category prediction model is trained to improve the accuracy of nodule prediction.

[0044] In an optional embodiment, the training ultrasound dataset is input into a preset training nodule category prediction model to obtain a global threshold, including: The training ultrasound dataset is input into a preset training nodule category prediction model to predict the nodule category, and the prediction probability value of the nodule category corresponding to each ultrasound image in the training ultrasound dataset is obtained. Based on the predicted probability values ​​of all nodule categories, the predicted probability value of the largest nodule category corresponding to each ultrasound image is obtained; The global threshold is obtained by predicting the probability values ​​of all the largest category nodules.

[0045] In this optional embodiment, the predicted probability value for each type of nodule can be different. For example, the probability of a type I nodule is 10%, the probability of a type II nodule is 15%, the probability of a type III nodule is 60%, and the probability of a type IV nodule is 15%.

[0046] In this optional embodiment, the maximum category nodule prediction probability value can be selected from the probability values ​​of different categories corresponding to each nodule.

[0047] In this optional embodiment, the global threshold can be calculated by filtering out all the maximum probability values ​​in a batch of ultrasound image datasets and then calculating the average value.

[0048] As can be seen, in this optional embodiment, by calculating the nodule category prediction probability value, the global threshold of the entire model is calculated, thereby reflecting the current overall confidence level of the model, filtering the training data for inputting the initial nodule category prediction model, and improving the training quality of the model.

[0049] In another optional embodiment, the training ultrasound dataset is input into a preset training nodule category prediction model to obtain a local threshold, including: Perform the following operations for each type of nodule: When the predicted probability value of a category nodule is determined to be greater than the preset first target nodule probability threshold, the first target nodule is obtained; based on all the first target nodules, the number of first target nodules is obtained. When it is determined that the number of first target nodules is less than the preset threshold for the number of first target nodules, the probability threshold for the first target nodules is adjusted to obtain the adjusted probability threshold. The local threshold is obtained based on the adjusted probability threshold corresponding to each type of nodule.

[0050] In this optional embodiment, the first target nodule probability threshold can be a probability value set for any one of the following nodules: nodule of type I, nodule of type II, nodule of type III, and nodule of type IV. This probability threshold can be 70%, 80%, 90%, etc.

[0051] In this optional embodiment, the first target nodule can be a nodule with a probability greater than the first target nodule probability threshold. For example, if the probability threshold for the four types of nodules is set to 90%, when the predicted probability value for the four types of nodules is determined to be 95%, the nodule category corresponding to that probability value is determined, which is equivalent to the four types of nodules being the first target nodule; when the predicted probability value for the four types of nodules is determined to be 75%, that value is discarded, and the four types of nodules are not included in the first target nodule; the same applies to other types I, II, and III nodules.

[0052] In this optional embodiment, the first target nodule number threshold can be a minimum value for the first target nodule number. It can be set to a specific number, such as 50, 100, 200, or 1000.

[0053] In this optional embodiment, adjusting the probability threshold can be done by recognizing that if the number of first target nodules is less than a preset first target nodule number threshold, it indicates that the initial system-set first target nodule probability threshold was too high, resulting in a reduction in the sample size of such nodules. Therefore, the system can automatically adjust the first target nodule probability threshold to lower its value. For example, the first target nodule probability threshold can be adjusted from 90% to 85%, and the number of such nodules can be judged again until the number of such nodules meets the desired number.

[0054] In this optional embodiment, the local threshold may be a nodule probability threshold that includes nodules of types I, II, III, and IV; or it may be a nodule probability threshold that includes nodules of types III and IV; or it may be a probability threshold that includes nodules of type IV.

[0055] As can be seen, in this optional embodiment, by setting a specific value for the threshold of each type of nodule, nodules that meet the requirements can be filtered out, and then the specific number of nodules of this type can be obtained; then, by judging the specific number of nodules of this type through the quantity threshold, it can be determined whether the obtained number of nodules meets the requirements; the threshold of nodules that do not meet the requirements is adjusted to obtain the adjustment threshold of each type of nodule, and then the overall local threshold is obtained; through the overall local threshold, the number of nodules of each type can be guaranteed, thereby alleviating class imbalance and improving the quality of model training.

[0056] In another optional embodiment, the training ultrasound dataset is input into a preset training nodule category prediction model to obtain a conservative threshold, including: Based on the predicted probability values ​​of all nodules, the predicted probability value of the second target nodule is obtained; When the predicted probability value of the second target nodule is determined to be less than the preset conservative nodule probability threshold, the second target nodule corresponding to the predicted probability value of the second target nodule is marked as a safe nodule. When the predicted probability value of the second target nodule is determined to be greater than or equal to the preset conservative nodule probability threshold, the second target nodule corresponding to the predicted probability value of the second target nodule is marked as the first biopsy nodule. The nodule corresponding to the predicted probability value of the largest category nodule is marked as a contrast nodule; By comparing the nodule with the safe nodule, a second biopsy nodule was obtained; Based on the second biopsy nodule and the first biopsy nodule, the proportion of the second biopsy nodule is obtained; when it is determined that the proportion of the second biopsy nodule is less than the preset proportion threshold, the conservative nodule probability threshold is adjusted to obtain the conservative threshold.

[0057] In this optional embodiment, the second target nodule prediction probability value can be any of the four types of nodule probability values; statistical analysis is performed on all nodule probability values, and the four types of nodule probability values ​​are selected as the second target nodule prediction probability value; the second target nodule prediction probability value can also be the probability value corresponding to any one of the three types of nodules.

[0058] In this optional embodiment, the conservative nodule probability threshold can be set as follows: when the probability value of a nodule is greater than the conservative nodule probability threshold, the nodule corresponding to the probability value at this time is defined as a fourth-class nodule; when the probability value is less than the conservative nodule probability threshold, the nodule corresponding to the probability value at this time is defined as a third-class nodule and is no longer included in the fourth-class nodule category. This mode can be considered as the first mode.

[0059] For example, the highest probability prediction value for the first type of nodule is 95%, the highest probability prediction value for the second type of nodule is 85%, and the conservative nodule probability threshold is 90%; then the first type of nodule can be classified into four types of nodules; and the second type of nodule can be classified into three types of nodules.

[0060] In this optional embodiment, a safe nodule can be defined as all nodules classified into three categories; a first biopsy nodule can be defined as all nodules classified into four categories.

[0061] In this optional embodiment, the nodules to be compared can be in another mode, namely the second mode, which directly outputs the nodule corresponding to the highest probability value. Regardless of whether it is a third or fourth category, the nodule with the higher probability is output as the category of nodule.

[0062] In this optional embodiment, the second biopsy nodule can be determined by comparing a nodule classified as safe with one classified as a contrast nodule. Both classification modes are used to determine the same nodule. If the same nodule is a Class IV nodule, it is classified as a safe nodule in the first mode (it is actually a Class IV nodule but is classified as a Class III nodule), and as a contrast nodule in the second mode (in this case, a Class IV nodule). In this case, it is determined whether the contrast nodule corresponds to the category of the safe nodule, because the original category of a safe nodule could be either Class III or Class IV, but within the category of safe nodules, it is uniformly considered a Class III nodule. When a safe nodule (it is actually a Class IV nodule but is considered a Class III nodule) is determined to be a Class IV nodule in the contrast nodule classification, the safe nodule is re-marked as the second biopsy nodule.

[0063] Furthermore, if a safe nodule (which is itself a type IV nodule but is considered a type III nodule) is identified as a type III nodule in the comparison nodule, then the nodule is considered a type III nodule and is no longer marked as a type IV nodule.

[0064] For example, let the conservative nodule probability threshold be 90%.

[0065] Define A as a nodule. In the first mode, nodule A has a 60% probability of being classified as a Class III nodule and a 30% probability of being classified as a Class IV nodule, while the conservative nodule probability threshold is set to 80%. In this case, nodule A is defined as a safe nodule, as the nodule itself is a Class III nodule. In the second mode, nodule A has the highest probability of being classified as a Class III nodule, so it is defined as a contrastive nodule (Class III nodule). In this case, the contrastive nodule is a Class III nodule, which is also a safe nodule.

[0066] Define B as a nodule. In the first mode, nodule B is determined to be a Class III nodule with a 30% probability and a Class IV nodule with a 60% probability, while the conservative nodule probability threshold is set to 80%. At this time, nodule B is defined as a safe nodule (it has a high probability of being a Class IV nodule), which is a Class III nodule, but the nodule itself is a Class IV nodule. In the second mode, nodule B outputs a Class IV nodule, which is defined as a contrast nodule (Class IV nodule). The contrast nodule (Class IV nodule) is then matched and compared with the safe nodule (it is a Class IV nodule, but it is defined as a Class III nodule). It is found that the contrast nodule with the highest output probability is a Class IV nodule. In the first mode, this Class IV nodule was classified as a Class III nodule, so this safe nodule is reclassified as a Class IV nodule, which is the second biopsy nodule.

[0067] Define C as a nodule. In the first mode, the probability of nodule C being classified as a Class III nodule is 5%, and the probability of it being a Class IV nodule is 90%, while the conservative nodule probability threshold is set to 80%. Nodule C is defined as the first biopsy nodule. In the second mode, the output of nodule C is a Class IV nodule, which is defined as a comparison nodule. At this time, the nodule is a Class IV nodule, and after matching, it is the first biopsy nodule.

[0068] In this optional embodiment, the proportion of the second biopsy nodules can be the ratio of the number of the second biopsy nodules to the total number of biopsy nodules.

[0069] In this optional embodiment, the percentage threshold can be a manually set value; specifically, it can be 90%, 80%, etc.; when the percentage of the second biopsy nodule is lower than this percentage threshold, it indicates that the set conservative nodule probability threshold is too high and needs to be adjusted. This can reduce the model's computational power for matching data, thereby improving the model's speed; the adjusted conservative nodule probability threshold is the conservative threshold.

[0070] As can be seen, in this optional embodiment, by pre-setting a conservative nodule probability threshold, relatively certain dangerous nodules can be screened out first; the dangerous nodules excluded after screening are defined as safe nodules, which are used to match the nodules with the highest probability directly output; after matching the nodules with the highest probability output, the matching results are judged to determine whether the matched nodules with the highest probability are dangerous nodules, to prevent dangerous nodules from being misjudged; finally, the ratio of the dangerous nodules modified after matching to all dangerous nodules is judged, and the conservative nodule probability threshold is adjusted to obtain a conservative threshold, which is used to reduce the computational load of the model and improve the operation speed of the model; by directly judging nodules as safe nodules and first biopsy nodules, dangerous nodules can be directly screened out, thereby reducing the probability of false positives; then, matching safe nodules with the nodules with the highest probability directly output can prevent dangerous nodules from being missed, thereby improving the sensitivity of the model.

[0071] In another optional embodiment, the initial nodule category prediction model is trained using a first ultrasound image dataset to obtain a first nodule category prediction model, including: Data augmentation is performed on the first ultrasound image dataset to obtain an enhanced ultrasound image dataset; The initial nodule category prediction model is trained using an enhanced ultrasound image dataset and a pre-defined first cross-entropy loss function to obtain the first nodule category prediction model.

[0072] In this optional embodiment, the first cross-entropy loss function is used to constrain the training of the initial nodule category prediction model during the training process; the first cross-entropy loss function can be a cross-entropy loss function, which, as those skilled in the art know, can be a scoring tool used to measure how much the model's prediction results differ from the actual results; by processing and constraining the first ultrasound image dataset with existing labels through the first cross-entropy loss function, the model can acquire initial recognition capabilities.

[0073] In this optional embodiment, data augmentation may involve randomly horizontally flipping the first ultrasound image dataset to maintain label integrity.

[0074] As can be seen, this optional embodiment ensures the integrity of the nodule labels in the first ultrasound image dataset by randomly flipping the dataset horizontally; then, by using the first cross-entropy loss function to constrain the initial training process of the model, the model can acquire preliminary classification ability, thereby improving the nodule recognition ability of the entire model.

[0075] In another optional embodiment, the first nodule category prediction model is trained using a second ultrasound image dataset to obtain a trained nodule category prediction model, including: The second ultrasound image dataset is subjected to a first image enhancement process to obtain a first enhanced image; A second image enhancement process is performed on the second ultrasound image dataset to obtain a second enhanced image; The first enhanced image is input into the first nodule category prediction model to obtain nodule prediction labels; when it is determined that the nodule prediction label data is greater than the global threshold and the local threshold, the nodule prediction label is used to participate in the training of the first nodule category prediction model. Based on the nodule prediction label, the first nodule category prediction model is trained using the second enhanced image and the preset second cross-entropy loss function to obtain the trained nodule category prediction model.

[0076] In this optional embodiment, the first image enhancement process may be to apply weak enhancement to an ultrasound image, that is, to make minor changes to the image; the second image enhancement process may be to apply strong enhancement to an ultrasound image, that is, to make significant changes to the image; the second cross-entropy loss function may be used to constrain the training of the first nodule category prediction model during the training process.

[0077] In this optional embodiment, the nodule prediction label may be: predicting the nodules in the first enhanced image, predicting the different types of prediction values ​​for each nodule pair, and using the nodule category corresponding to the largest prediction value as the label of the nodule in the first enhanced image, that is, the nodule prediction label.

[0078] In this optional embodiment, the nodule prediction label being greater than both the global and local thresholds can be used as training image data only when all the results of the nodule prediction labels are greater than both the global and local thresholds. This is because not all pseudo-labels are reliable. If the model's prediction for a certain image is uncertain, using this uncertain prediction as a label will introduce noise.

[0079] When the nodule prediction label is less than the global threshold and the local threshold, or less than the global threshold but greater than the local threshold, or greater than the global threshold but less than the local threshold, the ultrasound image data will be discarded to improve the quality of the training image data.

[0080] In this optional embodiment, the predicted nodule label after judgment is used as a control group, and the first nodule category prediction model is trained using the training dataset of the second enhanced image; during the training process, the second cross-entropy loss function is used to constrain the model training; the second cross-entropy loss function can be the cross-entropy loss function, in which case the cross-entropy loss function applies to the dataset of nodule predicted labels and the training dataset of the second enhanced image.

[0081] As can be seen, by processing the second ultrasound image dataset with different intensities, a set of first enhanced images for comparison and a set of second enhanced images for training can be obtained. By first processing the first enhanced images to obtain nodule prediction labels, the nodule prediction labels are only used as a control for training when the data of the nodule prediction labels is greater than the global threshold and the local threshold. Finally, using the nodule prediction labels as control data, the second enhanced images are used to train and compare the first nodule category prediction model. At the same time, the second cross-entropy loss function is used for constraint to obtain a well-trained nodule category prediction model, which improves the accuracy of the model in nodule identification.

[0082] Example 2 Please see Figure 2 , Figure 2 This is a schematic flowchart of an ultrasound image nodule identification method disclosed in an embodiment of the present invention. Figure 2 The described ultrasound image nodule identification method can be applied to an ultrasound image nodule identification device, which can be integrated into a cloud server or a local server; this embodiment of the invention is not limited thereto. Figure 2 As shown, the ultrasound image nodule identification method may include the following operations: Step 201: Obtain the ultrasound image to be predicted.

[0083] In this optional embodiment, the ultrasound image to be predicted can be an already captured ultrasound image, such as a thyroid ultrasound image or a breast ultrasound image.

[0084] Step 202: Input the ultrasound image to be predicted into the nodule category prediction model to obtain the output of the nodule category prediction model; based on the output, obtain the category of the nodule in the ultrasound image to be predicted.

[0085] In this optional embodiment, the nodule category prediction model can be trained using any of the nodule category prediction model training methods in Embodiment 1.

[0086] In this optional embodiment, the output result may be that the ultrasound image has been processed and the result is already in the ultrasound image, but the result has not yet been converted into visual information.

[0087] As can be seen, in this optional embodiment, the ultrasound image to be predicted is first obtained, and the ultrasound image is input into the trained nodule category prediction model, which can accurately predict the nodules in the ultrasound image and obtain the category of the nodules in the ultrasound image.

[0088] Example 3 Please see Figure 3 , Figure 3 This is a schematic diagram of a nodule category prediction model training device disclosed in an embodiment of the present invention. The device may include: The pre-training module 301 is used to acquire the training ultrasound dataset; the training ultrasound dataset is input into the pre-set training nodule category prediction model to obtain the global threshold, local threshold, and conservative threshold; the global threshold is used to reflect the overall confidence level of the model; the local threshold is used to adjust the number of nodule categories identified; the conservative threshold is used to reduce the false positive rate of the target nodule; the global threshold, local threshold, and conservative threshold are fed back into the training nodule category prediction model to obtain the initial nodule category prediction model; The training data acquisition module 302 is used to acquire a first ultrasound image dataset and a second ultrasound image dataset; the first ultrasound image dataset consists of labeled ultrasound images, and the second ultrasound image dataset consists of unlabeled ultrasound images. The formal training module 303 is used to train the initial nodule category prediction model using the first ultrasound image dataset to obtain the first nodule category prediction model; and to train the first nodule category prediction model using the second ultrasound image dataset to obtain the trained nodule category prediction model; the nodule category prediction model is used to identify the category of nodules in ultrasound images.

[0089] As can be seen, in this embodiment of the invention, a training ultrasound dataset is first acquired, and a preset training nodule category prediction model is trained to obtain a global threshold, a local threshold, and a conservative threshold. These thresholds are then fed back into the training nodule category prediction model to obtain an initial nodule category prediction model. Next, an ultrasound image dataset is acquired. The preset initial nodule category prediction model is first trained using a labeled first ultrasound image dataset to obtain a first nodule category prediction model. Through this initial training, the first nodule category prediction model can acquire preliminary classification capabilities. Then, the first nodule category prediction model is trained using an unlabeled second ultrasound image dataset to obtain a trained nodule category prediction model. By first obtaining the global threshold, local threshold, and conservative threshold through training, the confidence level of the initial nodule category prediction model is improved, the number of different nodule categories is increased, and the false positive rate of nodules is reduced. Finally, the nodule category prediction model is trained to improve the accuracy of nodule prediction.

[0090] In an optional embodiment, the pre-training module 301 inputs the training ultrasound dataset into a preset training nodule category prediction model to obtain the global threshold. The specific operation may include: The training ultrasound dataset is input into a preset training nodule category prediction model to predict the nodule category, and the prediction probability value of the nodule category corresponding to each ultrasound image in the training ultrasound dataset is obtained. Based on the predicted probability values ​​of all nodule categories, the predicted probability value of the largest nodule category corresponding to each ultrasound image is obtained; The global threshold is obtained by predicting the probability values ​​of all the largest category nodules.

[0091] As can be seen, in this optional embodiment, by calculating the nodule category prediction probability value, the global threshold of the entire model is calculated, thereby reflecting the current overall confidence level of the model, filtering the training data for inputting the initial nodule category prediction model, and improving the training quality of the model.

[0092] In another optional embodiment, the pre-training module 301 inputs the training ultrasound dataset into a preset training nodule category prediction model to obtain the specific operation method of the local threshold, which may include: Perform the following operations for each type of nodule: When the predicted probability value of a category nodule is determined to be greater than the preset first target nodule probability threshold, the first target nodule is obtained; based on all the first target nodules, the number of first target nodules is obtained. When it is determined that the number of first target nodules is less than the preset threshold for the number of first target nodules, the probability threshold for the first target nodules is adjusted to obtain the adjusted probability threshold. The local threshold is obtained based on the adjusted probability threshold corresponding to each type of nodule.

[0093] As can be seen, in this optional embodiment, by setting a specific value for the threshold of each type of nodule, nodules that meet the requirements can be filtered out, and then the specific number of nodules of this type can be obtained; then, by judging the specific number of nodules of this type through the quantity threshold, it can be determined whether the obtained number of nodules meets the requirements; the threshold of nodules that do not meet the requirements is adjusted to obtain the adjustment threshold of each type of nodule, and then the overall local threshold is obtained; through the overall local threshold, the number of nodules of each type can be guaranteed, thereby alleviating class imbalance and improving the quality of model training.

[0094] In another optional embodiment, the pre-training module 301 inputs the training ultrasound dataset into a preset training nodule category prediction model to obtain the specific operation method of the conservative threshold, which may include: Based on the predicted probability values ​​of all nodules, the predicted probability value of the second target nodule is obtained; When the predicted probability value of the second target nodule is determined to be less than the preset conservative nodule probability threshold, the second target nodule corresponding to the predicted probability value of the second target nodule is marked as a safe nodule. When the predicted probability value of the second target nodule is determined to be greater than or equal to the preset conservative nodule probability threshold, the second target nodule corresponding to the predicted probability value of the second target nodule is marked as the first biopsy nodule. The nodule corresponding to the predicted probability value of the largest category nodule is marked as a contrast nodule; By comparing the nodule with the safe nodule, a second biopsy nodule was obtained; Based on the second biopsy nodule and the first biopsy nodule, the proportion of the second biopsy nodule is obtained; when it is determined that the proportion of the second biopsy nodule is less than the preset proportion threshold, the conservative nodule probability threshold is adjusted to obtain the conservative threshold.

[0095] As can be seen, by pre-setting a conservative nodule probability threshold, relatively certain dangerous nodules can be screened out first. The dangerous nodules excluded after screening are defined as safe nodules, used to match the highest probability nodules directly output. After matching the highest probability nodules, the matching results are evaluated to determine whether the matched highest probability nodules are dangerous nodules, preventing false positives. Finally, the ratio of the modified dangerous nodules after matching to all dangerous nodules is compared, and the conservative nodule probability threshold is adjusted to obtain a conservative threshold, which reduces the computational load of the model and improves its processing speed. By directly classifying nodules as safe nodules and first biopsy nodules, dangerous nodules can be directly screened out, thereby reducing the probability of false positives. Then, matching safe nodules with the highest probability nodules directly output can prevent dangerous nodules from being missed, thus improving the model's sensitivity.

[0096] In another optional embodiment, the formal training module 303 uses the first ultrasound image dataset to train the initial nodule category prediction model to obtain the specific operation mode of the first nodule category prediction model, which may include: Data augmentation is performed on the first ultrasound image dataset to obtain an enhanced ultrasound image dataset; The initial nodule category prediction model is trained using an enhanced ultrasound image dataset and a preset first cross-entropy loss function to obtain the first nodule category prediction model; wherein, the first cross-entropy loss function is used to constrain the training of the initial nodule category prediction model during the training process.

[0097] As can be seen, by randomly flipping the first ultrasound image dataset horizontally, the integrity of the nodule labels in the first ultrasound image dataset is ensured; then, by using the first cross-entropy loss function to constrain the initial training process of the model, the model can acquire preliminary classification ability, thereby improving the nodule recognition ability of the entire model.

[0098] In another optional embodiment, the formal training module 303 uses a second ultrasound image dataset to train the first nodule category prediction model, and the specific operation of obtaining the trained nodule category prediction model may include: The second ultrasound image dataset is subjected to a first image enhancement process to obtain a first enhanced image; A second image enhancement process is performed on the second ultrasound image dataset to obtain a second enhanced image; The first enhanced image is input into the first nodule category prediction model to obtain nodule prediction labels; when it is determined that the nodule prediction label data is greater than the global threshold and the local threshold, the nodule prediction label is used to participate in the training of the first nodule category prediction model. Based on the nodule prediction label, the first nodule category prediction model is trained using the second enhanced image and the preset second cross-entropy loss function to obtain the trained nodule category prediction model. The second cross-entropy loss function is used to constrain the training of the first nodule category prediction model during the training process.

[0099] As can be seen, by processing the second ultrasound image dataset with different intensities, a set of first enhanced images for comparison and a set of second enhanced images for training can be obtained. By first processing the first enhanced images to obtain nodule prediction labels, the nodule prediction labels are only used as a control for training when the data of the nodule prediction labels is greater than the global threshold and the local threshold. Finally, using the nodule prediction labels as control data, the second enhanced images are used to train and compare the first nodule category prediction model. At the same time, the second cross-entropy loss function is used for constraint to obtain a well-trained nodule category prediction model, which improves the accuracy of the model in nodule identification.

[0100] Example 4 Please see Figure 4 , Figure 4 This is a schematic diagram of the structure of an ultrasound image nodule identification device disclosed in an embodiment of the present invention. The device may include: Image acquisition module 401 is used to acquire ultrasound images to be predicted.

[0101] The nodule identification module 402 is used to input the ultrasound image to be predicted into the nodule category prediction model and obtain the output result of the nodule category prediction model; based on the output result, the category of the nodule in the ultrasound image to be predicted is obtained.

[0102] In this optional embodiment, the nodule category prediction model is trained using any of the nodule category prediction model training methods in Embodiment 1.

[0103] As can be seen, in this optional embodiment, the ultrasound image to be predicted is first obtained, and the ultrasound image is input into the trained nodule category prediction model to accurately predict the nodules in the ultrasound image and obtain the category of the nodules in the ultrasound image.

[0104] Example 5 Please see Figure 5 , Figure 5 This is a schematic diagram of the structure of a device including a memory and a processor, as disclosed in an embodiment of the present invention. Figure 5 As shown, the device may include: Memory 501 storing executable program code; Processor 502 coupled to memory 501; The processor 502 calls the executable program code stored in the memory 501 to execute some or all of the steps in the nodule category prediction model training method described in Embodiment 1 of the present invention or the ultrasound image nodule recognition method described in Embodiment 2.

[0105] Example 6 The steps in Examples 1 and 2 are applied to breast nodule identification. The specific operation process is as follows: 1. Task Data: Given a breast ultrasound image, the system automatically outputs its BI-RADS classification (classes 1-4). During the screening process, after the ultrasound physician acquires the image, BiCAT (BI-RADS Classification Adaptive Consistency Transformer model) outputs predicted classes, confidence levels, and heatmaps to assist the radiologist in making the final assessment.

[0106] In screening practice, a large number of ultrasound images are acquired and stored, but only a small fraction receive BI-RADS (Nodule Classification) annotations from radiologists. For example, in the development data of this protocol, 121,695 images from 63,326 screening patients were unannotated, while only 1,570 images (from 1,443 patients) had BI-RADS annotations. The class distribution in these annotated data is severely unbalanced: BI-RADS class 1 accounts for 49.0%, class 2 for 4.5%, class 3 for 44.0%, and class 4 for only 1.7%. The core idea of ​​this protocol is to utilize this large amount of unannotated data to assist in training the model with a small amount of annotated data.

[0107] 2. Model Architecture: Breast ultrasound images commonly exhibit speckle noise, which can interfere with the identification of key diagnostic features such as microcalcifications and irregular edges. To address this, BiCAT employs a DnCNN (Denoising Convolutional Neural Network) as its front-end module. This module first extracts and denoises the images, enhancing diagnostically relevant features. The output of the DnCNN is fused with the original input through residual connections, thus preserving the original image information while incorporating enhanced features after denoising. The fused features are then fed into a ResNet (Residual Network) for deep learning, and finally, a softmax activation function is used to output the predicted probabilities for four BI-RADS categories.

[0108] 3. Two-stage semi-supervised training process: If supervised learning is trained using only 1,570 labeled images, the model's performance is limited by insufficient data and class imbalance. To utilize 121,695 unlabeled images, this approach employs a two-stage semi-supervised training strategy.

[0109] like Figure 6 As shown, Figure 6This is a schematic diagram of a two-stage semi-supervised training process disclosed in this scheme; the first stage is pre-training: supervised learning is performed using only labeled data, optimized using a multi-class cross-entropy loss function, and data augmentation is limited to random horizontal flipping to maintain label integrity. The purpose of this stage is to enable the model to acquire preliminary classification capabilities.

[0110] The second stage is augmentation training: based on the pre-trained model, unlabeled data is introduced for semi-supervised learning. The basic idea is consistency regularization—applying different levels of enhancement to an ultrasound image should not change its BI-RADS category. Specifically, weak enhancement and strong enhancement are applied to each unlabeled image, and the model's prediction for the weakly enhanced image is used as a "pseudo-label" to train the model to give the same prediction for the strongly enhanced image. Weak enhancement includes mild transformations such as random speckle noise, random occlusion, and Rand Augment (random automatic enhancement); strong enhancement is based on weak enhancement with additional transformations.

[0111] However, not all pseudo-labels are reliable. If the model's prediction for a particular image is uncertain, using this uncertain prediction as a label will introduce noise. To address this, our approach employs an adaptive threshold filtering mechanism to filter low-quality pseudo-labels. This mechanism includes two thresholds: a global threshold reflecting the model's overall confidence level, calculated based on the exponential moving average of the maximum predicted probability within a batch; and a local threshold set individually for each BI-RADS category, adjusted based on the difference between the predicted distribution and the empirical distribution for that category—for categories with fewer model predictions (e.g., BI-RADS 4 categories), the threshold is adaptively lowered, allowing more pseudo-labels for that category to be retained, thus mitigating class imbalance. Only pseudo-labels exceeding both thresholds are allowed to participate in training; the final loss function is the sum of the supervised loss of the labeled data and the loss of the filtered pseudo-labels.

[0112] 4. Adjustable classification mode: In screening practice, the distinction between BI-RADS categories 3 and 4 is the most critical decision point—category 3 implies short-term follow-up, while category 4 implies biopsy. Different clinical environments tend to favor this distinction: in routine physical examinations, due to low prevalence, physicians tend to raise the threshold to reduce unnecessary biopsies; while in high-risk population screening, they tend to lower the threshold to avoid missed diagnoses.

[0113] Please see Figure 7 , Figure 7This diagram illustrates the differences in classification and borderline case changes between the conservative and open modes disclosed in this scheme. It provides BiCAT with two modes: the conservative mode sets a 90% probability threshold for BI-RADS 4 categories, automatically classifying borderline cases below the threshold as category 3, reducing the false positive rate from approximately 2% (radiologist level) to approximately 1.1%; the open mode directly outputs the category with the highest probability, retaining all borderline cases as category 4, thus improving sensitivity by 9%–21% compared to the conservative mode. The two modes are switched via a single threshold parameter.

[0114] 5. Confidence stratification and interpretability: In large-scale screening, it is impractical for radiologists to review all AI predictions one by one. BiCAT uses adaptive thresholds generated by semi-supervised training to categorize predictions into three levels: "very confident," "confident," and "not confident." High-confidence predictions can be directly accepted, while low-confidence cases prompt physicians to focus on reviewing them, achieving an efficient selective review workflow.

[0115] Furthermore, this solution uses Grad-CAM++ to generate heatmaps, visualizing the image regions that the model focuses on during classification decisions. For example... Figure 8 As shown, Figure 8 This is a schematic diagram of an ultrasound image nodule thermogram disclosed in this solution. The thermogram can locate BI-RADS diagnostic features such as microcalcifications, irregular edges, and abnormal aspect ratios, helping physicians understand the basis of the model's judgment and solving the AI ​​"black box" problem.

[0116] 6. Advantages and positive effects of this plan: First, we verified the effectiveness of semi-supervised learning. As shown in Table 1, Table 1 compares the performance of purely supervised learning and semi-supervised learning. It can be seen that after introducing unlabeled images, the accuracy improved by 6 percentage points and the AUC improved by 2.8 percentage points (p<0.001), indicating that semi-supervised learning can effectively utilize unlabeled data to overcome the bottleneck of insufficient labeling.

[0117] Table 1. Performance comparison of semi-supervised learning and purely supervised learning (internal test set, 314 images)

[0118] As shown in Table 2, BiCAT's performance on prospective and external validation is demonstrated. On prospective cohorts preserving the true prevalence, BiCAT achieved an accuracy of 79.7% and an AUC of 0.943; on external multicenter data, the AUC was 0.835, validating its cross-center generalization ability.

[0119] Table 2. Results of prospective and external multicenter validation

[0120] As shown in Table 3, BiCAT was compared with the image interpretation of 10 radiologists. BiCAT showed better accuracy and consistency than the three groups of radiologists, and also had higher sensitivity in BI-RADS Class 4.

[0121] Table 3. Comparison of BiCAT results with radiologists (234 cases of image reading)

[0122] As shown in Table 4, this table compares the two classification models with radiologists in a prospective cohort. The false positive rate of the conservative model is only about half that of radiologists.

[0123] Table 4. Comparison of classification patterns and radiologists in a prospective cohort.

[0124] In terms of confidence stratification, high-confidence predictions covered 85.3% of cases with an accuracy rate of 84.4% (out of an overall 79.7%). This means that in large-scale screening, radiologists only need to focus on reviewing about 15% of low-confidence cases, significantly improving efficiency.

[0125] The results above show that this method effectively overcomes the problem of insufficient annotation through semi-supervised learning, and achieves performance comparable to or even better than that of radiologists in BI-RADS four-class classification. At the same time, the flexible dual-mode and confidence stratification can be adapted to different clinical screening scenarios.

[0126] Through the detailed description of the above embodiments, those skilled in the art can clearly understand that each implementation method can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, including read-only memory (ROM), random access memory (RAM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), one-time programmable read-only memory (OTPROM), electrically-Erasable Programmable Read-Only Memory (EEPROM), compact disc read-only memory (CD-ROM) or other optical disc storage, disk storage, magnetic tape storage, or any other computer-readable medium that can be used to carry or store data.

[0127] Finally, it should be noted that the above embodiments are merely preferred embodiments of the present invention and are only used to illustrate the technical solutions of the present invention, not to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for training a nodule category prediction model, characterized in that, The method includes: Obtain the training ultrasound dataset; The training ultrasound dataset is input into a preset training nodule category prediction model to obtain a global threshold and a local threshold. The global threshold is used to reflect the overall confidence level of the model. The local threshold is used to adjust the number of nodule categories identified. The training ultrasound dataset is input into the preset training nodule category prediction model to perform nodule category prediction and obtain the nodule category prediction probability value corresponding to each ultrasound image in the training ultrasound dataset. Based on the predicted probability values ​​of all the nodules of the aforementioned categories, the predicted probability value of the second target nodule and the predicted probability value of the largest category nodule corresponding to each ultrasound image are obtained. When it is determined that the predicted probability value of the second target nodule is less than the preset conservative nodule probability threshold, the second target nodule corresponding to the predicted probability value of the second target nodule is marked as a safe nodule; when it is determined that the predicted probability value of the second target nodule is greater than or equal to the preset conservative nodule probability threshold, the second target nodule corresponding to the predicted probability value of the second target nodule is marked as a first biopsy nodule. The nodule corresponding to the predicted probability value of the largest category nodule is marked as a contrast nodule; the contrast nodule and the safe nodule are matched to obtain the second biopsy nodule; the proportion of the second biopsy nodule is obtained based on the second biopsy nodule and the first biopsy nodule; when it is determined that the proportion of the second biopsy nodule is less than a preset proportion threshold, the conservative nodule probability threshold is adjusted to obtain a conservative threshold; the conservative threshold is used to reduce the false positive rate of the target nodule; The global threshold, the local threshold, and the conservative threshold are fed back into the trained nodule category prediction model to obtain the initial nodule category prediction model; Obtain a first ultrasound image dataset and a second ultrasound image dataset; the first ultrasound image dataset consists of labeled ultrasound images, and the second ultrasound image dataset consists of unlabeled ultrasound images. The initial nodule category prediction model is trained using the first ultrasound image dataset to obtain a first nodule category prediction model; the first nodule category prediction model is trained using the second ultrasound image dataset to obtain a trained nodule category prediction model; the nodule category prediction model is used to identify the category of nodules in ultrasound images.

2. The training method for the nodule category prediction model according to claim 1, characterized in that, The step of inputting the training ultrasound dataset into a preset training nodule category prediction model to obtain a global threshold includes: A global threshold is obtained based on the predicted probability values ​​of all the largest category nodules.

3. The training method for the nodule category prediction model according to claim 2, characterized in that, The step of inputting the training ultrasound dataset into a preset training nodule category prediction model to obtain a local threshold includes: Perform the following operations for each type of nodule: When it is determined that the predicted probability value of the nodule category is greater than the preset first target nodule probability threshold, a first target nodule is obtained; based on all the first target nodules, the number of first target nodules is obtained. When it is determined that the number of the first target nodules is less than the preset threshold for the number of the first target nodules, the probability threshold for the first target nodules is adjusted to obtain the adjusted probability threshold. A local threshold is obtained based on the adjusted probability threshold corresponding to each type of nodule.

4. The training method for the nodule category prediction model according to claim 3, characterized in that, The step of training the initial nodule category prediction model using the first ultrasound image dataset to obtain the first nodule category prediction model includes: The first ultrasound image dataset is augmented to obtain an augmented ultrasound image dataset. The initial nodule category prediction model is trained using the enhanced ultrasound image dataset and a preset first cross-entropy loss function to obtain a first nodule category prediction model; wherein, the first cross-entropy loss function is used to constrain the training of the initial nodule category prediction model during the training process.

5. The training method for the nodule category prediction model according to claim 4, characterized in that, The step of training the first nodule category prediction model using the second ultrasound image dataset to obtain a trained nodule category prediction model includes: The second ultrasound image dataset is subjected to a first image enhancement process to obtain a first enhanced image; The second ultrasound image dataset is subjected to a second image enhancement process to obtain a second enhanced image; The first enhanced image is input into the first nodule category prediction model to obtain nodule prediction labels; when it is determined that the data of the nodule prediction label is greater than the global threshold and the local threshold, the nodule prediction label is used to participate in the training of the first nodule category prediction model. Based on the nodule prediction label, the first nodule category prediction model is trained using the second enhanced image and a preset second cross-entropy loss function to obtain a trained nodule category prediction model. The second cross-entropy loss function is used to constrain the training of the first nodule category prediction model during the training process.

6. A method for nodule identification in ultrasound images, characterized in that, The method includes: Acquire the ultrasound image to be predicted; The ultrasound image to be predicted is input into the nodule category prediction model to obtain the output result of the nodule category prediction model; based on the output result, the category of the nodule in the ultrasound image to be predicted is obtained; wherein, the nodule category prediction model is trained by any of the nodule category prediction model training methods of claims 1-5.

7. A training device for a nodule category prediction model, characterized in that, The device includes: A pre-training module is used to acquire a training ultrasound dataset; input the training ultrasound dataset into a preset training nodule category prediction model to obtain a global threshold and a local threshold; the global threshold is used to reflect the overall confidence level of the model; the local threshold is used to adjust the number of nodule categories identified; input the training ultrasound dataset into the preset training nodule category prediction model to perform nodule category prediction, and obtain the nodule category prediction probability value corresponding to each ultrasound image in the training ultrasound dataset; based on all the nodule category prediction probability values, obtain the second target nodule prediction probability value and the maximum nodule category prediction probability value corresponding to each ultrasound image; when it is determined that the second target nodule prediction probability value is less than a preset conservative nodule probability threshold, the second target nodule corresponding to the second target nodule prediction probability value is... The nodules are marked as safe nodules. When the predicted probability value of the second target nodule is greater than or equal to a preset conservative nodule probability threshold, the second target nodule corresponding to the predicted probability value of the second target nodule is marked as a first biopsy nodule. The nodule corresponding to the predicted probability value of the largest category nodule is marked as a contrast nodule. The contrast nodule and the safe nodule are matched to obtain the second biopsy nodule. The proportion of the second biopsy nodule is obtained based on the second biopsy nodule and the first biopsy nodule. When the proportion of the second biopsy nodule is less than a preset proportion threshold, the conservative nodule probability threshold is adjusted to obtain a conservative threshold. The conservative threshold is used to reduce the false positive rate of the target nodule. The global threshold, the local threshold, and the conservative threshold are fed back into the training nodule category prediction model to obtain the initial nodule category prediction model. The training data acquisition module is used to acquire a first ultrasound image dataset and a second ultrasound image dataset; the first ultrasound image dataset consists of labeled ultrasound images, and the second ultrasound image dataset consists of unlabeled ultrasound images. The formal training module is used to train the initial nodule category prediction model using the first ultrasound image dataset to obtain a first nodule category prediction model; and to train the first nodule category prediction model using the second ultrasound image dataset to obtain a trained nodule category prediction model; the nodule category prediction model is used to identify the category of nodules in ultrasound images.

8. An ultrasound image nodule identification device, characterized in that, The device includes: Image acquisition module, used to acquire the ultrasound image to be predicted; The nodule identification module is used to input the ultrasound image to be predicted into the nodule category prediction model to obtain the output result of the nodule category prediction model; and to obtain the category of the nodule in the ultrasound image to be predicted based on the output result; wherein the nodule category prediction model is trained by any of the nodule category prediction model training methods of claims 1-5.

9. An apparatus comprising a memory and a processor, characterized in that, The device includes: A memory storing executable program code, and a processor coupled to the memory; The processor calls the executable program code stored in the memory to execute the nodule category prediction model training method as described in any one of claims 1-5 or the ultrasound image nodule recognition method as described in claim 6.

Images