Thyroid nodule detection system

By using a thyroid nodule detection system to detect and track thyroid ultrasound videos frame by frame, and combining this with a trajectory management module for quality assessment, the system solves the problems of existing systems being unable to process video temporal information and lacking automatic assessment capabilities. This enables end-to-end automated processing from video to detection results, improving screening efficiency and accuracy.

CN122023259APending Publication Date: 2026-05-12聆数医疗科技(苏州)有限公司 +1
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202512007476.9
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2025-12-29
Publication Date
2026-05-12

AI Technical Summary

Technical Problem

Existing thyroid nodule diagnostic systems cannot effectively process temporal information in video sequences, lack a complete screening workflow, and lack the ability to automatically assess scan quality, making it impossible to distinguish between real nodules and transient artifacts in videos.

Method used

A thyroid nodule detection system is adopted, including a receiving module, a multi-target instance segmentation module, an instance tracking module, and a determination module. The system uses a trained target instance segmentation model and a multi-target tracking algorithm to detect and track thyroid ultrasound videos frame by frame. Combined with a trajectory management module, the system performs quality assessment and filters unstable tracking trajectories, achieving end-to-end automated processing from video to detection results.

Benefits of technology

It significantly improves the efficiency and automation of thyroid nodule screening, can adapt to large-scale screening scenarios, improves the accuracy and consistency of test results, and supports batch processing and automated report generation.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN122023259A_ABST
    Figure CN122023259A_ABST
Patent Text Reader

Abstract

The invention provides a thyroid nodule detection system, and the system comprises a receiving module which is configured to receive a thyroid ultrasound video; the multi-target instance segmentation module is configured to perform frame-by-frame detection on target objects in each video frame through a trained target instance segmentation model to obtain detection data of each target object in each video frame, the detection data comprises a detection frame and confidence, and each target object comprises a carotid artery, a thyroid gland and a thyroid nodule; the instance tracking module is configured to track each target object in the thyroid ultrasound video by adopting a multi-target tracking algorithm based on target detection based on each detection data to obtain each tracking trajectory of each target object; and the determination module is configured to determine a target thyroid nodule detection result corresponding to the thyroid ultrasound video based on each tracking trajectory of each target object. According to the invention, the video can be processed for an end-to-end screening scene, and the detection accuracy is improved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the fields of image processing technology and medical device technology, and in particular to a thyroid nodule detection system. Background Technology

[0002] Current thyroid nodule diagnostic systems are mainly classification diagnostic systems based on single-frame images and report generation systems that rely on pre-segmentation.

[0003] Classification and diagnostic systems based on single-frame images are essentially static image analyses and cannot handle temporal information in video sequences. Furthermore, report generation systems relying on pre-segmentation heavily depend on a separate, pre-defined nodule detection and segmentation module, making the entire system not a true end-to-end screening solution. Therefore, an effective solution is urgently needed to address these issues. Summary of the Invention

[0004] To address the above problems, the present invention provides a thyroid nodule detection system.

[0005] This invention provides a thyroid nodule detection system, comprising: A receiving module is configured to receive thyroid ultrasound video, the thyroid ultrasound video comprising multiple video frames; The multi-target instance segmentation module is configured to perform frame-by-frame detection of target objects in each video frame using a trained target instance segmentation model, and obtain detection data of each target object in each video frame. The detection data includes detection boxes and confidence scores. Each target object includes carotid arteries, thyroid glands and thyroid nodules. The instance tracking module is configured to track each target object in the thyroid ultrasound video based on the detection data and using a multi-target tracking algorithm based on target detection, thereby obtaining each tracking trajectory of each target object; The determination module is configured to determine the detection result of the target thyroid nodule corresponding to the thyroid ultrasound video based on each tracking trajectory of the target object. A thyroid nodule detection system according to the present invention further includes a trajectory management module, configured to: For each target object, a quality assessment is performed on each tracking trajectory of the target object to obtain a target quality assessment result for each tracking trajectory of the target object; The tracking trajectories whose target quality assessment results are not satisfactory are filtered out.

[0006] According to the thyroid nodule detection system provided by the present invention, the quality assessment includes a first quality assessment; The trajectory management module is specifically configured as follows: For each target object, the first quality assessment is performed on each of the tracking trajectories of the target object to obtain the first quality assessment result of each of the tracking trajectories of the target object. The first quality assessment index includes at least one of trajectory length, continuity and motion smoothness. Based on the first quality assessment result of each tracking trajectory of the target object, the target quality assessment result of each tracking trajectory of the target object is determined.

[0007] According to the thyroid nodule detection system provided by the present invention, the trajectory management module is specifically configured as follows: For each tracking trajectory of the target object, the number of video frames representing the trajectory length corresponding to the tracking trajectory is obtained, and the trajectory length evaluation result of the tracking trajectory is determined based on the number of video frames. Obtain the number of interval frames representing the continuity of the tracking trajectory, and determine the continuity evaluation result of the tracking trajectory based on the number of interval frames; Obtain the continuous rate of change representing the motion smoothness corresponding to the tracking trajectory, and determine the motion smoothness evaluation result of the tracking trajectory based on the continuous rate of change. The continuous rate of change is the rate of change of the overlap of the detection boxes and / or the rate of change of the position of the detection boxes in consecutive video frames. The first quality assessment result of the tracking trajectory is determined based on at least one of the trajectory length assessment result, the continuity assessment result, and the motion smoothness assessment result.

[0008] According to the thyroid nodule detection system provided by the present invention, the quality assessment includes a second quality assessment; The trajectory management module is specifically configured as follows: For each target object, obtain the confidence sequence corresponding to each tracking trajectory of the target object, the confidence sequence containing the confidence of all detection boxes corresponding to the tracking trajectory; Based on the confidence sequence corresponding to each of the tracking trajectories of the target object, a second quality assessment is performed on each of the tracking trajectories of the target object to obtain a second quality assessment result for each of the tracking trajectories of the target object. The indicators of the second quality assessment include at least one of the confidence mean, confidence minimum, and confidence volatility. Based on the second quality assessment results of each tracking trajectory of the target object, the target quality assessment results of each tracking trajectory of the target object are determined.

[0009] According to the thyroid nodule detection system provided by the present invention, the trajectory management module is specifically configured as follows: Based on the first quality assessment result and / or the second quality assessment result, determine the confidence score of each of the tracking trajectories of the target object; The target quality assessment result of the tracking trajectory whose confidence score is less than the scoring threshold is determined to be unsuccessful, and the target quality assessment result of the tracking trajectory whose confidence score is greater than or equal to the scoring threshold is determined to be successful.

[0010] According to the thyroid nodule detection system provided by the present invention, the instance tracking module is specifically configured as follows: For each of the target objects, perform the following steps: For each video frame, the position of each existing trajectory of the target object in the video frame is predicted based on Kalman filtering, and the target object's tracking trajectory is obtained by two associations based on the detection data corresponding to the video frame and the position of each existing trajectory in the video frame.

[0011] According to the thyroid nodule detection system provided by the present invention, the instance tracking module is specifically configured as follows: Based on the confidence level of the target object in the video frame, determine the detection box type of the target object in the video frame; When the detection box type is a high-confidence detection box, the high-confidence detection box is associated with each of the existing trajectories based on the position of each of the existing trajectories in the video frame; When the detection box type is a low confidence detection box, the low confidence detection box is associated with each of the remaining unmatched trajectories in the video frame based on the position of each of the existing trajectories. If the detection box type is the high-confidence detection box and the association is unsuccessful, a new trajectory is created for the target object based on the high-confidence detection box.

[0012] According to the thyroid nodule detection system provided by the present invention, the instance tracking module is specifically configured as follows: If the confidence level of the target object in the video frame is greater than or equal to the high confidence threshold, the detection box type of the target object in the video frame is determined to be the high confidence detection box. If the confidence level of the target object in the video frame is less than the high confidence threshold and greater than or equal to the low confidence threshold, the detection box type of the target object in the video frame is determined to be the low confidence detection box.

[0013] A thyroid nodule detection system according to the present invention further includes a training module configured to: Acquire multiple initial images containing a specified object, wherein the specified object is at least one of a carotid artery, a thyroid gland, and a thyroid nodule; Each specified object in each of the initial images is labeled, and data augmentation is performed on each of the initial images to obtain a training image set; Based on the training image set, the initial instance segmentation model is trained to obtain the target instance segmentation model.

[0014] According to the thyroid nodule detection system provided by the present invention, the training module is specifically configured as follows: Random brightness and contrast adjustments and / or Gaussian noise additions are performed on each of the initial images to achieve data augmentation, resulting in the training image set. The Gaussian noise addition includes simulated ultrasound artifacts.

[0015] According to the thyroid nodule detection system provided by the present invention, the detection data further includes a pixel-level segmentation mask; The determining module is specifically configured as follows: Based on the tracking trajectories of the thyroid nodules, the initial thyroid nodule detection results are determined; Based on at least one of the following: the tracking trajectory of the carotid artery, the pixel-level segmentation mask of the carotid artery, the tracking trajectory of the thyroid gland, the pixel-level segmentation mask of the thyroid gland, and the pixel-level segmentation mask of the thyroid nodule, the initial thyroid nodule detection result is subjected to false positive filtering to obtain the target thyroid nodule detection result.

[0016] This invention also provides a method for detecting thyroid nodules, comprising the following modules: Receive thyroid ultrasound video, wherein the thyroid ultrasound video includes multiple video frames; By using a trained target instance segmentation model, the target objects in each video frame are detected frame by frame to obtain the detection data of each target object in each video frame. The detection data includes detection boxes and confidence scores. Each target object includes carotid arteries, thyroid glands and thyroid nodules. Based on the detection data, a multi-target tracking algorithm based on target detection is used to track each target object in the thyroid ultrasound video to obtain each tracking trajectory of each target object; Based on the tracking trajectories of each target object, the detection result of the target thyroid nodule corresponding to the thyroid ultrasound video is determined.

[0017] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the thyroid nodule detection method as described above.

[0018] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the thyroid nodule detection method as described above.

[0019] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the thyroid nodule detection method as described above.

[0020] The thyroid nodule detection system provided by this invention includes a receiving module configured to receive thyroid ultrasound videos, which comprise multiple video frames; a multi-target instance segmentation module configured to perform frame-by-frame detection of target objects in each video frame using a trained target instance segmentation model, obtaining detection data for each target object in each video frame, the detection data including bounding boxes and confidence scores, and each target object including carotid arteries, thyroid glands, and thyroid nodules; an instance tracking module configured to track each target object in the thyroid ultrasound video based on the detection data using a multi-target tracking algorithm based on target detection, obtaining tracking trajectories for each target object; and a determination module configured to determine the target thyroid nodule detection result corresponding to the thyroid ultrasound video based on the tracking trajectories of each target object. This invention can significantly improve screening efficiency, greatly enhance automation, and achieve end-to-end automated processing from raw video to thyroid nodule detection results; it supports batch processing and is suitable for large-scale screening scenarios. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0022] Figure 1 This is a schematic diagram of the thyroid nodule detection system provided by the present invention.

[0023] Figure 2 This is a schematic diagram of multi-target instance segmentation provided by the present invention.

[0024] Figure 3 This is a flowchart illustrating the thyroid nodule detection method provided by the present invention.

[0025] Figure 4 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0026] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0027] First, a brief description of the relevant content involved in this invention will be given.

[0028] The classification and diagnostic system based on single-frame images mainly introduces frequency domain channel attention into each residual block of the residual network and uses an encoder-decoder with multi-scale deformable attention for final prediction. While this approach improves detection performance to some extent, it is essentially still based on static image analysis and cannot handle temporal information in video sequences.

[0029] The report generation system, which relies on pre-segmentation, combines a feature extraction network, a semantic mapping module, and a text generation model based on a next-generation dialogue pre-trained model. Through fine-tuning methods and function call mechanisms, it automatically generates structured diagnostic reports containing Thyroid Imaging Reporting and Data System (TI-RADS) scores. However, this system cannot process raw thyroid ultrasound images / videos and heavily depends on a separate, pre-installed nodule detection and segmentation module. Therefore, the entire system is not a true "end-to-end" screening solution.

[0030] Furthermore, current thyroid nodule diagnostic systems lack a complete workflow for screening scenarios. Most existing systems focus on specific aspects of the diagnostic process, such as nodule detection, benign / malignant classification, or report generation, but lack a complete automated process from video acquisition to final screening conclusions. In real-world screening scenarios, the system needs to be able to process the entire scan video and provide clear screening conclusions and recommendations.

[0031] Current thyroid nodule diagnostic systems are ill-suited to the specific needs of screening scenarios. Screening scenarios have distinct requirements compared to diagnostic scenarios: screening demands higher specificity to reduce unnecessary further testing, requires clear classification criteria to guide subsequent treatment, and necessitates automated report generation to improve efficiency. Most existing systems are designed for diagnostic scenarios and cannot adequately meet the specific requirements of screening.

[0032] Current thyroid nodule diagnostic systems lack video quality control and stability assessment. Most existing methods assume good input data quality, but in actual screening, issues such as non-standard scanning, excessively rapid probe movement, and poor image quality frequently arise. The systems lack the ability to automatically assess scan quality and cannot distinguish between real nodules and transient artifacts based on the stability of nodule appearance in the video.

[0033] To address at least one of the aforementioned problems, the present invention provides a thyroid nodule detection system.

[0034] The following is combined with Figures 1-4 The thyroid nodule detection system of the present invention is described.

[0035] Figure 1 This is a schematic diagram of the thyroid nodule detection system provided by the present invention, as shown below. Figure 1 As shown, the thyroid nodule detection system includes the following: The receiving module 101 is configured to receive thyroid ultrasound video, which includes multiple video frames. The multi-target instance segmentation module 102 is configured to perform frame-by-frame detection of target objects in each video frame using a trained target instance segmentation model to obtain detection data of each target object in each video frame. The detection data includes detection boxes and confidence scores. Each target object includes carotid arteries, thyroid glands, and thyroid nodules. Instance tracking module 103 is configured to track each target object in the thyroid ultrasound video based on the detection data and using a multi-target tracking algorithm based on target detection, thereby obtaining each tracking trajectory of each target object; The determination module 104 is configured to determine the detection result of the target thyroid nodule corresponding to the thyroid ultrasound video based on each tracking trajectory of the target object.

[0036] Specifically, the target instance segmentation model can use models such as YouOnly Look Once Version 11 with Segmentation Capability (YOLOv11-seg) or Real-Time Detection Transformer with MaskPrediction (Mask-RT-DETR) as its basic architecture. The target instance segmentation model can also use Mask-attention Mask Transformer with unified architecture (Mask2Former) or Mask Detection Transformer with Improved deNoising anchOr boxes (Mask DINO) as its basic architecture for instance segmentation. The target instance segmentation model can also use Shifted Windows Transformer (Swin Transformer) as its backbone network to improve feature extraction capabilities.

[0037] Specifically, multi-target tracking algorithms based on target detection can be ByteTrack algorithms, Observation-Centric Simple Online and Realtime Tracking (OC-SOR) algorithms, or graph neural networks can be used for multi-target association to enhance tracking stability in complex scenarios. This invention does not impose any limitations on these methods.

[0038] In practical applications, users can input thyroid ultrasound videos that require thyroid nodule detection into the thyroid nodule detection system via the receiving module.

[0039] Then, the thyroid nodule detection system uses a multi-target instance segmentation module to call a trained target instance segmentation model to perform frame-by-frame detection of the carotid artery, thyroid gland, and thyroid nodules in the thyroid ultrasound video, obtaining the detection results of the carotid artery, thyroid gland, and thyroid nodules in each video frame. The detection data includes detection boxes and confidence scores, and may also include pixel-level segmentation masks.

[0040] Figure 2 This is a schematic diagram of multi-target instance segmentation provided by the present invention. See [link / reference] Figure 2 , Figure 2The left image shows a video frame without multi-object instance segmentation, and the right image shows a video frame with multi-object instance segmentation. In the right image, the blue area is the pixel-level segmentation mask for thyroid gland T, and the blue detection box is the detection box for thyroid gland T. "T-1: 0.88" indicates that the confidence level of thyroid gland T is 0.88. The red area is the pixel-level segmentation mask for thyroid gland N, and the red detection box is the detection box for thyroid gland N. "N-1: 0.85" indicates that the confidence level of thyroid gland N is 0.85. The green area is the pixel-level segmentation mask for carotid artery C, and the green detection box is the detection box for carotid artery C. "C-1: 0.89" indicates that the confidence level of carotid artery C is 0.89.

[0041] After obtaining the detection results of the carotid artery, thyroid gland and thyroid nodules in each video frame, the thyroid nodule detection system uses the instance tracking module to track each target object based on the detection data of the target object in each video frame using a multi-target tracking algorithm, and obtains at least one tracking trajectory of the target object.

[0042] See Figure 2 In “T-1: 0.88”, 1 represents the tracking trajectory 1 of the thyroid gland T; in “N-1: 0.85”, 1 represents the tracking trajectory 1 of the thyroid gland N; in “C-1: 0.89”, 1 represents the tracking trajectory 1 of the carotid artery C.

[0043] Furthermore, the thyroid nodule detection system determines the detection result of the target thyroid nodule by identifying the tracking module, based on the respective tracking trajectories of the carotid artery, thyroid gland, and thyroid nodule, i.e. whether the thyroid ultrasound video indicates the presence of a thyroid nodule.

[0044] The thyroid nodule detection system provided by this invention includes a receiving module configured to receive thyroid ultrasound videos, which comprise multiple video frames; a multi-target instance segmentation module configured to perform frame-by-frame detection of target objects in each video frame using a trained target instance segmentation model, obtaining detection data for each target object in each video frame, the detection data including bounding boxes and confidence scores, and each target object including carotid arteries, thyroid glands, and thyroid nodules; an instance tracking module configured to track each target object in the thyroid ultrasound video based on the detection data using a multi-target tracking algorithm based on target detection, obtaining tracking trajectories for each target object; and a determination module configured to determine the target thyroid nodule detection result corresponding to the thyroid ultrasound video based on the tracking trajectories of each target object. This invention can significantly improve screening efficiency, greatly enhance automation, and achieve end-to-end automated processing from raw video to thyroid nodule detection results; it supports batch processing and is suitable for large-scale screening scenarios.

[0045] Optionally, the thyroid nodule detection system further includes a trajectory management module, configured to: For each target object, a quality assessment is performed on each tracking trajectory of the target object to obtain a target quality assessment result for each tracking trajectory of the target object; The tracking trajectories whose target quality assessment results are not satisfactory are filtered out.

[0046] In practical applications, the quality of each tracking trajectory can be evaluated to obtain the target quality evaluation result. The evaluation indicators can include multiple aspects, such as trajectory smoothness and trajectory continuity.

[0047] Furthermore, trajectory filtering is performed based on the target quality assessment results, i.e., filtering out tracking trajectories that fail the target quality assessment, i.e., transient and unstable detections. Further, the target thyroid nodule detection result is determined based on the remaining tracking trajectories.

[0048] In this embodiment of the invention, the tracking trajectory can be quality evaluated, and unstable tracking trajectories that appear briefly can be deleted to avoid interference with the detection results of thyroid nodules, thereby improving the accuracy of the detection results of the target thyroid nodules.

[0049] Optionally, the quality assessment includes a first quality assessment; The trajectory management module is specifically configured as follows: For each target object, the first quality assessment is performed on each of the tracking trajectories of the target object to obtain the first quality assessment result of each of the tracking trajectories of the target object. The first quality assessment index includes at least one of trajectory length, continuity and motion smoothness. Based on the first quality assessment result of each tracking trajectory of the target object, the target quality assessment result of each tracking trajectory of the target object is determined.

[0050] Specifically, the trajectory length can be the number of video frames corresponding to the tracking trajectory; the continuity can be the number of interval frames between the video frames corresponding to the tracking trajectory; and the motion smoothness can be the rate of change of the overlap of detection boxes and / or the rate of change of the position of detection boxes in the consecutive video frames corresponding to the tracking trajectory.

[0051] In practical applications, at least one of the following factors—trajectories length, continuity, and motion smoothness—can be evaluated to obtain a first quality evaluation result.

[0052] Furthermore, the target quality assessment result is determined based on the first quality assessment result. For example, if the first quality assessment result is "fail", the target quality assessment result will also be "fail"; if only the first quality assessment is performed, and the first quality assessment result is "pass", the target quality assessment result will also be "pass".

[0053] In this embodiment of the invention, the quality of the tracking trajectory can be evaluated from aspects such as trajectory length, continuity and motion smoothness, which can ensure the accuracy of the tracking trajectory that appears briefly and is unstable, thereby improving the accuracy of filtering.

[0054] Optionally, the trajectory management module is specifically configured as follows: For each tracking trajectory of the target object, the number of video frames representing the trajectory length corresponding to the tracking trajectory is obtained, and the trajectory length evaluation result of the tracking trajectory is determined based on the number of video frames. Obtain the number of interval frames representing the continuity of the tracking trajectory, and determine the continuity evaluation result of the tracking trajectory based on the number of interval frames; Obtain the continuous rate of change representing the motion smoothness corresponding to the tracking trajectory, and determine the motion smoothness evaluation result of the tracking trajectory based on the continuous rate of change. The continuous rate of change is the rate of change of the overlap of the detection boxes and / or the rate of change of the position of the detection boxes in consecutive video frames. The first quality assessment result of the tracking trajectory is determined based on at least one of the trajectory length assessment result, the continuity assessment result, and the motion smoothness assessment result.

[0055] In practical applications, for each tracking trajectory, the number of video frames corresponding to the tracking trajectory can be obtained. If the number of video frames is greater than a first number, the trajectory length evaluation result of the tracking trajectory is determined to be passed. If the number of video frames is less than or equal to the first number, the trajectory length evaluation result of the tracking trajectory is determined to be failed.

[0056] For each tracking trajectory, the number of interval frames corresponding to the tracking trajectory can be obtained. If the number of interval frames is greater than a second number, the continuity evaluation result of the tracking trajectory is determined to be unsuccessful; if the number of video frames is less than or equal to the second number, the continuity evaluation result of the tracking trajectory is determined to be successful. Here, the first number is greater than the second number.

[0057] For each tracking trajectory, the continuous rate of change corresponding to the tracking trajectory can be obtained. If the continuous rate of change is less than the rate of change threshold, the motion smoothness evaluation result of the tracking trajectory is determined to be passed. If the continuous rate of change is greater than or equal to the rate of change threshold, the motion smoothness evaluation result of the tracking trajectory is determined to be failed.

[0058] If the continuous rate of change includes the rate of change of overlap and the rate of change of position, then if at least one of the rate of change of overlap and the rate of change of position is less than its corresponding rate of change threshold, the continuous rate of change is determined to be less than the rate of change threshold. The rate of change thresholds for the rate of change of overlap and the rate of change of position can be the same or different.

[0059] Furthermore, a first quality assessment result for the tracking trajectory can be determined based on at least one of the trajectory length assessment result, continuity assessment result, and motion smoothness assessment result.

[0060] For example, if at least one of the trajectory length assessment result, continuity assessment result, and motion smoothness assessment result is "fail," it indicates that the tracking trajectory has failed the first quality assessment, i.e., the first quality assessment result is "fail." Otherwise, it indicates that the tracking trajectory has passed the first quality assessment, i.e., the first quality assessment result is "pass." Through quality control standardization, non-conforming tracking trajectories are automatically identified, reducing unnecessary work, and through unified assessment standards, the consistency of test results is ensured.

[0061] In this embodiment of the invention, quality assessment is performed from multiple dimensions such as trajectory length, continuity, and motion smoothness, which can ensure the accuracy of short-lived unstable tracking trajectories, thereby improving the accuracy of filtering.

[0062] Optionally, the quality assessment includes a second quality assessment; The trajectory management module is specifically configured as follows: For each target object, obtain the confidence sequence corresponding to each tracking trajectory of the target object, the confidence sequence containing the confidence of all detection boxes corresponding to the tracking trajectory; Based on the confidence sequence corresponding to each of the tracking trajectories of the target object, a second quality assessment is performed on each of the tracking trajectories of the target object to obtain a second quality assessment result for each of the tracking trajectories of the target object. The indicators of the second quality assessment include at least one of the confidence mean, confidence minimum, and confidence volatility. Based on the second quality assessment results of each tracking trajectory of the target object, the target quality assessment results of each tracking trajectory of the target object are determined.

[0063] Specifically, the confidence level volatility can be a parameter characterizing the degree of confidence level volatility, such as the confidence level standard deviation and / or confidence level variance.

[0064] In practical applications, the confidence sequence corresponding to each tracking trajectory can be evaluated for quality to obtain a second quality evaluation result.

[0065] First, for each tracking trajectory, the confidence level of the detection box associated with that trajectory is calculated to obtain the confidence level sequence for that trajectory.

[0066] Then, based on the confidence sequence corresponding to the tracking trajectory, a quality assessment is performed to obtain the second quality assessment result.

[0067] Furthermore, the target quality assessment result is determined based on the second quality assessment result. For example, if the second quality assessment result is unsuccessful, the target quality assessment result is also unsuccessful; if only the second quality assessment is performed, and the second quality assessment result is successful, the target quality assessment result is also successful.

[0068] In this embodiment of the invention, the quality of the tracking trajectory can be evaluated from aspects such as the mean confidence level, the minimum confidence level, and the confidence level volatility. This can ensure the accuracy of short-lived and unstable tracking trajectories, thereby improving the accuracy of filtering.

[0069] Optionally, the trajectory management module is specifically configured as follows: For each tracking trajectory of the target object, based on the confidence sequence corresponding to the tracking trajectory, determine the mean confidence level, minimum confidence level, and confidence level volatility corresponding to the tracking trajectory; Based on the mean confidence level corresponding to the tracking trajectory, the mean evaluation result of the tracking trajectory is determined; Based on the minimum confidence value corresponding to the tracking trajectory, the minimum value evaluation result of the tracking trajectory is determined; Based on the confidence volatility corresponding to the tracking trajectory, the volatility assessment result of the tracking trajectory is determined; The second quality assessment result of the tracking trajectory is determined based on at least one of the mean assessment result, the minimum assessment result, and the volatility assessment result.

[0070] In practical applications, for each tracking trajectory, the mean confidence level, minimum confidence level, and confidence level volatility can be determined based on the confidence level sequence.

[0071] If the mean confidence level is greater than the mean threshold, the mean evaluation result of the tracking trajectory is determined to be passed; if the mean confidence level is less than or equal to the mean threshold, the mean evaluation result of the tracking trajectory is determined to be failed.

[0072] If the minimum confidence level is greater than the minimum threshold, the minimum evaluation result of the tracking trajectory is determined to be passed; if the minimum confidence level is less than or equal to the minimum threshold, the minimum evaluation result of the tracking trajectory is determined to be failed.

[0073] If the confidence level volatility is less than the volatility threshold, the volatility result of the tracking trajectory is determined to be passed; if the confidence level volatility is greater than or equal to the volatility threshold, the volatility result of the tracking trajectory is determined to be failed.

[0074] If confidence level volatility includes confidence level standard deviation and confidence level variance, then if at least one of the confidence level standard deviation and confidence level variance is less than its corresponding volatility threshold, the confidence level volatility is determined to be less than the volatility threshold. The volatility thresholds for the confidence level standard deviation and confidence level variance can be the same or different.

[0075] Furthermore, a second quality assessment result for the tracking trajectory can be determined based on at least one of the mean assessment result, the minimum assessment result, and the volatility assessment result.

[0076] For example, if at least one of the mean assessment result, minimum value assessment result, and volatility assessment result is "fail," it indicates that the tracking trajectory has failed the second quality assessment, i.e., the second quality assessment result is "fail." Otherwise, it indicates that the tracking trajectory has passed the second quality assessment, i.e., the second quality assessment result is "pass." Through quality control standardization, non-conforming tracking trajectories are automatically identified, reducing unnecessary work, and through unified assessment standards, the consistency of test results is ensured.

[0077] In this embodiment of the invention, quality assessment is performed from multiple dimensions, including mean assessment results, minimum assessment results, and volatility assessment results. This ensures the accuracy of tracking trajectories that are temporarily unstable, thereby improving the accuracy of filtering.

[0078] Optionally, the trajectory management module is specifically configured as follows: Based on the first quality assessment result and / or the second quality assessment result, determine the confidence score of each of the tracking trajectories of the target object; The target quality assessment result of the tracking trajectory whose confidence score is less than the scoring threshold is determined to be unsuccessful, and the target quality assessment result of the tracking trajectory whose confidence score is greater than or equal to the scoring threshold is determined to be successful.

[0079] In practical applications, a trajectory confidence scoring mechanism can be established: for each tracking trajectory, the first quality assessment result and / or the second quality assessment result of the tracking trajectory are substituted into the trajectory confidence scoring strategy to obtain the confidence score of the tracking trajectory.

[0080] Furthermore, if the confidence score is less than the scoring threshold, it means that the tracking trajectory has not passed the quality assessment, and the tracking trajectory is filtered out; if the confidence score is greater than or equal to the scoring threshold, it means that the tracking trajectory has passed the quality assessment, and the tracking trajectory is retained.

[0081] In this invention, a trajectory confidence scoring mechanism is constructed to determine the confidence score of the tracking trajectory. Based on the confidence score and the scoring threshold, transient and unstable tracking trajectories can be filtered, thus improving the accuracy of the filtering.

[0082] Optionally, the instance tracking module 103 is specifically configured as follows: For each of the target objects, perform the following steps: For each video frame, the position of each existing trajectory of the target object in the video frame is predicted based on Kalman filtering, and the target object's tracking trajectory is obtained by two associations based on the detection data corresponding to the video frame and the position of each existing trajectory in the video frame.

[0083] In practical applications, the ByteTrack algorithm can be used for instance tracking. All bounding boxes in each detection result are correlated.

[0084] Specifically, from the first video frame to the last video frame, the following steps are performed sequentially: for each target object, use Kalman filtering to predict the position of the existing trajectory of the target object in the current video frame, and then associate the detection box of the target object in the current video frame with the existing trajectory in two steps; traverse all video frames to obtain each tracking trajectory of the target object.

[0085] In this invention, the accuracy of the tracking trajectory can be improved by using the ByteTrack algorithm to determine the tracking trajectory.

[0086] Optionally, the instance tracking module 103 is specifically configured as follows: Based on the confidence level of the target object in the video frame, determine the detection box type of the target object in the video frame; When the detection box type is a high-confidence detection box, the high-confidence detection box is associated with each of the existing trajectories based on the position of each of the existing trajectories in the video frame; When the detection box type is a low confidence detection box, the low confidence detection box is associated with each of the remaining unmatched trajectories in the video frame based on the position of each of the existing trajectories. If the detection box type is a high-confidence detection box and the association is unsuccessful, a new trajectory is created for the target object based on the high-confidence detection box.

[0087] Specifically, the detection box types include three categories: high-confidence detection boxes, low-confidence detection boxes (usually corresponding to fuzzy target objects), and invalid detection boxes.

[0088] In practical applications, in the current video frame, the type of the detection box of the target object is determined based on the confidence level of the target object, i.e., the detection box type.

[0089] Then, two associations are performed: First, the high-confidence detection box of the target object is associated with the corresponding predicted trajectory (existing trajectory). Second, the low-confidence detection box of the target object (usually corresponding to a blurred nodule) is associated again with the remaining unmatched trajectories, effectively reducing missed detections. The remaining unmatched trajectories are existing trajectories that failed to match successfully after the first association (associated with the high-confidence detection box). Finally, new trajectories are created for the still unmatched high-confidence detection boxes. If a high-confidence detection box appears in a single video frame and cannot be overlapped, it is treated as an independent trajectory, thus achieving stable and continuous nodule tracking in the ultrasound video sequence.

[0090] Furthermore, if the detection box type is an invalid detection box, the invalid detection box is discarded; if the detection box type is a low-confidence detection box and the association is unsuccessful, the low-confidence detection box is discarded.

[0091] In this embodiment of the invention, comprehensive tracking of the target object in the thyroid ultrasound video is ensured by secondary association and the creation of new trajectories.

[0092] Optionally, the instance tracking module 103 is specifically configured as follows: If the confidence level of the target object in the video frame is greater than or equal to the high confidence threshold, the detection box type of the target object in the video frame is determined to be the high confidence detection box. If the confidence level of the target object in the video frame is less than the high confidence threshold and greater than or equal to the low confidence threshold, the detection box type of the target object in the video frame is determined to be the low confidence detection box.

[0093] Specifically, the high confidence threshold is greater than the low confidence threshold.

[0094] In practical applications, detection box types can be categorized using high-confidence and low-confidence thresholds: detection boxes with a confidence level greater than or equal to the high-confidence threshold are classified as high-confidence detection boxes, those with a confidence level less than the low-confidence threshold are classified as invalid detection boxes, and the remaining detection boxes are classified as low-confidence detection boxes. In other words, detection boxes with a confidence level less than the high-confidence threshold but greater than or equal to the low-confidence threshold are classified as low-confidence detections. This allows for the determination of the accuracy and reliability of the detection box type.

[0095] Optionally, the thyroid nodule detection system further includes a training module configured to: Acquire multiple initial images containing a specified object, wherein the specified object is at least one of a carotid artery, a thyroid gland, and a thyroid nodule; Each specified object in each of the initial images is labeled, and data augmentation is performed on each of the initial images to obtain a training image set; Based on the training image set, the initial instance segmentation model is trained to obtain the target instance segmentation model.

[0096] In practical applications, models such as YOLOv11-seg or Mask-RT-DETR can be used as initial instance segmentation models due to their advantages in real-time instance segmentation.

[0097] You can first initialize the image, and then annotate the initial image, including three types of specified objects: carotid arteries, thyroid glands, and thyroid nodules.

[0098] Then, data augmentation is performed on each initial image (characteristics of ultrasound images) to obtain multiple backup images. The labeled initial and backup images are used as training images to construct a training image set.

[0099] Furthermore, the initial instance segmentation model is trained based on the training image set, wherein the output of the initial instance segmentation model includes a pixel-level segmentation mask, bounding box coordinates, and confidence score for each specified object.

[0100] In this embodiment of the invention, by performing image annotation and data augmentation, model training is carried out from two dimensions: result comparison and sample quantity, which helps to improve training efficiency and model robustness.

[0101] Optionally, the training module is specifically configured as follows: Random brightness and contrast adjustments and / or Gaussian noise additions are performed on each of the initial images to achieve data augmentation, resulting in the training image set. The Gaussian noise addition includes simulated ultrasound artifacts.

[0102] Specifically, data augmentation can be performed on the initial image through methods such as random brightness and contrast adjustment and / or Gaussian noise addition. Gaussian noise addition can include ordinary Gaussian noise addition and / or simulating ultrasound artifacts. By using different data augmentation methods, the diversity of training samples can be ensured, thereby improving model training efficiency and model robustness.

[0103] Optionally, the detection data may further include a pixel-level segmentation mask; The determining module 104 is specifically configured as follows: Based on the tracking trajectories of the thyroid nodules, the initial thyroid nodule detection results are determined; Based on at least one of the following: the tracking trajectory of the carotid artery, the pixel-level segmentation mask of the carotid artery, the tracking trajectory of the thyroid gland, the pixel-level segmentation mask of the thyroid gland, and the pixel-level segmentation mask of the thyroid nodule, the initial thyroid nodule detection result is subjected to false positive filtering to obtain the target thyroid nodule detection result.

[0104] In practical applications, the detection (tracing trajectories) of the carotid artery and thyroid gland is used for false positive filtering of thyroid nodules: First, based on the various trajectories of the thyroid nodule, an initial detection result is preliminarily determined to indicate the presence of a thyroid nodule, i.e., the initial thyroid nodule detection result. Then, based on at least one of the following: the various trajectories of the carotid artery, the pixel-level segmentation masks of the carotid artery, the various trajectories of the thyroid gland, the pixel-level segmentation masks of the thyroid gland, and the pixel-level segmentation masks of the thyroid nodule, a false positive filtering strategy is applied to the initial thyroid nodule detection result, thereby obtaining the target thyroid nodule detection result. This improves the accuracy of the target thyroid nodule detection result.

[0105] For example, based on the relationship between the thyroid nodule trajectory and the thyroid artery trajectory and the carotid artery trajectory, if the mask corresponding to the thyroid nodule trajectory is not within the mask of any segment of the thyroid trajectory, the thyroid nodule trajectory needs to be filtered out, that is, false positive filtering is performed on the initial thyroid nodule detection results; in addition, if the mask corresponding to the thyroid nodule trajectory overlaps significantly with the mask of a segment of the carotid artery trajectory, the thyroid nodule trajectory also needs to be filtered out, that is, false positive filtering is performed on the initial thyroid nodule detection results.

[0106] The thyroid nodule detection method provided by the present invention is described below. The thyroid nodule detection method described below can be referred to in correspondence with the thyroid nodule detection system described above.

[0107] Figure 3 This is a schematic flowchart of the thyroid nodule detection method provided by the present invention, as shown below. Figure 3 As shown, the method includes the following: Step 301: Receive thyroid ultrasound video, which includes multiple video frames; Step 302: Using the trained target instance segmentation model, the target objects in each video frame are detected frame by frame to obtain the detection data of each target object in each video frame. The detection data includes detection boxes and confidence scores. Each target object includes carotid arteries, thyroid glands and thyroid nodules. Step 303: Based on the detection data, a multi-target tracking algorithm based on target detection is used to track each target object in the thyroid ultrasound video to obtain each tracking trajectory of each target object; Step 304: Based on the tracking trajectories of each target object, determine the detection result of the target thyroid nodule corresponding to the thyroid ultrasound video.

[0108] The thyroid nodule detection method provided by this invention involves receiving a thyroid ultrasound video, which includes multiple video frames. A trained target instance segmentation model is used to detect target objects in each video frame frame by frame, obtaining detection data for each target object in each video frame. The detection data includes bounding boxes and confidence scores. The target objects include carotid arteries, thyroid glands, and thyroid nodules. Based on the detection data, a multi-target tracking algorithm based on target detection is used to track each target object in the thyroid ultrasound video, obtaining tracking trajectories for each target object. Based on the tracking trajectories of each target object, the detection result of the target thyroid nodule corresponding to the thyroid ultrasound video is determined. This invention can significantly improve screening efficiency, greatly enhance automation, and achieve end-to-end automated processing from raw video to thyroid nodule detection results. It supports batch processing and is suitable for large-scale screening scenarios.

[0109] Optionally, before determining the detection result of the target thyroid nodule corresponding to the thyroid ultrasound video based on the tracking trajectories of each of the target objects, the method further includes: For each target object, a quality assessment is performed on each tracking trajectory of the target object to obtain a target quality assessment result for each tracking trajectory of the target object; The tracking trajectories whose target quality assessment results are not satisfactory are filtered out.

[0110] Optionally, the quality assessment includes a first quality assessment; The step of performing a quality assessment on each tracking trajectory of each target object to obtain a target quality assessment result for each tracking trajectory of the target object includes: For each target object, a first quality assessment is performed on each of the tracking trajectories of the target object to obtain a first quality assessment result for each of the tracking trajectories of the target object. The first quality assessment index includes at least one of trajectory length, continuity, and motion smoothness. Based on the first quality assessment result of each tracking trajectory of the target object, the target quality assessment result of each tracking trajectory of the target object is determined.

[0111] Optionally, the first quality assessment of each tracking trajectory of the target object to obtain the first quality assessment result of each tracking trajectory of the target object includes: For each tracking trajectory of the target object, the number of video frames representing the trajectory length corresponding to the tracking trajectory is obtained, and the trajectory length evaluation result of the tracking trajectory is determined based on the number of video frames. Obtain the number of interval frames representing the continuity of the tracking trajectory, and determine the continuity evaluation result of the tracking trajectory based on the number of interval frames; Obtain the continuous rate of change representing the motion smoothness corresponding to the tracking trajectory, and determine the motion smoothness evaluation result of the tracking trajectory based on the continuous rate of change. The continuous rate of change is the rate of change of the overlap of the detection boxes and / or the rate of change of the position of the detection boxes in consecutive video frames. The first quality assessment result of the tracking trajectory is determined based on at least one of the trajectory length assessment result, the continuity assessment result, and the motion smoothness assessment result.

[0112] Optionally, the quality assessment includes a second quality assessment; The step of performing a quality assessment on each tracking trajectory of each target object to obtain a target quality assessment result for each tracking trajectory of the target object includes: For each target object, obtain the confidence sequence corresponding to each tracking trajectory of the target object, the confidence sequence containing the confidence of all detection boxes corresponding to the tracking trajectory; Based on the confidence sequence corresponding to each of the tracking trajectories of the target object, a second quality assessment is performed on each of the tracking trajectories of the target object to obtain the second quality assessment result of each of the tracking trajectories of the target object. The indicators of the second quality assessment include at least one of the confidence mean, confidence minimum and confidence volatility. Based on the second quality assessment results of each tracking trajectory of the target object, the target quality assessment results of each tracking trajectory of the target object are determined.

[0113] Optionally, obtaining the target quality assessment results for each of the tracking trajectories of the target object includes: Based on the first quality assessment result and / or the second quality assessment result, determine the confidence score of each of the tracking trajectories of the target object; The target quality assessment result of the tracking trajectory whose confidence score is less than the scoring threshold is determined to be unsuccessful, and the target quality assessment result of the tracking trajectory whose confidence score is greater than or equal to the scoring threshold is determined to be successful.

[0114] Optionally, based on the detection data, a multi-target tracking algorithm based on target detection is used to track each target object in the thyroid ultrasound video to obtain each tracking trajectory of each target object, including: For each of the target objects, perform the following steps: For each video frame, the position of each existing trajectory of the target object in the video frame is predicted based on Kalman filtering, and the target object's tracking trajectory is obtained by two associations based on the detection data corresponding to the video frame and the position of each existing trajectory in the video frame.

[0115] Optionally, the association based on the detection data corresponding to the video frame and the position of each of the existing trajectories in the video frame is performed twice, including: Based on the confidence level of the target object in the video frame, determine the detection box type of the target object in the video frame; When the detection box type is a high-confidence detection box, the high-confidence detection box is associated with each of the existing trajectories based on the position of each of the existing trajectories in the video frame; When the detection box type is a low confidence detection box, the low confidence detection box is associated with each of the remaining unmatched trajectories in the video frame based on the position of each of the existing trajectories. If the detection box type is a high-confidence detection box and the association is unsuccessful, a new trajectory is created for the target object based on the high-confidence detection box.

[0116] Optionally, determining the detection box type of the target object in the video frame based on the confidence level of the target object in the video frame includes: If the confidence level of the target object in the video frame is greater than or equal to the high confidence threshold, the detection box type of the target object in the video frame is determined to be the high confidence detection box. If the confidence level of the target object in the video frame is less than the high confidence threshold and greater than or equal to the low confidence threshold, the detection box type of the target object in the video frame is determined to be the low confidence detection box.

[0117] Optionally, the training process of the target instance segmentation model includes: Acquire multiple initial images containing a specified object, wherein the specified object is at least one of a carotid artery, a thyroid gland, and a thyroid nodule; Each specified object in each of the initial images is labeled, and data augmentation is performed on each of the initial images to obtain a training image set; Based on the training image set, the initial instance segmentation model is trained to obtain the target instance segmentation model.

[0118] Optionally, the data augmentation of each of the initial images includes: Random brightness and contrast adjustments and / or Gaussian noise additions are performed on each of the initial images to achieve data augmentation, resulting in the training image set. The Gaussian noise addition includes simulated ultrasound artifacts.

[0119] Optionally, the detection data may further include a pixel-level segmentation mask; The determination of the target thyroid nodule detection result corresponding to the thyroid ultrasound video based on the tracking trajectory of each of the target objects includes: Based on the tracking trajectories of the thyroid nodules, the initial thyroid nodule detection results are determined; Based on at least one of the following: the tracking trajectory of the carotid artery, the pixel-level segmentation mask of the carotid artery, the tracking trajectory of the thyroid gland, the pixel-level segmentation mask of the thyroid gland, and the pixel-level segmentation mask of the thyroid nodule, the initial thyroid nodule detection result is subjected to false positive filtering to obtain the target thyroid nodule detection result.

[0120] Figure 4 This is a schematic diagram of the structure of the electronic device provided by the present invention, such as... Figure 4 As shown, the electronic device may include: a processor 410, a communication interface 420, a memory 430, and a communication bus 440, wherein the processor 410, the communication interface 420, and the memory 430 communicate with each other through the communication bus 440. The processor 410 can call logical instructions in the memory 430 to execute a thyroid nodule detection method, which includes: receiving a thyroid ultrasound video, the thyroid ultrasound video including multiple video frames; using a trained target instance segmentation model, performing frame-by-frame detection on target objects in each video frame to obtain detection data for each target object in each video frame, the detection data including detection boxes and confidence scores, and each target object including carotid arteries, thyroid glands, and thyroid nodules; based on each detection data, using a multi-target tracking algorithm based on target detection to track each target object in the thyroid ultrasound video to obtain each tracking trajectory of each target object; and based on each tracking trajectory of each target object, determining the target thyroid nodule detection result corresponding to the thyroid ultrasound video.

[0121] Furthermore, the logical instructions in the aforementioned memory 430 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, 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 personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0122] On the other hand, the present invention also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer is able to execute the thyroid nodule detection method provided by the above methods. The method includes: receiving a thyroid ultrasound video, the thyroid ultrasound video including multiple video frames; performing frame-by-frame detection on target objects in each video frame using a trained target instance segmentation model to obtain detection data for each target object in each video frame, the detection data including a detection box and confidence score, and each target object including a carotid artery, a thyroid gland, and a thyroid nodule; based on each detection data, using a multi-target tracking algorithm based on target detection to track each target object in the thyroid ultrasound video to obtain each tracking trajectory of each target object; and based on each tracking trajectory of each target object, determining the target thyroid nodule detection result corresponding to the thyroid ultrasound video.

[0123] In another aspect, the present invention also provides a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the thyroid nodule detection method provided by the above methods. The method includes: receiving a thyroid ultrasound video, the thyroid ultrasound video comprising multiple video frames; performing frame-by-frame detection of target objects in each video frame using a trained target instance segmentation model to obtain detection data for each target object in each video frame, the detection data including a detection box and confidence level, and each target object including a carotid artery, a thyroid gland, and a thyroid nodule; based on the detection data, employing a multi-target tracking algorithm based on target detection to track each target object in the thyroid ultrasound video to obtain tracking trajectories for each target object; and determining the target thyroid nodule detection result corresponding to the thyroid ultrasound video based on the tracking trajectories of each target object.

[0124] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0125] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0126] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A thyroid nodule detection system, characterized in that, include: A receiving module is configured to receive thyroid ultrasound video, the thyroid ultrasound video comprising multiple video frames; The multi-target instance segmentation module is configured to perform frame-by-frame detection of target objects in each video frame using a trained target instance segmentation model, and obtain detection data of each target object in each video frame. The detection data includes detection boxes and confidence scores. Each target object includes carotid arteries, thyroid glands and thyroid nodules. The instance tracking module is configured to track each target object in the thyroid ultrasound video based on the detection data and using a multi-target tracking algorithm based on target detection, thereby obtaining each tracking trajectory of each target object; The determination module is configured to determine the detection result of the target thyroid nodule corresponding to the thyroid ultrasound video based on each tracking trajectory of the target object.

2. The thyroid nodule detection system according to claim 1, characterized in that, It also includes a trajectory management module, which is configured as follows: For each target object, a quality assessment is performed on each tracking trajectory of the target object to obtain a target quality assessment result for each tracking trajectory of the target object; The tracking trajectories whose target quality assessment results are not satisfactory are filtered out.

3. The thyroid nodule detection system according to claim 2, characterized in that, The quality assessment includes a first quality assessment; The trajectory management module is specifically configured as follows: For each target object, the first quality assessment is performed on each of the tracking trajectories of the target object to obtain the first quality assessment result of each of the tracking trajectories of the target object. The first quality assessment index includes at least one of trajectory length, continuity and motion smoothness. Based on the first quality assessment result of each tracking trajectory of the target object, the target quality assessment result of each tracking trajectory of the target object is determined.

4. The thyroid nodule detection system according to claim 3, characterized in that, The trajectory management module is specifically configured as follows: For each tracking trajectory of the target object, the number of video frames representing the trajectory length corresponding to the tracking trajectory is obtained, and the trajectory length evaluation result of the tracking trajectory is determined based on the number of video frames. Obtain the number of interval frames representing the continuity of the tracking trajectory, and determine the continuity evaluation result of the tracking trajectory based on the number of interval frames; Obtain the continuous rate of change representing the motion smoothness corresponding to the tracking trajectory, and determine the motion smoothness evaluation result of the tracking trajectory based on the continuous rate of change. The continuous rate of change is the rate of change of the overlap of the detection boxes and / or the rate of change of the position of the detection boxes in consecutive video frames. The first quality assessment result of the tracking trajectory is determined based on at least one of the trajectory length assessment result, the continuity assessment result, and the motion smoothness assessment result.

5. The thyroid nodule detection system according to claim 2, characterized in that, The quality assessment includes a second quality assessment; The trajectory management module is specifically configured as follows: For each target object, obtain the confidence sequence corresponding to each tracking trajectory of the target object, the confidence sequence containing the confidence of all detection boxes corresponding to the tracking trajectory; Based on the confidence sequence corresponding to each of the tracking trajectories of the target object, a second quality assessment is performed on each of the tracking trajectories of the target object to obtain a second quality assessment result for each of the tracking trajectories of the target object. The indicators of the second quality assessment include at least one of the confidence mean, confidence minimum, and confidence volatility. Based on the second quality assessment results of each tracking trajectory of the target object, the target quality assessment results of each tracking trajectory of the target object are determined.

6. The thyroid nodule detection system according to claim 3, characterized in that, The trajectory management module is specifically configured as follows: Based on the first quality assessment result and / or the second quality assessment result, determine the confidence score of each of the tracking trajectories of the target object; The target quality assessment result of the tracking trajectory whose confidence score is less than the scoring threshold is determined to be unsuccessful, and the target quality assessment result of the tracking trajectory whose confidence score is greater than or equal to the scoring threshold is determined to be successful.

7. The thyroid nodule detection system according to claim 1, characterized in that, The instance tracking module is specifically configured as follows: For each of the target objects, perform the following steps: For each video frame, the position of each existing trajectory of the target object in the video frame is predicted based on Kalman filtering; Based on the confidence level of the target object in the video frame, determine the detection box type of the target object in the video frame; When the detection box type is a high-confidence detection box, the high-confidence detection box is associated with each of the existing trajectories based on the position of each of the existing trajectories in the video frame; When the detection box type is a low confidence detection box, the low confidence detection box is associated with each of the remaining unmatched trajectories in the video frame based on the position of each of the existing trajectories. If the detection box type is the high-confidence detection box and the association is unsuccessful, a new trajectory is created for the target object based on the high-confidence detection box.

8. The thyroid nodule detection system according to claim 1, characterized in that, It also includes a training module, configured as follows: Acquire multiple initial images containing a specified object, wherein the specified object is at least one of a carotid artery, a thyroid gland, and a thyroid nodule; Each specified object in each of the initial images is labeled, and data augmentation is performed on each of the initial images to obtain a training image set; Based on the training image set, the initial instance segmentation model is trained to obtain the target instance segmentation model.

9. The thyroid nodule detection system according to claim 8, characterized in that, The training module is specifically configured as follows: Random brightness and contrast adjustments and / or Gaussian noise additions are performed on each of the initial images to achieve data augmentation, resulting in the training image set. The Gaussian noise addition includes simulated ultrasound artifacts.

10. The thyroid nodule detection system according to any one of claims 1-9, characterized in that, The detection data also includes pixel-level segmentation masks; The determining module is specifically configured as follows: Based on the tracking trajectories of the thyroid nodules, the initial thyroid nodule detection results are determined; Based on at least one of the following: the tracking trajectory of the carotid artery, the pixel-level segmentation mask of the carotid artery, the tracking trajectory of the thyroid gland, the pixel-level segmentation mask of the thyroid gland, and the pixel-level segmentation mask of the thyroid nodule, the initial thyroid nodule detection result is subjected to false positive filtering to obtain the target thyroid nodule detection result.