Training method of tumor lesion target detection model for endometrial cancer CT image and application thereof

Abstract

The application provides a training method and application of a tumor lesion target detection model for endometrial carcinoma CT images, and comprises the following steps: acquiring endometrial carcinoma CT images labeled with real lesion boundary boxes as training images; constructing a tumor lesion target detection model; grouping multiple training images into a training data set; iteratively training the tumor lesion target detection model using the training data set; constructing a total loss function based on a classification loss and a lesion-aware boundary box loss; and training the tumor lesion target detection model based on the total loss function until the tumor lesion target detection model converges. In the scheme, the risk weight is calculated based on the lesion area and the local contrast of the real lesion boundary box, the lesion-aware boundary box loss is obtained by weighting the boundary box regression loss using the risk weight, the positioning supervision intensity of high-miss-risk lesions such as small-scale and low-contrast lesions is oriented and strengthened, and the feature attention of the model to difficult lesions is improved.

Description

Technical Field

[0001] This application relates to the field of medical image processing, and in particular to a training method and application of a tumor lesion target detection model for CT images of endometrial cancer. Background Technology

[0002] Endometrial cancer is one of the most common malignant tumors of the female reproductive system. With the popularization of medical imaging technology, CT imaging, with its high spatial resolution, has become the core imaging tool for lesion localization, clinical staging assessment, and postoperative follow-up monitoring of endometrial cancer. However, in actual clinical scenarios, endometrial cancer lesions often appear on CT images as having blurred boundaries, little difference in grayscale from surrounding normal soft tissue, and lacking clear boundary markers. This not only increases the difficulty of manual image interpretation but also places higher demands on the recognition capabilities of automated detection algorithms. Traditional manual image interpretation relies heavily on the physician's clinical experience. When faced with large-scale continuous image slices, it is easily affected by image fatigue, experience differences, and subjective judgment biases. The stability of identifying small lesions, lesions with irregular boundaries, and low-contrast abnormal areas is insufficient, posing a potential risk of missed diagnosis and misjudgment.

[0003] With the development of deep learning technology, object detection algorithms represented by Faster R-CNN, RetinaNet, and the YOLO series have achieved mature applications in the field of natural images and are gradually migrating to the field of medical image analysis. Among them, the YOLO series models have the characteristics of end-to-end detection, fast inference speed, and convenient deployment, and are often used as the basic framework for lesion detection in medical images. However, medical images and natural images differ significantly in terms of data scale, object morphological features, boundary clarity, and annotation consistency. When general object detection models are directly applied, the models are more likely to learn anatomical structures and background textures with clear boundaries and significant features, but they do not pay enough attention to difficult lesion features with small scale and low contrast. Ultimately, this can easily lead to problems such as missed lesion detection, low confidence prediction, and boundary localization errors, making it difficult to meet the clinical needs of accurate detection of lesions on CT scans for endometrial cancer.

[0004] Current optimization schemes for lesion detection in medical imaging mostly focus on adjusting the backbone network structure, embedding attention mechanisms, or improving a single loss function, with few systematically addressing the issue of missed detection of difficult lesions from the perspective of a supervision mechanism throughout the entire training process. On the one hand, existing training strategies generally use a uniform supervision intensity for different lesion samples, making it difficult to provide targeted and enhanced localization supervision for small or low-contrast lesions. On the other hand, existing schemes lack a coherent optimization chain to address the problems of insufficient positive sample allocation and inadequate response of lesion region features for difficult lesions, and the intensity of auxiliary supervision is difficult to dynamically adapt with the training process, easily leading to insufficient guidance in the early stages of training or interference with the convergence of the main task in the later stages. Therefore, there is an urgent need for a target detection training scheme for endometrial cancer CT lesions with the ability to perceive the risk of missed detection, to systematically improve the detection performance of difficult lesions. Summary of the Invention

[0005] This application provides a training method and application for a tumor lesion target detection model for CT images of endometrial cancer. The scheme calculates risk weights based on the lesion area and local contrast of the real lesion bounding box, and uses the risk weights to weight the bounding box regression loss to obtain the lesion perception bounding box loss, thereby directionally strengthening the localization supervision of small-scale, low-contrast lesions with high risk of missed detection, and improving the model's feature attention to difficult lesions.

[0006] In a first aspect, embodiments of this application provide a method for training a tumor lesion target detection model for CT images of endometrial cancer, the method comprising:

[0007] CT images of endometrial cancer labeled with actual lesion bounding boxes are obtained as training images. The risk weight of each actual lesion bounding box in the training image is calculated based on the lesion area and local contrast. The lesion area is the area of ​​the corresponding actual lesion bounding box, and the local contrast is the absolute value of the difference between the gray value of the corresponding actual lesion bounding box and the gray value of the preset surrounding area. A tumor lesion detection model is constructed by using multiple training images to form a training dataset. The model is then iteratively trained using this dataset. Classification loss and bounding box regression loss are calculated using the predicted lesion bounding boxes and corresponding ground truth lesion bounding boxes obtained in each training round. The bounding box regression loss is weighted using corresponding risk weights to obtain the lesion-aware bounding box loss. A total loss function is constructed based on the classification loss and the lesion-aware bounding box loss. The trainable parameters of the tumor lesion detection model are updated based on this total loss function. The training is iteratively continued until the model converges, resulting in a well-trained tumor lesion detection model.

[0008] Secondly, embodiments of this application provide an application method for a tumor lesion target detection model, including: Acquire CT images of the endometrial cancer to be detected, and input the CT images of the endometrial cancer to be detected into a trained tumor lesion target detection model to obtain the tumor lesion detection results.

[0009] Thirdly, embodiments of this application provide an electronic device, including a memory and a processor, wherein the memory stores a computer program, and the processor is configured to run the computer program to execute a training method for a tumor lesion target detection model for CT images of endometrial cancer or an application method for a tumor lesion target detection model.

[0010] Fourthly, embodiments of this application provide a readable storage medium storing a computer program, which, when executed by a processor, implements, for example, a training method for a tumor lesion target detection model for CT images of endometrial cancer or an application method for a tumor lesion target detection model.

[0011] The main contributions and innovations of this invention are as follows: This application's embodiments calculate risk weights based on the lesion area and local contrast of the real lesion bounding box. These risk weights are then used to weight the bounding box regression loss to obtain the lesion perception bounding box loss. This approach specifically strengthens the localization supervision of small-scale, low-contrast lesions with high false negative risk, enhancing the model's focus on the features of difficult lesions and reducing their false negative risk. Furthermore, this application's embodiments utilize risk weights to correct the positive sample matching score of candidate anchor boxes, determining positive samples based on the corrected matching score. This increases the number of effective positive samples for high-risk lesions, alleviating the problem of insufficient positive sample allocation for small and low-contrast lesions, while maintaining training stability and ensuring sufficient training and learning for difficult lesions. This application's embodiments generate a soft coverage weight map based on the pixel distance to the lesion center, construct a coverage supervision loss and add it to the total loss function, thereby constraining the model's predictive response to the lesion region, guiding the model to focus on the lesion location, enhancing the feature learning effect of the lesion region, and helping to improve the detection ability of difficult lesions; this application's embodiments use a coverage supervision weight coefficient that decays linearly with the training rounds to dynamically weight and schedule the coverage supervision loss, so that in the early stage of training, a stronger coverage supervision guides the model to quickly focus on the lesion region, and in the later stage of training, the coverage supervision intensity is gradually reduced to avoid the auxiliary constraint interfering with the convergence of the main detection tasks such as classification and bounding box regression, thus balancing the guiding role and the optimization effect of the main task.

[0012] Details of one or more embodiments of this application are set forth in the following drawings and description to make other features, objects and advantages of this application more readily apparent. Attached Figure Description

[0013] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings: Figure 1 This is a flowchart illustrating a method for training a tumor lesion target detection model for CT images of endometrial cancer according to an embodiment of this application. Figure 2 This is a schematic diagram of the structure of a tumor lesion target detection model according to an embodiment of this application; Figure 3 This is a schematic diagram illustrating the training of a lesion-aware bounding box loss according to an embodiment of this application; Figure 4 This is a schematic diagram illustrating a training process that combines positive sample augmentation with coverage supervision loss according to an embodiment of this application. Figure 5 This is a comparison chart of the main detection indicators under different ablation experimental settings according to an embodiment of this application; Figure 6 This is a detection result diagram of a typical sample according to an embodiment of this application; Figure 7 This is a schematic diagram of the hardware structure of an electronic device according to an embodiment of this application. Detailed Implementation

[0014] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numerals in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with one or more embodiments of this specification. Rather, they are merely examples of apparatuses and methods consistent with some aspects of one or more embodiments of this specification as detailed in the appended claims.

[0015] It should be noted that the steps of the corresponding methods are not necessarily performed in the order shown and described in this specification in other embodiments. In some other embodiments, the methods may include more or fewer steps than described in this specification. Furthermore, a single step described in this specification may be broken down into multiple steps in other embodiments; and multiple steps described in this specification may be combined into a single step in other embodiments.

[0016] Example 1 This application provides a training method for a tumor lesion target detection model in CT images of endometrial cancer. This method calculates risk weights based on the lesion area and local contrast of the real lesion bounding box. These risk weights are then used to weight the bounding box regression loss to obtain the lesion perception bounding box loss. This method specifically strengthens the localization supervision of small-scale, low-contrast lesions with high risk of missed detection, improving the model's focus on the features of difficult-to-detect lesions. Specifically, refer to... Figure 1 The method includes: CT images of endometrial cancer labeled with actual lesion bounding boxes are obtained as training images. The risk weight of each actual lesion bounding box in the training image is calculated based on the lesion area and local contrast. The lesion area is the area of ​​the corresponding actual lesion bounding box, and the local contrast is the absolute value of the difference between the gray value of the corresponding actual lesion bounding box and the gray value of the preset surrounding area. A tumor lesion detection model is constructed by using multiple training images to form a training dataset. The model is then iteratively trained using this dataset. Classification loss and bounding box regression loss are calculated using the predicted lesion bounding boxes and corresponding ground truth lesion bounding boxes obtained in each training round. The bounding box regression loss is weighted using corresponding risk weights to obtain the lesion-aware bounding box loss. A total loss function is constructed based on the classification loss and the lesion-aware bounding box loss. The trainable parameters of the tumor lesion detection model are updated based on this total loss function. The training is iteratively continued until the model converges, resulting in a well-trained tumor lesion detection model.

[0017] In the current embodiment, this method obtains endometrial cancer CT images from the ECPC-IDS endometrial cancer CT tumor lesion dataset, and obtains usable training images by preprocessing, generating, annotating, and converting the annotation format of the endometrial cancer CT images.

[0018] Specifically, the preprocessing and annotation steps used in this scheme are all completed using existing technologies, which can be used to train the tumor lesion target detection model. This scheme does not limit the preprocessing methods or annotation methods.

[0019] In the current embodiment, an area weight value is calculated based on the lesion area, and a contrast weight value is calculated based on the local contrast. The sum of the default weight value, the area weight value, and the contrast weight value is used as the risk weight. The area weight value is negatively correlated with the lesion area, and the contrast weight value is negatively correlated with the local contrast.

[0020] Furthermore, preset area reference values ​​and contrast reference values ​​are defined. When the lesion area is greater than or equal to the area reference value, the area weight value is 0; when the local contrast is greater than or equal to the contrast reference value, the contrast weight value is 0.

[0021] Specifically, the default weight value in this scheme is 1, that is, when the lesion area is greater than or equal to the preset area reference value and the local contrast is greater than or equal to the preset contrast reference value, the corresponding risk weight is 1.

[0022] Specifically, the formula for calculating the lesion area within the actual lesion boundary is:

[0023] in, This is the index of the actual lesion's bounding box number. The area of ​​the lesion. The actual lesion boundary box The width is the actual lesion boundary box. The height.

[0024] Specifically, the formula for calculating the local contrast of the actual lesion boundary is:

[0025] in, This is the index of the actual lesion's bounding box number. For local contrast, The grayscale value is the value within the bounding box of the actual lesion. The preset surrounding area is a bounding box formed by extending n pixels outward from the actual lesion boundary box, and the area remaining after deducting the actual lesion boundary box itself. The preset surrounding area can be adjusted according to the actual situation.

[0026] Specifically, area reference values ​​and contrast reference values ​​are obtained by statistically analyzing the lesion area and local contrast of multiple training images in the dataset. Specifically, the area weight value is the product of the area adjustment factor and the area mapping value. The area adjustment factor is a preset value used to adjust the influence of the area weight value in the risk weight. The area mapping value is obtained based on the area risk function, and the formula of the area risk function is expressed as follows:

[0027] in, The area mapping value is obtained based on the area risk function. This is the index of the actual lesion's bounding box number. The area of ​​the lesion. The area reference value means that the area risk function can simultaneously control the negative correlation between the area weight value and the lesion area, and the area weight value is 0 when the lesion area is greater than or equal to the preset area reference value.

[0028] Specifically, the contrast weight value is the product of the contrast adjustment factor and the contrast mapping value. The contrast adjustment factor is a preset value used to adjust the influence of the contrast weight value in the risk weight. The contrast mapping value is obtained based on the contrast risk function, and the formula of the contrast risk function is expressed as follows:

[0029] in, The contrast mapping value is obtained based on the contrast risk function. This is the index of the actual lesion's bounding box number. For local contrast, The contrast reference value means that the contrast risk function can simultaneously adjust the contrast weight value to be negatively correlated with the local contrast, and the contrast weight value to be 0 when the local contrast is greater than or equal to the contrast reference value.

[0030] In other words, when the lesion area is smaller than the reference area, it means that the lesion is closer to the small target lesion, and its risk value increases. However, when the lesion area is not smaller than the reference area, the risk value of this part is not increased. Similarly, when the local contrast is lower than the reference contrast, it means that the lesion is less different from the surrounding tissue and the boundary is not obvious, thus increasing the risk value of this part.

[0031] The formula for calculating risk weights can be expressed as:

[0032] in, This is the index of the actual lesion's bounding box number. This represents the risk weight, with 1 being the default weight value. As an area adjustment factor, This is the contrast adjustment factor. This is based on the area mapping value obtained from the area risk function. This is the contrast mapping value obtained based on the contrast risk function.

[0033] In this current embodiment, the scheme uses YOLO as the basic architecture to construct a tumor lesion target detection model, and the structural diagram of the tumor lesion target detection model is shown below. Figure 2 As shown, In the current embodiment, the classification loss and bounding box regression loss mentioned in this scheme are both commonly used losses in the field of object detection and can be calculated using any loss calculation method. The classification loss is used to constrain the accuracy of the model in classifying lesion regions, and the bounding box regression loss is used to constrain the positional deviation between the predicted lesion bounding box and the real lesion bounding box.

[0034] Specifically, a schematic diagram illustrating the training process using a loss function to construct a lesion-sensing bounding box is shown below. Figure 3 As shown, the formula for obtaining the lesion-sensing bounding box loss by weighting the bounding box regression loss with corresponding risk weights is expressed as follows:

[0035] in, Loss of sensing bounding box for lesions The actual lesion boundary box Risk weights, The actual lesion boundary box The bounding box regression loss.

[0036] In the current embodiment, during the training process of the tumor lesion target detection model, multiple candidate anchor boxes are obtained in each round of training. The positive sample matching score between each candidate anchor box and the corresponding real lesion bounding box is calculated. The risk weight of the corresponding real lesion bounding box is obtained. The product of the corresponding positive sample matching score and the corresponding risk weight is used as the corrected matching score. When the corrected matching score is greater than the positive sample score threshold, the corresponding candidate anchor box is used as a positive sample.

[0037] In other words, this scheme adjusts the positive sample matching score by using risk weights during training, thereby effectively improving the training samples for small targets and allowing edge candidate positions that the model would easily ignore to participate in training, thus directly alleviating the problem of insufficient positive sample allocation for small-scale, low-contrast lesions.

[0038] Specifically, for candidate anchor boxes Its boundary with the corresponding real lesion The positive sample matching score is Then, the formula for obtaining its corrected matching score can be expressed as:

[0039] in, To correct the matching score, The actual lesion boundary box Risk weights, The score represents the positive sample matching score.

[0040] In other words, in this scheme, the higher the risk weight, the more likely the corresponding real lesion bounding box is to be missed. Therefore, by calculating and correcting the matching score, some candidate positions that originally had slightly lower matching scores but were still close to high-risk lesions are more likely to be assigned as positive samples, thereby increasing the number of effective training samples for difficult lesions.

[0041] Specifically, the positive sample matching score is calculated based on the intersection-union ratio of the candidate anchor box and the actual lesion bounding box, or the center distance between the two. The positive sample matching score is a conventional calculation method of the YOLO model, and this scheme does not impose detailed limitations.

[0042] For example, the actual lesion boundary box For the two candidate anchor boxes, the positive sample score threshold is 0.53. Calculations show that candidate anchor box 1 has a positive sample matching score of 0.55, and candidate anchor box 2 has a positive sample matching score of 0.52. Under the standard YOLO positive sample allocation rule, only candidate anchor box 1 has a positive sample matching score greater than the positive sample score threshold. Therefore, candidate anchor box 1 is determined to be a positive sample, and the true lesion bounding box... Only one valid training sample can be obtained. In this scheme, if the corresponding risk weight is 1.5, the corrected matching score of candidate anchor box 1 is 1.5×0.55=0.825, and the corrected matching score of candidate anchor box 2 is 1.5×0.52=0.78. Both candidate anchor box 1 and candidate anchor box 2 are judged as positive samples.

[0043] In the current embodiment, during the training of the tumor lesion target detection model, the center pixel coordinates of each real lesion bounding box are obtained. For each pixel in the training image corresponding to the real lesion bounding box, the distance between the pixel and the center pixel coordinates of the corresponding real lesion bounding box is used as the soft coverage distance. Soft coverage weights are assigned to each pixel based on the corresponding soft coverage distance. The soft coverage weights of each pixel in the training image are integrated to obtain the soft coverage weight map of the corresponding real lesion bounding box. A coverage supervision loss is constructed based on the soft coverage weight map as a loss term of the total loss function. The coverage supervision loss is used to constrain the prediction response of the tumor lesion target detection model to the real lesion bounding box during the training process. The larger the soft coverage distance, the more negatively correlated it is with the soft coverage weight.

[0044] Specifically, let the first The center pixel coordinates of the bounding box of each real lesion are The corresponding pixel in the training image is The formula for the soft cover weight of a pixel is expressed as:

[0045] in, To correspond to the actual lesion boundary box pixels soft cover weights, The distance attenuation coefficient is used to control the attenuation range of the real lesion bounding box, so as to constrain the soft coverage distance to be negatively correlated with the soft coverage weight. The distance attenuation coefficient can be predetermined according to the width and height of the real lesion bounding box. The distance attenuation coefficient can make the pixels close to the center pixel coordinates of the real lesion bounding box have high soft coverage weight, while the soft coverage weight of the training image boundary and the adjacent area gradually decreases.

[0046] Specifically, the formula for covering the supervision loss is expressed as:

[0047] in, To cover the losses from oversight, To train the number of true lesion bounding boxes within the image, To define the boundary of the actual lesion The corresponding soft-cover weight map, For tumor lesion target detection models, the bounding boxes of real lesions are used. The predicted response.

[0048] Furthermore, based on the training epochs, coverage supervision weight coefficients are assigned to the coverage supervision loss, and the formula for the coverage supervision weight coefficients is expressed as:

[0049] in, For the first The coverage supervision weight coefficient for each round, The preset initial coverage supervision weight coefficients, This refers to the total number of training rounds.

[0050] Specifically, based on the formula for the coverage supervision weight coefficient, it can be seen that the coverage supervision weight coefficient is larger in the early stage of training, so that the coverage supervision loss can help the model focus on the lesion area more quickly. In the later stage of training, the coverage supervision weight coefficient gradually decreases to avoid the coverage supervision being too strong and interfering with the optimization of the main detection tasks such as classification and bounding box regression.

[0051] Specifically, the positive sample enhancement described above can increase the number of positive samples for small-scale, low-contrast lesions that are difficult to detect and prone to being missed. Combined with coverage-supervised loss, this further guides the model to focus on the lesion region, thereby improving the model's feature learning ability for these difficult lesions. A schematic diagram illustrating the optimization training process combining positive sample enhancement and coverage-supervised loss is shown below. Figure 4 As shown.

[0052] In the current embodiment, the distribution focus loss is calculated as a loss term in the total loss function, and the formula for the total loss function is expressed as follows:

[0053] in, For the total loss function, For classifying losses, Loss of sensing bounding box for lesions For the distribution focus loss, For training rounds Coverage supervision weight coefficient at time, To cover the losses from oversight.

[0054] Specifically, the distribution focus loss is used to constrain the probability distribution of the predicted lesion bounding box.

[0055] Feasibility experiments were conducted to verify the feasibility of the tumor lesion target detection model trained in this scheme. YOLOv8s was used as a comparison model, and comparisons were made under the same data partitioning, input size, and formal evaluation protocol. The feasibility experiments consisted of four parts: First, YOLOv8s and the proposed method were compared under the original annotation protocol to verify whether the proposed method still had an improvement effect without annotation consistency verification; Second, the main experimental results were reported under the annotation consistency verification protocol as the final performance basis of this paper; Third, the effects of different technical means of this model were analyzed through ablation experiments; Fourth, the applicability of the proposed method in the current task was illustrated through comparison of different detection models and visualization case analysis.

[0056] The performance targets of this scheme are shown in Table 1. Considering the high sensitivity of medical lesion detection to false negatives, this paper focuses not only on mAP50 but also on Precision, Recall, and F1-score to comprehensively evaluate the model's detection capability and false positive control capability. Specifically, mAP50 mainly reflects the basic detection capability under the IoU=0.50 standard, Recall measures the proportion of lesions detected, Precision reflects the false positive control capability, and F1-score comprehensively evaluates the balance between Precision and Recall. Subsequent sections will present the core indicators in tables, along with bar charts and visualizations to illustrate the experimental phenomena.

[0057] Table 1 Preset Performance Targets

[0058] The original annotation protocol is used to verify the effectiveness of the proposed method without annotation consistency verification. The results of YOLOv8s and the model trained by the proposed method under the original annotation protocol are shown in Table 2.

[0059] Table 2 shows the results of YOLOv8s and the model trained using the proposed method under the original annotation protocol. Original YOLOv8sbaseline 0.7623 0.3874 0.7893 0.7549 0.7717 Original The model trained in this scheme 0.7858 0.3760 0.7983 0.7510 0.7739 As shown in Table 2, under the original annotation protocol, the proposed method improves mAP50 from 0.7623 to 0.7858, an increase of 0.0235; Precision from 0.7893 to 0.7983; and F1-score from 0.7717 to 0.7739. Recall decreases slightly from 0.7549 to 0.7510, but the overall F1-score still improves. This indicates that the proposed method does not solely rely on the annotation consistency verification protocol and can improve the model's overall detection performance under the original annotation conditions. It should also be noted that the model trained under this scheme has a slightly lower mAP50-95 under the original annotation protocol than YOLOv8s, indicating that the method significantly improves the basic detection capability under the IoU=0.50 standard, but still has shortcomings under the more stringent high IoU localization standard. This phenomenon is consistent with the task characteristics of blurred CT lesion boundaries and difficulty in locating small lesions. The purpose of retaining the original annotation protocol results is to demonstrate that the training-side improvements in this paper also have a certain effect under the original data conditions, and at the same time to provide a reference for the main results under the subsequent annotation consistency verification protocol.

[0060] The boundaries of lesions in medical images are somewhat ambiguous, and different annotators or data slices use significantly different scales to delineate these boundaries. Especially in the labeling of small lesions and low-contrast lesions, even a slight deviation in the bounding box position can cause large fluctuations in the Intersection over Union (IoU), thus significantly impacting the training performance and evaluation accuracy of the model. To reduce the interference that may be caused by labeling bias, while maintaining the original proportions of the training, validation, and test sets, this paper systematically reviewed samples of small lesions, low-contrast regions, and controversial boundaries. Based on the principle of standardization, a new, consistent-labeled dataset called clean_v2 was established.

[0061] The core objective of this review process is not to change the data distribution or to re-divide the test set, but rather to reduce annotation noise according to unified boundary rules. Candidate samples fall into three categories: first, small target regions caused by inaccurate bounding boxes; second, lesions due to unclear details or insufficient contrast; and third, near-hit boundary candidate samples where the predicted bounding box is close to the lesion but the IoU does not reach the threshold. Based on these categories, area statistics, grayscale change detection, and positional offset quantification analysis are comprehensively considered to select questionable lesion instances. The lesions are then corrected according to unified professional lesion identification standards, thereby significantly improving the consistency and accuracy of annotation results across various subsets.

[0062] It is important to note that annotation consistency verification is a crucial step in data quality control, but it is not the main algorithmic innovation discussed in this paper. Its primary purpose is to eliminate interference from factors such as blurred boundaries and inconsistent detection box sizes that affect model training effectiveness and performance evaluation. To ensure the objectivity and comparability of experimental results, the accuracy of object detection was re-evaluated according to a consistent annotation standard in all YOLOv8-based, optimized models and comparative experiments. The data performance obtained in this way is not defined as the performance of the original dataset, but rather described as "results under the annotation consistency verification protocol." The main experimental results under the clean_v2 protocol are shown in Table 3.

[0063] Table 3 shows the main experimental results under the clean_v2 protocol. clean_v2 YOLOv8sbaseline 0.7863 0.3985 0.8238 0.7391 0.7792 clean_v2 The model trained in this scheme 0.8020 0.3875 0.8067 0.7589 0.7821 As shown in Table 3, the proposed method achieved mAP50 of 0.8020, Precision of 0.8067, Recall of 0.7589, and F1-score of 0.7821, all meeting the preset performance targets. Compared with the YOLOv8s baseline, the proposed method improved mAP50 by 0.0157, Recall by 0.0198, and F1-score by 0.0029. Precision decreased from 0.8238 to 0.8067, but remained above the preset target of 0.78, indicating that the proposed method improved lesion recall within an acceptable false positive rate.

[0064] These results indicate that our proposed method primarily improves lesion detection and recall capabilities at the AP50 level. In medical lesion detection tasks, increased recall means that more real lesions can be detected, which is of practical significance in reducing the risk of missed detections. On the other hand, the mAP50-95 decreased from 0.3985 to 0.3875, indicating that our proposed method has not yet improved the fine boundary quality under strict localization thresholds. Therefore, our conclusion should be stated as "improving mAP50 and recall while meeting the requirements of precision and F1-score," rather than stating that all metrics are comprehensively superior to the baseline.

[0065] To verify the effectiveness of each technique described in this paper, ablation experiments were conducted under the annotation consistency review protocol. The ablation settings revolve around BoxReweight, RiskAssign, Coverage, and Annealing. BoxReweight corresponds to "weighting the bounding box regression loss with the corresponding risk weights," RiskAssign corresponds to positive sample allocation, Coverage corresponds to coverage supervision loss, and Annealing corresponds to the allocation of coverage supervision weight coefficients. The results of the ablation experiments are shown in Table 4.

[0066] Table 4 Results of the ablation experiment Base 0.7863 0.3985 0.8238 0.7391 0.7792 BR yes 0.7901 0.3867 0.7680 0.7589 0.7634 BR+RA yes yes 0.7878 0.3857 0.8166 0.7391 0.7759 BR+Cov yes yes yes 0.7882 0.3875 0.8161 0.7194 0.7647 Full-A yes yes yes 0.7767 0.3879 0.8161 0.7194 0.7647 Full yes yes yes yes 0.8020 0.3875 0.8067 0.7589 0.7821 Table 4 shows that after using Box Reweight alone, Recall increased from 0.7391 to 0.7589, and mAP50 increased from 0.7863 to 0.7901, but Precision decreased to 0.7680. This indicates that reweighting for missed detection risk can enhance the learning of difficult lesions, but its use alone can easily lead to an increase in false positives. After adding Risk Assign, Precision remained at 0.8166, but Recall did not improve significantly, indicating that risk-aware positive sample allocation helps maintain training stability, but is still insufficient to solve the recall problem of difficult lesions alone. To more intuitively illustrate the changing trends of the main indicators under different ablation settings, this paper further plots a comparison chart of the main detection indicators under different ablation experimental settings, as shown below. Figure 5 As shown, by Figure 5 It can be seen that when BoxReweight+Cov is used without Risk Assign, Recall drops to 0.7194, indicating that coverage supervision cannot be used in isolation and still requires reasonable positive sample allocation. Full-A's mAP50 drops to 0.7767, indicating that if coverage constraints are not staged and scheduled in the later stages of training, it will affect detection ranking and final performance. Full achieves the highest mAP50 and F1-score, indicating that BoxReweight increases attention to difficult samples, Risk Assign stabilizes positive sample learning, Coverage enhances response in lesion regions, and Annealing avoids excessively strong coverage constraints in the later stages. These results also demonstrate that the value of this approach lies not in a single step, but in using the same prior risk of missed detections to connect the supervised object, supervised region, and supervised time sequence, forming a continuous training supervision mechanism.

[0067] To further verify the effectiveness of the training method proposed in this scheme, different types of detection models were selected for comparison, including a lightweight YOLO model, a larger YOLO model, a two-stage detector, and a one-stage anchor-based detector. All models were trained and evaluated under the same annotation consistency verification protocol and adopted the same formal evaluation process. The experimental results of the different detection models are shown in Table 5.

[0068] Table 5 Comparative experimental results of different detection models YOLOv8n Phase 1 YOLO 0.5632 0.2083 0.6736 0.6364 0.6545 YOLOv8m Phase 1 YOLO 0.4641 0.1712 0.6000 0.5336 0.5649 YOLO11s Phase 1 YOLO 0.5216 0.2056 0.6751 0.5257 0.5911 FasterR-CNN Two-stage detector 0.5149 0.1831 0.3712 0.7233 0.4906 RetinaNet One-stage anchor-based 0.5217 0.1840 0.5229 0.6324 0.5725 The model trained in this scheme Training based on YOLOv8s 0.8020 0.3875 0.8067 0.7589 0.7821 As shown in Table 5, our proposed method outperforms other models in the comparison in terms of mAP50, Precision, Recall, and F1-score. YOLOv8n has a smaller parameter set and limited feature representation capabilities; YOLOv8m has a larger parameter set but did not achieve better results with the current data scale, indicating that simply increasing the model size does not consistently improve performance on this task. Faster R-CNN has a relatively high Recall but low Precision, indicating a higher number of false positives; the overall metrics of RetinaNet and YOLO11s also do not surpass our proposed method. These results do not mean that the aforementioned detectors are weaker than our proposed method in all medical image tasks, but rather that, under the current dataset, training settings, and formal evaluation protocol, the improved YOLOv8s, designed to address the risk of missed detections, is more suitable for our task.

[0069] To further observe the actual detection performance of the model on difficult lesions, this paper selects typical samples for visualization analysis, focusing on the samples that YOLOv8 missed but the method in this paper successfully detected. The detection results of typical samples are as follows: Figure 6 As shown, by Figure 6 It can be seen that the method in this paper can detect some lesions that YOLOv8sbaseline failed to detect, especially lesions with small area or low local contrast. This shows that training with a sense of missed detection risk can enhance the model's attention to difficult lesions and help improve lesion recall.

[0070] Example 2 Based on the same concept, this application also proposes an application method for a tumor lesion target detection model, including: Acquire CT images of endometrial cancer to be detected, and input these images into the tumor lesion target detection model trained in Example 1 to obtain the tumor lesion detection results.

[0071] Example 3 This embodiment also provides an electronic device, see reference. Figure 7 It includes a memory 402 and a processor 401, the memory 402 storing a computer program and the processor 401 being configured to run the computer program to perform the steps in any of the above method embodiments.

[0072] Specifically, the processor 401 may include a central processing unit (CPU), an application specific integrated circuit (ASIC), or one or more integrated circuits that can be configured to implement the embodiments of this application.

[0073] The memory 402 may include a mass storage device for data or instructions. For example, and not limitingly, the memory 402 may include a hard disk drive (HDD), a floppy disk drive, a solid-state drive (SSD), flash memory, an optical disk drive, a magneto-optical disk drive, magnetic tape, or a Universal Serial Bus (USB) drive, or a combination of two or more of these. Where appropriate, the memory 402 may include removable or non-removable (or fixed) media. Where appropriate, the memory 402 may be internal or external to a data processing device. In a particular embodiment, the memory 402 is non-volatile memory. In a particular embodiment, the memory 402 includes read-only memory (ROM) and random access memory (RAM). Where appropriate, the ROM may be a mask-programmed ROM, a programmable read-only memory (PROM), an erasable read-only memory (EPROM), an electrically erasable read-only memory (EEPROM), an electrically alterable read-only memory (EAROM), or flash memory, or a combination of two or more of these. Where appropriate, the RAM can be Static Random-Access Memory (SRAM) or Dynamic Random-Access Memory (DRAM). DRAM can be Fast Page Mode Dynamic Random Access Memory (FPMDRAM), Extended Data Out Dynamic Random Access Memory (EDODRAM), Synchronous Dynamic Random-Access Memory (SDRAM), etc.

[0074] The memory 402 can be used to store or cache various data files that need to be processed and / or communicated, as well as possible computer program instructions executed by the processor 401.

[0075] The processor 401 reads and executes computer program instructions stored in the memory 402 to implement any of the training methods for tumor lesion target detection models for endometrial cancer CT images in the above embodiments.

[0076] Optionally, the electronic device may further include a transmission device 403 and an input / output device 404, wherein the transmission device 403 is connected to the processor 401 and the input / output device 404 is connected to the processor 401.

[0077] The transmission device 403 can be used to receive or send data via a network. Specific examples of the network described above may include wired or wireless networks provided by the communication provider of the electronic device. In one example, the transmission device includes a Network Interface Controller (NIC), which can connect to other network devices via a base station to communicate with the Internet. In another example, the transmission device 403 may be a Radio Frequency (RF) module used for wireless communication with the Internet.

[0078] The input / output device 404 is used to input or output information. In this embodiment, the input information may be a CT image of endometrial cancer with the actual lesion bounding box marked, and the output information may be the update results of the trainable parameters of the tumor lesion target detection model, etc.

[0079] Optionally, in this embodiment, the processor 401 can be configured to perform the following steps via a computer program: CT images of endometrial cancer labeled with actual lesion bounding boxes are obtained as training images. The risk weight of each actual lesion bounding box in the training image is calculated based on the lesion area and local contrast. The lesion area is the area of ​​the corresponding actual lesion bounding box, and the local contrast is the absolute value of the difference between the gray value of the corresponding actual lesion bounding box and the gray value of the preset surrounding area. A tumor lesion detection model is constructed by using multiple training images to form a training dataset. The model is then iteratively trained using this dataset. Classification loss and bounding box regression loss are calculated using the predicted lesion bounding boxes and corresponding ground truth lesion bounding boxes obtained in each training round. The bounding box regression loss is weighted using corresponding risk weights to obtain the lesion-aware bounding box loss. A total loss function is constructed based on the classification loss and the lesion-aware bounding box loss. The trainable parameters of the tumor lesion detection model are updated based on this total loss function. The training is iteratively continued until the model converges, resulting in a well-trained tumor lesion detection model.

[0080] It should be noted that the specific examples in this embodiment can refer to the examples described in the above embodiments and optional implementations, and will not be repeated here.

[0081] Generally, various embodiments can be implemented in hardware or dedicated circuitry, software, logic, or any combination thereof. Some aspects of the invention can be implemented in hardware, while others can be implemented by firmware or software executed by a controller, microprocessor, or other computing device, but the invention is not limited thereto. Although various aspects of the invention may be shown and described as block diagrams, flowcharts, or using some other graphical representation, it should be understood that, by way of non-limiting example, these blocks, apparatuses, systems, techniques, or methods described herein can be implemented in hardware, software, firmware, dedicated circuitry or logic, general-purpose hardware or controllers or other computing devices, or some combination thereof.

[0082] Embodiments of the present invention can be implemented by computer software, which may be executable by a data processor of a mobile device, such as a processor entity, or by hardware, or by a combination of software and hardware. Computer software or programs (also referred to as program products) including software routines, applets, and / or macros can be stored in any device-readable data storage medium, and they include program instructions for performing specific tasks. The computer program product may include one or more computer-executable components configured to perform the embodiments when the program is run. The one or more computer-executable components may be at least one piece of software code or a portion thereof. Additionally, it should be noted in this respect that, as Figure 7 Any box in the logical flow can represent a program step, or interconnected logic circuits, boxes and functions, or a combination of program steps and logic circuits, boxes and functions. Software can be stored on physical media such as memory chips or blocks of storage implemented within a processor, magnetic media such as hard disks or floppy disks, and optical media such as DVDs and their data variants, CDs, etc. The physical medium is a non-transient medium.

[0083] Those skilled in the art should understand that the technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments have been described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0084] The above embodiments are merely illustrative of several implementation methods of this application, and their descriptions are relatively specific and detailed, but they should not be construed as limiting the scope of this application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A training method for a tumor lesion target detection model in CT images of endometrial cancer, characterized in that, Includes the following steps: CT images of endometrial cancer labeled with actual lesion bounding boxes are obtained as training images. The risk weight of each actual lesion bounding box in the training image is calculated based on the lesion area and local contrast. The lesion area is the area of ​​the corresponding actual lesion bounding box, and the local contrast is the absolute value of the difference between the gray value of the corresponding actual lesion bounding box and the gray value of the preset surrounding area. A tumor lesion detection model is constructed by using multiple training images to form a training dataset. The model is then iteratively trained using this dataset. Classification loss and bounding box regression loss are calculated using the predicted lesion bounding boxes and corresponding ground truth lesion bounding boxes obtained in each training round. The bounding box regression loss is weighted using corresponding risk weights to obtain the lesion-aware bounding box loss. A total loss function is constructed based on the classification loss and the lesion-aware bounding box loss. The trainable parameters of the tumor lesion detection model are updated based on this total loss function. The training is iteratively continued until the model converges, resulting in a well-trained tumor lesion detection model.

2. The training method for a tumor lesion target detection model for CT images of endometrial cancer according to claim 1, characterized in that, The area weight value is calculated based on the lesion area, and the contrast weight value is calculated based on the local contrast. The sum of the default weight value, the area weight value, and the contrast weight value is used as the risk weight. The area weight value is negatively correlated with the lesion area, and the contrast weight value is negatively correlated with the local contrast.

3. The training method for a tumor lesion target detection model for CT images of endometrial cancer according to claim 2, characterized in that, The preset area reference value and contrast reference value are used. When the lesion area is greater than or equal to the area reference value, the area weight value is 0; when the local contrast is greater than or equal to the contrast reference value, the contrast weight value is 0.

4. The training method for a tumor lesion target detection model for CT images of endometrial cancer according to claim 1, characterized in that, During the training of the tumor lesion target detection model, multiple candidate anchor boxes are obtained in each round of training. The positive sample matching score between each candidate anchor box and the corresponding real lesion bounding box is calculated, and the risk weight of the corresponding real lesion bounding box is obtained. The product of the corresponding positive sample matching score and the corresponding risk weight is used as the corrected matching score. When the corrected matching score is greater than the positive sample score threshold, the corresponding candidate anchor box is used as a positive sample.

5. The training method for a tumor lesion target detection model for CT images of endometrial cancer according to claim 1, characterized in that, During the training of the tumor lesion target detection model, the center pixel coordinates of each real lesion bounding box are obtained. For each pixel in the training image corresponding to the real lesion bounding box, the distance between the pixel and the center pixel coordinates of the corresponding real lesion bounding box is used as the soft coverage distance. Soft coverage weights are assigned to each pixel based on the corresponding soft coverage distance. The soft coverage weights of each pixel in the training image are integrated to obtain the soft coverage weight map of the corresponding real lesion bounding box. A coverage supervision loss is constructed based on the soft coverage weight map as a loss term of the total loss function. The coverage supervision loss is used to constrain the prediction response of the tumor lesion target detection model to the real lesion bounding box during the training process. The larger the soft coverage distance, the more negatively correlated it is with the soft coverage weight.

6. The training method for a tumor lesion target detection model for CT images of endometrial cancer according to claim 5, characterized in that, Based on the training epochs, a coverage supervision weight coefficient is assigned to the coverage supervision loss. The formula for the coverage supervision weight coefficient is as follows: in, For the first The coverage supervision weight coefficient for each round, The preset initial coverage supervision weight coefficients, This refers to the total number of training rounds.

7. A training method for a tumor lesion target detection model for CT images of endometrial cancer according to claim 1 or 5, characterized in that, If the distribution focus loss is calculated as the loss term in the total loss function, then the formula for the total loss function is as follows: in, For the total loss function, For classifying losses, Loss of sensing bounding box for lesions For the distribution focus loss, For training rounds Coverage supervision weight coefficient at time, To cover the losses from oversight.

8. An application method for a tumor lesion target detection model, characterized in that, include: Acquire CT images of endometrial cancer to be detected, and input the CT images of endometrial cancer to be detected into the tumor lesion target detection model trained by any of the methods described in claims 1-7 to obtain the tumor lesion detection results.

9. An electronic device comprising a memory and a processor, characterized in that, The memory stores a computer program, and the processor is configured to run the computer program to execute a training method for a tumor lesion target detection model for CT images of endometrial cancer as described in any one of claims 1-7, or an application method for a tumor lesion target detection model as described in claim 8.

10. A readable storage medium, characterized in that, The readable storage medium stores a computer program, which, when executed by a processor, implements a training method for a tumor lesion target detection model for CT images of endometrial cancer as described in any one of claims 1-7, or an application method for a tumor lesion target detection model as described in claim 8.

Images