Training set acquisition method, model training method, nodule region identification method and related products

Abstract

The application relates to the technical field of image processing, and discloses a training set acquisition method, a model training method, a nodule region identification method and related products. The method comprises the following steps: acquiring an initial training set; generating a plurality of sub-region images of each chest CT image; training a neural network model for N rounds to obtain an identification result of each sub-region image of each chest CT image; acquiring an effective detection frame corresponding to each chest CT image; for each chest CT image, calculating an image-level loss and a fast loss mean value corresponding to each round; calculating an evaluation score corresponding to each chest CT image according to the fast loss mean value corresponding to the Nth round; and acquiring an optimized training set according to the evaluation score sorting result of each chest CT image. The training set acquired by the method can be used to train a nodule region identification model of a CT image with high identification result accuracy.

Description

Technical Field

[0001] This application relates to the field of image processing technology, and in particular to a training set acquisition method, a model training method, a nodule region recognition method, and related products. Background Technology

[0002] Lung nodule detection is a crucial step in lung cancer screening and follow-up. Computed tomography (CT) of the lungs is the primary method for screening and diagnosing early-stage lung nodules in clinical practice. In recent years, with the rapid development of artificial intelligence technology, deep learning models are often used to identify nodule regions on chest CT images in order to reduce the workload of radiologists. However, deep learning models typically rely on a large amount of bounding box labeled data for supervised training, but real clinical data often suffers from problems such as missed labels, mislabeling, inconsistent labeling scales, and inconsistent labeling standards across multiple centers.

[0003] In existing deep learning model training methods, all collected chest CT image data is typically mixed and directly used for iterative training of the neural network. This conventional training method has the following obvious drawbacks: the neural network is prone to overfitting to these mislabeled or extremely abnormal "high-risk" samples in the later stages of training, causing the model to deviate from the correct feature learning direction, thereby reducing the model's generalization ability and detection accuracy in practical applications, resulting in low accuracy in identifying target regions using neural network models on chest CT images. Summary of the Invention

[0004] To address the technical problem of low accuracy in target region identification using neural network models in chest CT images in existing technologies, this invention provides technical solutions in the following aspects.

[0005] In a first aspect, the present invention provides a method for obtaining an optimized training set of chest CT images, comprising: Obtain an initial training set, which includes multiple chest CT images and labels corresponding to each chest CT image; Multiple sub-region images are generated for each of the chest CT images in the initial training set; The neural network model is trained for N rounds using the multiple sub-region images of each of the chest CT images to obtain the recognition result of each sub-region image of each chest CT image, and the recognition result includes a nodule region detection box; For each of the chest CT images, valid detection frames are selected from the corresponding nodule region detection frames; For each chest CT image, the mean of image-level loss and fast loss for each round is calculated. The image-level loss for round j is the weighted average of the losses of each sub-region of the chest CT image in round j, where the sub-region loss is the weighted average of the classification losses of each valid detection box in the sub-region, 1 ≤ j ≤ N. The mean of fast loss for round k is positively correlated with the mean of fast loss for round (k-1) and the image-level loss for round k, 2 ≤ k ≤ N. The mean of fast loss for round 1 is equal to the image-level loss for round 1. Based on the mean rapid loss of each chest CT image corresponding to the Nth round, an evaluation score is calculated for each chest CT image, and the evaluation score is positively correlated with the corresponding mean rapid loss. Based on the evaluation scores of each chest CT image, a predetermined number of target chest CT images are selected from the multiple chest CT images. The target chest CT images are removed from the initial training set, and the remaining chest CT images are used as the optimization training set.

[0006] Preferably, the step of selecting valid detection boxes from the corresponding nodule region detection boxes for each chest CT image includes: for each chest CT image, dividing all corresponding nodule region detection boxes into positive sample detection boxes and negative sample detection boxes, selecting valid negative sample detection boxes from all negative sample detection boxes using random sampling or positive-negative sample ratio sampling, and taking all corresponding positive sample detection boxes and the selected valid negative sample detection boxes as valid detection boxes.

[0007] Preferably, for each of the chest CT images, the mean of the fast loss for each round is calculated, including: Calculate the first product of the fast loss mean and the first preset multiple corresponding to the (k-1)th round, calculate the second product of the image-level loss and the second preset multiple corresponding to the kth round, and determine the sum of the first product and the second product as the fast loss mean of the kth round, where 2≤k≤N.

[0008] Preferably, calculating the evaluation score of each chest CT image based on the mean of the rapid loss of each chest CT image corresponding to the Nth round includes: After the k-th round of training, for each chest CT image, calculate the average slow loss of the (k-1)-th round and the third product of the third preset multiple, calculate the image-level loss of the k-th round and the fourth product of the fourth preset multiple, and determine the sum of the third product and the fourth product as the average slow loss of the k-th round, where 2≤k≤N; the average slow loss corresponding to the first round of training is equal to the image-level loss of the corresponding chest CT image in the first round of training; the value of the third preset multiple is greater than the value of the first preset multiple; For each of the chest CT images, the reduction ratio is calculated based on the difference between the mean slow loss and the mean fast loss in the kth round, and the reduction ratio is positively correlated with the difference. Based on the mean rapid loss and the decrease ratio of each chest CT image corresponding to the Nth round, the evaluation score of each chest CT image is calculated. The evaluation score is positively correlated with the corresponding mean rapid loss and negatively correlated with the corresponding decrease ratio.

[0009] Preferably, calculating the evaluation score for each chest CT image based on the mean rapid loss and the decrease ratio of each chest CT image corresponding to the Nth round includes: For each of the chest CT images, the loss deviation between the image-level loss in the k-th round and the image-level loss in the (k-1)-th round is calculated. The fifth product of the loss deviation and a fifth preset multiple is calculated. The jitter score in the (k-1)-th round and a sixth preset multiple are calculated. The sum of the fifth and sixth product values ​​is determined as the jitter score in the k-th round. The jitter score corresponding to the first round is 0, and 2≤k≤N. The evaluation score of each chest CT image is calculated based on the mean rapid loss, the decrease ratio, and the jitter score of each chest CT image corresponding to the Nth round. The evaluation score is positively correlated with the corresponding mean rapid loss and the corresponding jitter score, and negatively correlated with the corresponding decrease ratio.

[0010] Preferably, the step of calculating the evaluation score of each chest CT image based on the mean of rapid loss, the decrease ratio, and the jitter score of each chest CT image corresponding to the Nth round includes: Calculate the first difference between 1 and the stated decrease ratio; The mean of rapid loss, the jitter score, and the first difference of each chest CT image corresponding to the Nth round are added together to obtain the evaluation score of each chest CT image.

[0011] In a second aspect, the present invention provides a method for training a CT image nodule region recognition model, comprising: The optimized training set is obtained using the CT image optimization training set acquisition method of the present invention; The neural network model is further trained using the optimized training set to obtain a CT image nodule region recognition model.

[0012] In a third aspect, the present invention provides a method for identifying nodule regions in CT images, comprising: Acquire the chest CT image to be identified; The chest CT image to be identified is input into the CT image nodule region recognition model to obtain the nodule region recognition result. The CT image nodule region recognition model is pre-trained using the CT image nodule region recognition model training method of the present invention.

[0013] In a fourth aspect, the present invention provides an apparatus for acquiring an optimized training set of chest CT images, comprising: An initial training set acquisition module is used to acquire an initial training set, which includes multiple chest CT images and labels corresponding to each chest CT image. The sub-region image generation module is used to generate multiple sub-region images for each of the chest CT images in the initial training set. The nodule region detection box generation module is used to train the neural network model for N rounds using the multiple sub-region images of each chest CT image to obtain the recognition result of each sub-region image of each chest CT image, and the recognition result includes the nodule region detection box. The effective detection box acquisition module is used to filter out effective detection boxes from the corresponding nodule region detection boxes for each chest CT image. The loss calculation module is used to calculate the image-level loss and the mean fast loss for each round for each chest CT image. The image-level loss in round j is the weighted average of the losses of each sub-region of the chest CT image in round j, and the sub-region loss is the weighted average of the classification losses of each valid detection box in the sub-region, 1 ≤ j ≤ N. The mean fast loss in round k is positively correlated with the mean fast loss in round (k-1) and the image-level loss in round k, 2 ≤ k ≤ N. The mean fast loss in round 1 is equal to the image-level loss in round 1. The evaluation score calculation module is used to calculate the evaluation score of each chest CT image based on the mean of the rapid loss of each chest CT image corresponding to the Nth round, wherein the evaluation score is positively correlated with the mean of the corresponding rapid loss. The optimized training set acquisition module is used to select a preset number of target chest CT images from multiple chest CT images based on the evaluation score ranking results of each chest CT image, remove the target chest CT images from the initial training set, and use the remaining chest CT images as the optimized training set.

[0014] In a fifth aspect, the present invention provides a training apparatus for a CT image nodule region recognition model, comprising: An optimized training set acquisition module is used to acquire the optimized training set using the CT image optimized training set acquisition method of the present invention; The training module is used to further train the neural network model using the optimized training set to obtain a CT image nodule region recognition model.

[0015] In a sixth aspect, the present invention provides a device for identifying nodule regions in CT images, comprising: The chest CT image acquisition module is configured to acquire the chest CT image to be identified. The nodule region recognition result generation module is configured to input the chest CT image to be recognized into the CT image nodule region recognition model to obtain the nodule region recognition result. The CT image nodule region recognition model is pre-trained using the CT image nodule region recognition model training method of the present invention.

[0016] In a seventh aspect, the present invention provides a nodule region identification system for CT images, comprising: Image acquisition device, used to acquire chest CT images to be identified; The processing device is communicatively connected to the image acquisition device. The processing device stores computer instructions, which, when executed, implement the nodule region recognition method for CT images of the present invention to obtain nodule region recognition results. An output device is used to output the nodule region identification results.

[0017] In an eighth aspect, the present invention provides a computer device comprising a processor and a memory, the memory storing a computer program, the processor being configured to execute the computer program to implement the method for acquiring an optimized training set of chest CT images, or a method for training a CT image nodule region recognition model, or a method for recognizing nodule regions of CT images according to the present invention.

[0018] In a ninth aspect, the present invention provides a computer storage medium storing a computer program that, when executed on a processor, implements the method for obtaining an optimized training set of chest CT images, the method for training a CT image nodule region recognition model, or the method for recognizing nodule regions in CT images.

[0019] In a tenth aspect, the present invention provides a computer program product comprising a computer-readable storage medium having program instructions contained therein, the program instructions being executable by a processor to implement the method for acquiring an optimized training set of chest CT images, or a method for training a CT image nodule region recognition model, or a method for recognizing nodule regions of CT images.

[0020] This application offers the following advantages: The chest CT image denoising training method of this invention generates multiple sub-region images of the chest CT image and extracts nodule region detection boxes from them to filter out effective detection boxes. It constructs a refined processing logic from "global image" to "local effective features." This technique effectively overcomes the dilution effect of large areas of non-lung tissue and healthy background regions in chest CT images on the model gradient, avoids noise interference from massive redundant backgrounds, and forces the model to focus more on the real nodule feature distribution, significantly improving the targeting of feature extraction and the utilization rate of computational resources. Simultaneously, this invention breaks through the limitations of relying on a single round of loss to evaluate data quality, creatively introducing a fast loss mean calculated iteratively with each training round. This technique smooths loss fluctuations during training, enabling more stable and realistic quantification of the learning difficulty and anomaly degree of each chest CT image, thereby accurately eliminating "dirty data" caused by labeling errors or extremely low image quality. This provides a high-purity, high-quality denoised training set for subsequent model training, ensuring the upper limit of the model from the data source. Attached Figure Description

[0021] To more clearly illustrate the technical solutions of this application, the accompanying drawings used in the embodiments will be briefly described below. It should be understood that the following drawings only show some embodiments of this application and therefore should not be considered as a limitation on the scope of protection of this application. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 This illustrates a method for obtaining an optimized training set of chest CT images according to an embodiment of the present invention. Figure 1 ; Figure 2 A schematic diagram of chest CT images removed from the training set according to an embodiment of the present invention is shown. Figure 1 ; Figure 3 A schematic diagram of chest CT images removed from the training set according to an embodiment of the present invention is shown. Figure 2 ; Figure 4 The flowchart of the method for calculating the evaluation score of chest CT images according to an embodiment of the present invention is shown. Figure 1 ; Figure 5 The flowchart of the method for calculating the evaluation score of chest CT images according to an embodiment of the present invention is shown. Figure 2 ; Figure 6 A schematic diagram of the structure of a device for acquiring an optimized training set of chest CT images according to an embodiment of the present invention is shown; Figure 7 A flowchart of a CT image nodule region recognition model training method according to an embodiment of the present invention is shown; Figure 8 A schematic diagram of the structure of a CT image nodule region recognition model training device according to an embodiment of the present invention is shown; Figure 9 A flowchart of a method for identifying nodule regions in CT images according to an embodiment of the present invention is shown; Figure 10 A schematic diagram of the structure of a nodule region identification device for CT images according to an embodiment of the present invention is shown; Figure 11 A schematic diagram of the structure of a nodule region recognition system for CT images according to an embodiment of the present invention is shown; Figure 12 A schematic diagram of a computer device structure according to an embodiment of the present invention is shown. Detailed Implementation

[0023] The technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments.

[0024] The components of the embodiments of this application described and illustrated in the accompanying drawings can be arranged and designed in a variety of different configurations. Therefore, the following detailed description of the embodiments of this application provided in the drawings is not intended to limit the scope of the claimed application, but merely to illustrate selected embodiments of the application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0025] In the following, the terms “comprising,” “having,” and their cognates, which may be used in various embodiments of this application, are intended only to indicate a particular feature, number, step, operation, element, component, or combination thereof, and should not be construed as excluding, firstly, the presence of one or more other features, numbers, steps, operations, elements, components, or combinations thereof, or adding the possibility of one or more features, numbers, steps, operations, elements, components, or combinations thereof.

[0026] Furthermore, the terms "first," "second," and "third" are used only to distinguish descriptions and should not be interpreted as indicating or implying relative importance.

[0027] Unless otherwise specified, all terms used herein (including technical and scientific terms) shall have the same meaning as commonly understood by one of ordinary skill in the art to which the various embodiments of this application pertain. Terms (such as those defined in commonly used dictionaries) shall be interpreted as having the same meaning as in their contextual meaning in the relevant technical field and shall not be construed as having an idealized or overly formal meaning, unless clearly defined in the various embodiments of this application.

[0028] In existing technologies, when training a deep learning model to identify nodule regions in CT images, all collected labeled chest CT image data are typically mixed and directly used for iterative training of the neural network. However, due to issues such as missing labels, mislabeling, and inconsistent labeling scales in chest CT images, the accuracy of the trained deep learning model is low. This application addresses this by first training N rounds using all chest CT images in an initial training set, obtaining the loss corresponding to each chest CT image after the Nth round of training, and then sorting the chest CT images according to the magnitude of the loss. Several chest CT images with large losses are then removed, thereby eliminating noise samples (i.e., inaccurately or non-standardly labeled chest CT images) from the initial training set.

[0029] The following detailed description of some embodiments of this application is provided in conjunction with the accompanying drawings. Unless otherwise specified, the following embodiments and features can be combined with each other.

[0030] Example of a method for obtaining an optimized training set of chest CT images: In this embodiment, the chest CT image can be either a planar CT image or a 3D CT image. If the chest CT image is a planar CT image, the generated sub-region image is a planar image, the bounding box or detection box is a planar bounding box, and the intersection-union ratio is a two-dimensional intersection-union ratio. If the chest CT image is a 3D CT image, the generated sub-region image is a 3D image, the bounding box or detection box is a 3D bounding box, and the intersection-union ratio is a three-dimensional intersection-union ratio.

[0031] like Figure 1 As shown, the method for obtaining the optimized training set of chest CT images in this embodiment includes: S101. Obtain the initial training set, specifically: obtain the initial training set, which includes multiple chest CT images and the labels corresponding to each chest CT image; In this embodiment, the label is used to characterize whether there are pulmonary nodules in each chest CT image and to mark the corresponding pulmonary nodule region; In this embodiment, the lung nodule areas in the chest CT images are marked with a bounding box.

[0032] S102. Generate multiple sub-region images for each of the chest CT images in the initial training set; In this embodiment, for a given chest CT image, multiple sub-region images can be randomly generated. Alternatively, the image can be divided into several sub-region images at a fixed size. For example, if the chest CT image is a planar CT image, it can be divided into several sub-region images at a size of 256×256, with overlapping or non-overlapping portions between adjacent sub-region images. If the chest CT image is a 3D CT image, it can be divided into several sub-region images at a size of 160×160×80, with overlapping or non-overlapping portions between adjacent sub-region images.

[0033] A case identifier can be created for each chest CT image, and the data source identifier of the chest CT image can be recorded at the same time.

[0034] S103. Train the neural network model for N rounds, specifically by using the multiple sub-region images of each of the chest CT images to train the neural network model for N rounds, and obtaining the recognition result of each sub-region image of each of the chest CT images, wherein the recognition result includes a nodule region detection box. In this embodiment, N can be 25. In other embodiments, N can also be any other suitable value, such as any positive integer greater than 25.

[0035] In a single training round, after inputting the sub-region image of the chest CT image into the neural network model, the neural network model will output the corresponding recognition result. The recognition result is divided into two types: one is that there is a nodule region in the sub-region image, and the other is that there is no nodule region in the sub-region image. If the recognition result is that there is a nodule region, the neural network model will output the corresponding nodule region detection box.

[0036] In this embodiment, the neural network model uses the RetinaNet model. In other embodiments, other suitable neural network models can also be used, such as ResNet, 3D ResNet, ResNet-FPN, 3D ResNet-FPN, Faster R-CNN, RetinaNet, YOLO, SSD, U-Net, ResUNet, Mask R-CNN, etc.

[0037] In this embodiment, the training process corresponding to the j-th training round is as follows: Step 1: Batch partitioning of the entire dataset, including: At the beginning of the j-th training round, all sub-region images in the current training set (containing multiple chest CT images) are divided into M training batches according to a preset batch size. The neural network model will sequentially read these M training batches to perform forward and backward propagation, at which time the model inherits the initial weights updated in the previous round.

[0038] Step 2: Perform forward prediction and effective detection box selection for a single batch, including: For the currently executed m-th training batch (where m is less than or equal to M and greater than or equal to 1), input multiple sub-region images within this batch into the neural network model, and output the recognition results of each sub-region image within this batch (including predicted nodule region detection boxes and classification confidence). For each chest CT image within the m-th training batch, calculate the image-level loss based on the recognition results of each sub-region image.

[0039] Step 3: Update model parameters based on single-batch loss, including: calculating the overall network loss for the m-th batch based on the image-level loss of each chest CT image within the m-th batch; and performing backpropagation in the neural network model using gradient descent based on this overall network loss to immediately update the model's weight parameters. The updated model parameters will be directly used as the initial network parameters for forward prediction in the next batch (i.e., the (m+1)-th batch).

[0040] In one embodiment, the overall network loss of the m-th batch can be calculated by summing the image-level losses of each chest CT image to obtain a first sum, and then dividing the first sum by the batch size (i.e., averaging the image-level losses of each chest CT image) to obtain the overall network loss of the m-th batch. In other embodiments, other methods can also be used to calculate the overall network loss of the m-th batch.

[0041] Step 4: Repeat steps 2 to 3 until all M training batches in the j-th round have been completed. At this point, the neural network model has undergone M parameter iterations and updates in the j-th round.

[0042] S104. For each chest CT image, select valid detection frames from the corresponding nodule region detection frames; In this embodiment, the method for selecting effective detection boxes from the corresponding nodule region detection boxes for each chest CT image is as follows: for each chest CT image, all corresponding nodule region detection boxes are divided into positive sample detection boxes and negative sample detection boxes. Valid negative sample detection boxes are selected from all negative sample detection boxes by random sampling or positive-negative sample ratio sampling. All corresponding positive sample detection boxes and the selected valid negative sample detection boxes are taken as valid detection boxes. In this embodiment, a fixed IoU threshold matching method can be used to divide all nodule region detection boxes corresponding to the chest CT image into positive sample detection boxes and negative sample detection boxes, or an ATSS adaptive matching strategy can be used to divide all nodule region detection boxes corresponding to the chest CT image into positive sample detection boxes and negative sample detection boxes.

[0043] The method of using fixed IoU threshold matching to divide all nodule regions in chest CT images into positive and negative sample detection boxes includes: calculating the intersection-union ratio (i.e., the degree of overlap) between the nodule region detection box and the labeled box; if the intersection-union ratio is greater than the preset intersection-union ratio threshold, it is classified as a positive sample detection box; otherwise, it is classified as a negative sample detection box.

[0044] Selecting valid negative test boxes from all negative test boxes using random sampling means randomly selecting a predetermined proportion of negative test boxes from all negative test boxes, and using these selected negative test boxes as valid negative test boxes. The predetermined proportion refers to the ratio of the number of valid negative test boxes to the number of positive test boxes.

[0045] Using a positive-negative sample ratio sampling method to select effective negative sample detection frames from all negative sample detection frames means determining the number of negative samples to be selected based on the positive-negative sample ratio and the number of positive samples.

[0046] S105. Calculate the average image-level loss and fast loss for each chest CT image, specifically: for each chest CT image, calculate the average image-level loss and fast loss for each round, wherein the image-level loss in the j-th round is the weighted average of the losses of each sub-region of the chest CT image in the j-th round, and the sub-region loss is the weighted average of the classification losses of each valid detection box in the sub-region, 1≤j≤N; the average fast loss in the k-th round is positively correlated with the average fast loss in the (k-1)-th round and the image-level loss in the k-th round, 2≤k≤N; the average fast loss in the 1st round is equal to the image-level loss in the 1st round. In one embodiment, for each of the said chest CT images, the image-level loss corresponding to each round is calculated, including the following steps: S1051. Calculate the loss of each sub-region in the j-th round. Specifically, for the p-th sub-region in the j-th round, obtain each valid detection box participating in classification supervision within that sub-region, and calculate the classification loss of each valid detection box. In this embodiment, the classification loss can be calculated based on the cross-entropy loss function (e.g., BCEWithLogitsLoss) or other loss functions (e.g., Focal Loss, Gradient Harmonic Classification Loss (GHM-C), or Average Precision Loss (AP-Loss), etc.), according to the true label and model prediction value of each valid detection box. Subsequently, based on the weights corresponding to each valid detection box, calculate the weighted average of the classification losses of each valid detection box to obtain the sub-region loss of that sub-region.

[0047] The weights corresponding to the classification loss of each valid detection box can be set according to actual needs. In a preferred embodiment, the classification loss weights of each valid detection box can be made equal, that is, the arithmetic mean of the classification losses of all valid detection boxes in the sub-region can be used as the sub-region loss of the sub-region.

[0048] in, , , For the total number of training rounds, The total number of sub-regions divided into chest CT images.

[0049] S1052. Calculate the image-level loss for the j-th round. Specifically, when calculating the weighted average of the losses of each sub-region of the chest CT image in the j-th round, it is necessary to first determine the weight of the loss for each sub-region. In this embodiment, the weight of the loss for each sub-region can be determined based on the number of valid detection boxes in each sub-region, and the more valid detection boxes in each sub-region, the greater the corresponding weight. Multiply each sub-region loss by its corresponding weight and sum them, then divide by the sum of all weights to obtain the image-level loss for the j-th round.

[0050] In this embodiment, the expression for calculating image-level loss can be:

[0051] In the formula, This represents the image-level loss of the chest CT image corresponding to the j-th round; This represents the sub-region loss of the p-th sub-region in the j-th round; The weight represents the loss of the p-th sub-region, and its value is the number of valid detection boxes in the p-th sub-region (i.e. the number of valid elements that actually enter the classification supervision).

[0052] By using the weighted averaging method based on the number of effective detection boxes, it is possible to ensure that sub-regions containing more effective classification elements have a higher weight in the image-level loss calculation, thereby avoiding the shift in supervision signals caused by the cropping method of each sub-region or the uneven number of positive and negative samples.

[0053] In this embodiment, the fast loss mean is a rapidly updated exponential moving average of the image-level loss, used to reflect whether the corresponding chest CT image still maintains a high loss at the current stage (i.e., the current round). If a chest CT image has been consistently poorly learned recently, then the value of the fast loss mean of this chest CT image will be relatively high, and the probability that this chest CT image is incorrectly labeled is also higher.

[0054] In one embodiment, for each of the chest CT images, the mean rapid loss for each round is calculated, including: calculating the first product of the mean rapid loss for the (k-1)th round and a first preset multiple; calculating the second product of the image-level loss for the kth round and a second preset multiple; and determining the sum of the first product and the second product as the mean rapid loss for the kth round, where 2≤k≤N.

[0055] In this embodiment, the sum of the first preset multiple and the second preset multiple is 1, and the expression for calculating the mean of the fast loss is: ; In the formula, and Let represent the mean of the rapid loss for the i-th chest CT image corresponding to the k-th and (k-1)-th rounds, respectively. This represents the image-level loss of the i-th chest CT image corresponding to the k-th round; The first preset multiple, The first attenuation coefficient (i.e., the second preset multiple) can be 0.1, or other suitable values ​​in other embodiments.

[0056] When calculating the fast loss mean, this invention uses a weighted sum of the fast loss mean from the previous round and the image-level loss from the current round for iterative updates. This technique utilizes the concept of a moving average, requiring only the state data from the previous round during operation, without needing to reserve a large amount of memory resources to store all historical loss data. This scheme significantly reduces the space complexity and computational overhead in the data cleaning process, while effectively preserving the temporal evolution characteristics of historical losses, achieving a technical effect that balances computational efficiency and evaluation continuity.

[0057] S106. Calculate the evaluation score of each chest CT image, specifically: calculate the evaluation score of each chest CT image based on the mean of the rapid loss of each chest CT image corresponding to the Nth round, wherein the evaluation score is positively correlated with the mean of the corresponding rapid loss.

[0058] In this embodiment, the evaluation score is used to characterize the likelihood that a chest CT image belongs to a high-risk sample with significant annotation noise.

[0059] Since the higher the mean value of the rapid loss of a chest CT image, the higher the possibility that the annotation of the chest CT image is problematic, making the evaluation score positively correlated with the corresponding mean value of the rapid loss can more accurately calculate the probability that the chest CT image belongs to a high-risk sample with large annotation noise.

[0060] It should be noted that in this embodiment, the evaluation score can be calculated based solely on the mean of rapid loss of the chest CT image, or it can be calculated based on the mean of rapid loss combined with other parameters.

[0061] S107. Obtain the optimized training set, specifically: based on the evaluation score ranking results of each chest CT image, select a preset number of target chest CT images from the multiple chest CT images, remove the target chest CT images from the initial training set, and use the remaining chest CT images as the optimized training set.

[0062] In this embodiment, selecting a preset number of target chest CT images from multiple chest CT images based on the evaluation scores of each chest CT image includes: sorting each chest CT image in descending order of evaluation scores, and sequentially selecting a preset number of target chest CT images in ascending order of serial number.

[0063] In this embodiment, the preset number can be 5% of the total number of chest CT images in the initial training set. In other embodiments, the preset number can also be other suitable values.

[0064] The target chest CT images removed from the initial training set (i.e., high-risk samples with significant annotation noise) are as follows: Figure 2 and Figure 3 As shown.

[0065] It should be noted that steps S102 to S107 of the method for obtaining the chest CT image optimization training set in this embodiment can be executed iteratively multiple times. Each time it is executed, a portion of samples with high annotation noise will be removed from the training set, and the obtained optimized training set will be used as a new initial training set. Steps S102 to S107 will be executed for the chest CT images in the new initial training set. By iteratively executing multiple times, the quality of the samples in the optimized training set can be improved.

[0066] It should also be noted that the method for obtaining the optimized training set of chest CT images in this embodiment can be used not only to denoise the training set composed of chest CT images, but also to denoise the training set composed of other types of images (such as images taken in industrial scenes).

[0067] This embodiment provides a method for obtaining an optimized training set of chest CT images. By generating multiple sub-region images from the chest CT images and extracting nodule region detection boxes to filter out effective detection boxes, a refined processing logic from "global image" to "local effective features" is constructed. This technique effectively overcomes the dilution effect of large areas of non-lung tissue and healthy background regions in chest CT images on the model gradient, avoids noise interference caused by massive redundant backgrounds, and forces the model to focus more on the real nodule feature distribution, significantly improving the targeting of feature extraction and the utilization of computational resources. Simultaneously, this invention overcomes the limitations of relying on a single round of loss to evaluate data quality, creatively introducing a fast loss mean calculated iteratively with each training round. This technique smooths loss fluctuations during training, enabling more stable and realistic quantification of the learning difficulty and anomaly degree of each chest CT image, thereby accurately eliminating "dirty data" caused by labeling errors or extremely low image quality. This provides a high-purity, high-quality denoised training set for subsequent model training, ensuring the upper limit of the model from the data source.

[0068] As can be seen from the above embodiments, the evaluation score of a chest CT image can be calculated based on the fast loss mean combined with other parameters. In another embodiment, the slow loss mean, which characterizes the average loss level of the chest CT image over a longer time scale, can be further calculated. The evaluation score of the chest CT image is then calculated based on the fast loss mean and the slow loss mean. Figure 4 As shown, the specific process is as follows: S201. After the k-th round of training, calculate the average slow loss of each chest CT image. Specifically, after the k-th round of training, for each chest CT image, calculate the average slow loss of the (k-1)-th round and the third product of the third preset multiple, calculate the image-level loss of the k-th round and the fourth product of the fourth preset multiple, and determine the sum of the third product and the fourth product as the average slow loss of the k-th round; 2≤k≤N; the average slow loss corresponding to the first round of training is equal to the image-level loss of the corresponding chest CT image in the first round of training; the value of the third preset multiple is greater than the value of the first preset multiple. In this embodiment, the slow loss mean is a slower-updating exponential moving average of the image-level loss. It is smoother than the fast mean and less susceptible to short-term fluctuations.

[0069] In one implementation, the expression for calculating the mean of the slow loss is: ; In the formula, and Let represent the mean slow loss of the i-th chest CT image corresponding to the k-th and (k-1)-th rounds, respectively. This represents the image-level loss of the i-th chest CT image corresponding to the k-th round; The third preset multiple, This is the second attenuation coefficient (i.e., the fourth preset multiple). The value can be 0.01, and in other implementations, it can also be other suitable values.

[0070] S202. For each of the chest CT images, calculate the reduction ratio based on the difference between the mean slow loss and the mean fast loss in the kth round, wherein the reduction ratio is positively correlated with the difference. In this embodiment, the descent ratio is used to compare the relationship between the slow loss mean and the fast loss mean, to determine whether the loss of the chest CT image has decreased significantly. A larger descent ratio indicates that the current fast mean is lower than the long-term mean, meaning the model has learned better recently than before. Conversely, if this ratio is small, or even close to 0, it indicates that the neural network model's ability to recognize the chest CT image has not improved significantly, and the descent is insufficient. Therefore, a smaller descent ratio suggests a higher probability that the corresponding chest CT image belongs to a high-risk sample with significant annotation noise.

[0071] In this embodiment, the expression for calculating the decrease ratio is: ; In the formula, This represents the decrease rate of the i-th chest CT image corresponding to the k-th round. and Let represent the mean slow loss of the i-th chest CT image corresponding to the k-th round and the mean fast loss of the i-th chest CT image corresponding to the k-th round, respectively. It is a constant, and its value can be... .

[0072] S203. Based on the mean rapid loss and the decrease ratio of each chest CT image corresponding to the Nth round, calculate the evaluation score of each chest CT image. The evaluation score is positively correlated with the corresponding mean rapid loss and negatively correlated with the corresponding decrease ratio.

[0073] As can be seen from the above, the smaller the decrease ratio, the greater the probability that the corresponding chest CT image belongs to a high-risk sample with large annotation noise; the larger the mean of rapid loss, the greater the probability that the corresponding chest CT image belongs to a high-risk sample with large annotation noise; therefore, the evaluation score is positively correlated with the corresponding mean of rapid loss and negatively correlated with the corresponding decrease ratio, and the evaluation score of the chest CT image can be calculated relatively accurately.

[0074] This embodiment introduces a "slow loss mean" with higher weighting for historical features, and calculates the "decline ratio" by the difference between it and the fast loss mean, thus constructing a multi-dimensional evaluation system. During model training, both "difficult samples" with complex features and "dirty samples" with mislabeled labels often exhibit high loss values, easily leading to accidental deletion. This invention, by setting a mechanism where the evaluation score is positively correlated with the fast loss mean and negatively correlated with the decline ratio, can accurately capture the model's learning convergence trend. Specifically, for images with high loss but a significant decline ratio (i.e., difficult samples that the model is striving to converge), the system retains them to improve model robustness; for images with high loss and no decline trend (i.e., mislabeled data that the model cannot learn), the system resolutely removes them. This scheme completely solves the technical defect of conventional threshold filtering methods that easily and accidentally delete high-value difficult samples.

[0075] In addition to calculating the evaluation score of chest CT images based on the decline rate and the mean of rapid loss, in another embodiment, such as Figure 5 As shown, in calculating the evaluation score, in addition to the descent rate and the mean of rapid loss, the jitter score of the chest CT image can also be calculated. The evaluation score is then calculated based on the descent rate, the mean of rapid loss, and the jitter score. The specific process includes: S301. Calculate the jitter score of each chest CT image in the k-th round, 2≤k≤N. Specifically, for each chest CT image, calculate the loss deviation between the image-level loss in the k-th round and the image-level loss in the (k-1)-th round, calculate the fifth product of the loss deviation and a fifth preset multiple, calculate the jitter score in the (k-1)-th round and a sixth preset multiple, and determine the sum of the fifth and sixth product values ​​as the jitter score in the k-th round; the jitter score corresponding to the first round is 0; 2≤k≤N; In this embodiment, the expression for calculating the jitter score is: ; In the formula, and Let represent the jitter score of the i-th chest CT image corresponding to the k-th round and the jitter score of the i-th chest CT image corresponding to the (k-1)-th round, respectively. and Let represent the image-level loss of the i-th chest CT image corresponding to the k-th round and the image-level loss of the i-th chest CT image corresponding to the (k-1)-th round, respectively. It is the sixth preset multiple. This indicates the fifth preset multiple. In this embodiment, The value can be 0.1, and other suitable values ​​can be used in other embodiments.

[0076] S302. Calculate the evaluation score of each chest CT image based on the mean rapid loss, the decrease ratio, and the jitter score of each chest CT image corresponding to the Nth round. Specifically, calculate the evaluation score of each chest CT image based on the mean rapid loss, the decrease ratio, and the jitter score of each chest CT image corresponding to the Nth round. The evaluation score is positively correlated with the corresponding mean rapid loss and the corresponding jitter score, and negatively correlated with the corresponding decrease ratio.

[0077] Generally, a higher jitter score indicates greater fluctuation in the loss of the corresponding chest CT image, more unstable training, and a higher probability that the corresponding chest CT image belongs to a high-risk sample with significant annotation noise. Therefore, the evaluation score should be positively correlated with the jitter score when calculating the evaluation score.

[0078] In one embodiment, the evaluation score is calculated as follows: First, a first difference between 1 and the decrease ratio is calculated; then, the mean of rapid loss, the jitter score, and the first difference of each chest CT image corresponding to the Nth round are added together to obtain the evaluation score of each chest CT image.

[0079] It should be noted that in other embodiments, the mean of rapid loss, jitter score and first difference of each chest CT image can be weighted and summed to obtain the evaluation score of each chest CT image.

[0080] This embodiment constructs a "jitter score" by calculating the absolute deviation of the image-level loss between adjacent rounds and introduces it as a positive correlation factor into the evaluation system. Since mislabeled or strongly interfering data often cause gradient direction conflicts during backpropagation, the loss value of the sample oscillates dramatically across multiple training rounds. This invention quantifies and extracts this dynamic oscillation feature, further enhancing the algorithm's sensitivity to abnormally poor labels, effectively filling the gap in evaluation based solely on absolute loss values, and significantly reducing the false negative rate (missed deletion rate) during data cleaning.

[0081] Example of a device for acquiring an optimized training set of chest CT images: like Figure 6 As shown, this embodiment provides a device 100 for acquiring an optimized training set of chest CT images, including: The initial training set acquisition module 110 is used to acquire an initial training set, which includes multiple chest CT images and labels corresponding to each chest CT image. The labels are used to characterize whether there are lung nodules in each chest CT image and to mark the corresponding lung nodule regions. The sub-region image generation module 120 is used to generate multiple sub-region images for each of the chest CT images in the initial training set. The nodule region detection box generation module 130 is used to train the neural network model for N rounds using the multiple sub-region images of each chest CT image to obtain the recognition result of each sub-region image of each chest CT image, and the recognition result includes the nodule region detection box. The effective detection box acquisition module 140 is used to filter out effective detection boxes from the corresponding nodule region detection boxes for each chest CT image. The loss calculation module 150 is used to calculate the image-level loss and the mean fast loss for each round for each chest CT image. The image-level loss in the j-th round is the weighted average of the losses of each sub-region of the chest CT image in the j-th round, and the sub-region loss is the weighted average of the classification losses of each valid detection box in the sub-region, 1≤j≤N. The mean fast loss in the k-th round is positively correlated with the mean fast loss in the (k-1)-th round and the image-level loss in the k-th round, 2≤k≤N. The mean fast loss in the 1st round is equal to the image-level loss in the 1st round. The evaluation score calculation module 160 is used to calculate the evaluation score of each chest CT image based on the mean of the rapid loss of each chest CT image corresponding to the Nth round, wherein the evaluation score is positively correlated with the mean of the corresponding rapid loss. The optimized training set acquisition module 170 is used to select a preset number of target chest CT images from multiple chest CT images based on the evaluation score ranking results of each chest CT image, remove the target chest CT images from the initial training set, and use the remaining chest CT images as the optimized training set.

[0082] This embodiment provides a device for acquiring an optimized training set of chest CT images. Through the collaborative work of various functional modules, this device constructs an end-to-end, automated medical image label cleaning system. By dynamically capturing and analyzing the loss characteristics of samples during the initial model training phase, the device can accurately identify and remove mislabeled, missing, or extremely low-quality "dirty data" from the dataset, thereby improving the overall quality of the dataset from the source. Specifically, this device has the following significant technical effects: It is understood that the device for acquiring the optimized training set of chest CT images in this embodiment corresponds to the method for acquiring the optimized training set of chest CT images in the above embodiment. The options in the above embodiment are also applicable to this embodiment, so they will not be described again here.

[0083] Example of a training method for a CT image nodule region recognition model: like Figure 7 As shown, this embodiment provides a method for training a CT image nodule region recognition model, including: S401. Obtain the optimized training set, specifically by using the CT image optimized training set acquisition method described in the above embodiments to obtain the optimized training set; S402. The neural network model is further trained using an optimized training set to obtain a CT image nodule region recognition model.

[0084] When further training the neural network model, the training method described in the above-mentioned method embodiment for obtaining the optimized training set of chest CT images can be used, or other suitable training methods can be used.

[0085] On a 4mm–30mm lung nodule test set, the detection performance of two CT image nodule region recognition models was compared. One was the baseline model, which was trained directly using multiple chest 3D CT images as the training set via an incremental learning scheme without training set optimization. The other was the model used in this embodiment, which used the baseline model's training set as the initial training set, obtained an optimized training set through step S401, and then further trained the neural network model using step S402. After adopting the method of this embodiment, as shown in Table 1, the model recall increased from 0.8898 to 0.9526, an improvement of 6.28 percentage points; precision increased from 0.5474 to 0.6113, an improvement of 6.39 percentage points; and the F1 score (the harmonic mean of precision and recall) increased from 0.6778 to 0.7447, an improvement of 6.69 percentage points.

[0086] Table 1

[0087] The above results show that the present invention can effectively reduce the interference of abnormal annotations, missing annotations and unstable samples on the training process by identifying and removing high-risk noise samples through image-level long-term loss memory, and significantly improve the overall performance of the CT image nodule region recognition model.

[0088] This embodiment utilizes the high-purity "optimized training set" obtained by the chest CT image optimization training set acquisition method described in the above embodiments to further train the neural network model. This completely blocks the misleading influence of incorrect labels or poor-quality images in the initial dataset on the model parameter updates, enabling the neural network to converge more accurately and purely to the morphological feature distribution of the real lung nodules, avoiding feature shift, and significantly improving the robustness and feature representation ability of the final output nodule region recognition model.

[0089] Example of a training device for CT image nodule region recognition model: like Figure 8 As shown, this embodiment provides a CT image nodule region recognition model training device 300, including: The optimized training set acquisition module 310 is used to acquire the optimized training set using the CT image optimized training set acquisition method described in the above embodiments; The training module 320 is used to further train the neural network model using the optimized training set to obtain a CT image nodule region recognition model.

[0090] By optimizing the deep integration between the training set acquisition module and the training module, a closed-loop system of "high-quality data input - efficient model output" is formed. This hardware / software architecture ensures that the model receives high-quality data throughout its training lifecycle, guaranteeing the high accuracy and stability of the final deployed model.

[0091] It is understood that the CT image nodule region recognition model training device of this embodiment corresponds to the CT image nodule region recognition model training method of the above embodiment. The options in the above embodiment are also applicable to this embodiment, so they will not be described again here.

[0092] Example of a method for identifying nodule regions in CT images: like Figure 9 As shown, this embodiment provides a method for identifying nodule regions in CT images, including: S501. Acquire the chest CT image to be identified; S502. Input the chest CT image to be identified into the CT image nodule region recognition model to obtain the nodule region recognition result. The CT image nodule region recognition model is pre-trained using the CT image nodule region recognition model training method in the above embodiments.

[0093] The method employed in this embodiment significantly improves the clinical generalization performance and accuracy of the final nodule identification model: After obtaining a high-quality optimized training set, this invention utilizes it to further train the neural network model. By eliminating the misleading influence of incorrect labels, the model can fully and correctly converge to the true distribution of lung nodule features, completely eliminating the risk of "overfitting dirty data" in the later stages of training. The resulting CT image nodule region identification model possesses stronger feature generalization and discrimination capabilities when faced with real clinical CT images to be identified, enabling more accurate selection of true lung nodule regions and effectively reducing the false positive rate (misdiagnosis) and false negative rate (missed diagnosis) in clinical auxiliary diagnostic scenarios.

[0094] It can be understood that the method of this embodiment corresponds to the nodule region recognition device for CT images in the above embodiment, and can realize the functions of each module in the nodule region recognition device for CT images and achieve the corresponding technical effects.

[0095] Example of a device for identifying nodule regions in CT images: like Figure 10 As shown, this embodiment provides a nodule region identification device 200 for CT images, including: The chest CT image acquisition module 210 is configured to acquire the chest CT image to be identified. The nodule region recognition result generation module 220 is configured to input the chest CT image to be recognized into the CT image nodule region recognition model to obtain the nodule region recognition result. The CT image nodule region recognition model is pre-trained using the CT image nodule region recognition model training method in the above embodiments.

[0096] This device integrates a module for acquiring chest CT images to be identified and a module for generating nodule region identification results, deeply fusing a high-precision nodule identification core algorithm (included in the computer instructions of the processing device) with clinical physical equipment. It achieves seamless integration from patient CT image scanning and acquisition, low-level intelligent calculation, to front-end nodule annotation result output. In a real hospital setting, this physical device can provide radiologists with high-confidence targeted nodule indications in real time, greatly reducing doctors' image interpretation time and visual fatigue.

[0097] Example of a nodule region recognition system for CT images: like Figure 11 As shown, this embodiment provides a nodule region recognition system 400 for CT images, including: Image acquisition device 410 is used to acquire chest CT images to be identified; The processing device 420 is communicatively connected to the image acquisition device. The processing device stores computer instructions, which, when executed, implement the nodule region recognition method for CT images described in the above embodiments to obtain nodule region recognition results. Output device 430 is used to output the nodule region identification result.

[0098] Computer equipment example: like Figure 12 As shown, this application also provides a computer device, which, exemplary, includes a processor, a memory, and a communication interface. The memory stores a computer program, and the processor executes the computer program to enable the computer device to perform the methods described in the above embodiments for obtaining a chest CT image optimized training set, training a CT image nodule region recognition model, or identifying a nodule region in CT images.

[0099] The processor can be an integrated circuit chip with signal processing capabilities. The processor can be a general-purpose processor, including at least one of a Central Processing Unit (CPU), Graphics Processing Unit (GPU), Network Processor (NP), Digital Signal Processor (DSP), Application-Specific Integrated Circuit (ASIC), Field-Programmable Gate Array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. The general-purpose processor can be a microprocessor or any conventional processor, capable of implementing or executing the methods, steps, and logic block diagrams disclosed in the embodiments of this application.

[0100] The memory can be, but is not limited to, Random Access Memory (RAM), Read Only Memory (ROM), Programmable Read-Only Memory (PROM), Erasable Programmable Read-Only Memory (EPROM), Electrically Erasable Programmable Read-Only Memory (EEPROM), etc. The memory is used to store computer programs, and the processor can execute the computer programs accordingly after receiving execution instructions.

[0101] Examples of computer storage media: This application also provides a computer storage medium for storing the computer program used in the aforementioned computer device. The computer storage medium can be a readable storage medium, a non-volatile storage medium, or a volatile storage medium. For example, the computer storage medium may include, but is not limited to, various media capable of storing program code, such as a USB flash drive, a portable hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0102] Example of a computer program product: This embodiment provides a computer program product, including a computer-readable storage medium having program instructions contained therein, the program instructions being executable by a processor to implement the method for obtaining an optimized training set of chest CT images, the method for training a CT image nodule region recognition model, or the method for recognizing nodule regions of CT images as described in the above embodiments.

[0103] In the several embodiments provided in this application, it should be understood that the disclosed apparatus and methods can also be implemented in other ways. The apparatus embodiments described above are merely illustrative. For example, the flowcharts and block diagrams in the accompanying drawings show the architecture, functionality, and operation of possible implementations of apparatus, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that, in alternative implementations, the functions marked in the blocks may occur in a different order than those marked in the drawings. For example, two consecutive blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagram and / or flowchart, and combinations of blocks in the block diagram and / or flowchart, can be implemented using a dedicated hardware-based system that performs the specified function or action, or using a combination of dedicated hardware and computer instructions.

[0104] In addition, the functional modules or units in the various embodiments of this application can be integrated together to form an independent part, or each module can exist independently, or two or more modules can be integrated to form an independent part.

[0105] If the aforementioned functions are implemented as software functional modules and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a smartphone, personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application.

[0106] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A method for obtaining an optimized training set of chest CT images, characterized in that, include: Obtain an initial training set, which includes multiple chest CT images and labels corresponding to each chest CT image; Multiple sub-region images are generated for each of the chest CT images in the initial training set; The neural network model is trained for N rounds using the multiple sub-region images of each of the chest CT images to obtain the recognition result of each sub-region image of each chest CT image, and the recognition result includes a nodule region detection box; For each of the chest CT images, valid detection frames are selected from the corresponding nodule region detection frames; For each chest CT image, the mean of image-level loss and fast loss for each round is calculated. The image-level loss for round j is the weighted average of the losses of each sub-region of the chest CT image in round j, where the sub-region loss is the weighted average of the classification losses of each valid detection box in the sub-region, 1 ≤ j ≤ N. The mean of fast loss for round k is positively correlated with the mean of fast loss for round (k-1) and the image-level loss for round k, 2 ≤ k ≤ N. The mean of fast loss for round 1 is equal to the image-level loss for round 1. Based on the mean rapid loss of each chest CT image corresponding to the Nth round, an evaluation score is calculated for each chest CT image, and the evaluation score is positively correlated with the corresponding mean rapid loss. Based on the evaluation scores of each chest CT image, a predetermined number of target chest CT images are selected from the multiple chest CT images. The target chest CT images are removed from the initial training set, and the remaining chest CT images are used as the optimization training set.

2. The method for obtaining an optimized training set of chest CT images as described in claim 1, characterized in that, The step of selecting valid detection frames from the corresponding nodule region detection frames for each chest CT image includes: For each chest CT image, the corresponding nodule region detection boxes are divided into positive sample detection boxes and negative sample detection boxes. Valid negative sample detection boxes are selected from all negative sample detection boxes by random sampling or positive-negative sample ratio sampling. All the corresponding positive sample detection boxes and the selected valid negative sample detection boxes are taken as valid detection boxes.

3. The method for obtaining an optimized training set of chest CT images as described in claim 1, characterized in that, For the chest CT images, the mean of the fast loss for each round is calculated, including: Calculate the first product of the fast loss mean and the first preset multiple corresponding to the (k-1)th round, calculate the second product of the image-level loss and the second preset multiple corresponding to the kth round, and determine the sum of the first product and the second product as the fast loss mean of the kth round, where 2≤k≤N.

4. The method for obtaining an optimized training set of chest CT images as described in claim 3, characterized in that, The calculation of the evaluation score for each chest CT image based on the mean of the fast loss of each chest CT image corresponding to the Nth round includes: After the k-th round of training, for each chest CT image, calculate the average slow loss of the (k-1)-th round and the third product of the third preset multiple, calculate the image-level loss of the k-th round and the fourth product of the fourth preset multiple, and determine the sum of the third product and the fourth product as the average slow loss of the k-th round, where 2≤k≤N; the average slow loss corresponding to the first round of training is equal to the image-level loss of the corresponding chest CT image in the first round of training; the value of the third preset multiple is greater than the value of the first preset multiple; For each of the chest CT images, the reduction ratio is calculated based on the difference between the mean slow loss and the mean fast loss in the kth round, and the reduction ratio is positively correlated with the difference. Based on the mean rapid loss and the decrease ratio of each chest CT image corresponding to the Nth round, the evaluation score of each chest CT image is calculated. The evaluation score is positively correlated with the corresponding mean rapid loss and negatively correlated with the corresponding decrease ratio.

5. The method for obtaining an optimized training set of chest CT images as described in claim 4, characterized in that, The calculation of the evaluation score for each chest CT image based on the mean rapid loss and the decrease ratio of each chest CT image corresponding to the Nth round includes: For each of the chest CT images, the loss deviation between the image-level loss in the k-th round and the image-level loss in the (k-1)-th round is calculated. The fifth product of the loss deviation and a fifth preset multiple is calculated. The jitter score in the (k-1)-th round and a sixth preset multiple are calculated. The sum of the fifth and sixth product values ​​is determined as the jitter score in the k-th round. The jitter score corresponding to the first round is 0, and 2≤k≤N. The evaluation score of each chest CT image is calculated based on the mean rapid loss, the decrease ratio, and the jitter score of each chest CT image corresponding to the Nth round. The evaluation score is positively correlated with the corresponding mean rapid loss and the corresponding jitter score, and negatively correlated with the corresponding decrease ratio.

6. The method for obtaining an optimized training set of chest CT images as described in claim 5, characterized in that, The calculation of the evaluation score for each chest CT image based on the mean of rapid loss, the decrease ratio, and the jitter score corresponding to each chest CT image in the Nth round includes: Calculate the first difference between 1 and the stated decrease ratio; The mean of rapid loss, the jitter score, and the first difference of each chest CT image corresponding to the Nth round are added together to obtain the evaluation score of each chest CT image.

7. A method for training a CT image nodule region recognition model, characterized in that, include: The optimized training set is obtained using the method for obtaining the optimized training set of CT images as described in any one of claims 1 to 6; The neural network model is further trained using the optimized training set to obtain a CT image nodule region recognition model.

8. A method for identifying nodule regions in CT images, characterized in that, include: Acquire the chest CT image to be identified; The chest CT image to be identified is input into the CT image nodule region recognition model to obtain the nodule region recognition result, wherein the CT image nodule region recognition model is pre-trained using the method described in claim 7.

9. A device for acquiring an optimized training set of chest CT images, characterized in that, include: An initial training set acquisition module is used to acquire an initial training set, which includes multiple chest CT images and labels corresponding to each chest CT image. The sub-region image generation module is used to generate multiple sub-region images for each of the chest CT images in the initial training set. The nodule region detection box generation module is used to train the neural network model for N rounds using the multiple sub-region images of each chest CT image to obtain the recognition result of each sub-region image of each chest CT image, and the recognition result includes the nodule region detection box. The effective detection box acquisition module is used to filter out effective detection boxes from the corresponding nodule region detection boxes for each chest CT image. The loss calculation module is used to calculate the image-level loss and the mean fast loss for each round for each chest CT image. The image-level loss in round j is the weighted average of the losses of each sub-region of the chest CT image in round j, and the sub-region loss is the weighted average of the classification losses of each valid detection box in the sub-region, 1 ≤ j ≤ N. The mean fast loss in round k is positively correlated with the mean fast loss in round (k-1) and the image-level loss in round k, 2 ≤ k ≤ N. The mean fast loss in round 1 is equal to the image-level loss in round 1. The evaluation score calculation module is used to calculate the evaluation score of each chest CT image based on the mean of the rapid loss of each chest CT image corresponding to the Nth round, wherein the evaluation score is positively correlated with the mean of the corresponding rapid loss. The optimized training set acquisition module is used to select a preset number of target chest CT images from multiple chest CT images based on the evaluation score ranking results of each chest CT image, remove the target chest CT images from the initial training set, and use the remaining chest CT images as the optimized training set.

10. A training device for a CT image nodule region recognition model, characterized in that, include: An optimized training set acquisition module is used to acquire the optimized training set using the CT image optimized training set acquisition method according to any one of claims 1 to 6; The training module is used to further train the neural network model using the optimized training set to obtain a CT image nodule region recognition model.

11. A device for identifying nodule regions in CT images, characterized in that, include: The chest CT image acquisition module is configured to acquire the chest CT image to be identified. The nodule region recognition result generation module is configured to input the chest CT image to be recognized into the CT image nodule region recognition model to obtain the nodule region recognition result, wherein the CT image nodule region recognition model is pre-trained using the method as described in claim 7.

12. A nodule region recognition system for CT images, characterized in that, include: Image acquisition device, used to acquire chest CT images to be identified; The processing device is communicatively connected to the image acquisition device. The processing device stores computer instructions, which, when executed, implement the nodule region recognition method for CT images as described in claim 8, in order to obtain nodule region recognition results. An output device is used to output the nodule region identification results.

13. A computer device, characterized in that, The computer device includes a processor and a memory, the memory storing a computer program, and the processor executing the computer program to implement the method for obtaining the optimized training set of chest CT images according to any one of claims 1-6, or the method for training a CT image nodule region recognition model according to claim 7, or the method for recognizing nodule regions of CT images according to claim 8.

14. A computer-readable storage medium, characterized in that, It stores a computer program, which, when executed on a processor, implements the method for obtaining an optimized training set of chest CT images according to any one of claims 1-6, the method for training a CT image nodule region recognition model according to claim 7, or the method for recognizing nodule regions of CT images according to claim 8.

15. A computer program product comprising a computer-readable storage medium having program instructions contained therein, characterized in that, The program instructions can be executed by a processor to implement the method for obtaining the optimized training set of chest CT images according to any one of claims 1-6, the method for training the CT image nodule region recognition model according to claim 7, or the method for recognizing the nodule region of CT images according to claim 8.

Images