Intelligent segmentation system for tumor CT images based on multimodal and deep learning

Abstract

This invention discloses an intelligent tumor CT image segmentation system based on multimodal and deep learning, belonging to the field of image segmentation technology. It includes a data collection module, a data labeling module, a data processing and adjustment module, a loop control module, and a database module. The data collection module collects tumor CT images from glioma patients. The data labeling module acquires x manually pixel-level labeled tumor CT images. The data processing and adjustment module acquires the target segmented CT image, correlation information, regional feature information, and feature statistics, and obtains feature weights, regional feature scores, and adjusted segmentation thresholds. The loop control module determines and intelligently executes the segmentation results. Based on multimodal data fusion and deep learning, this invention can determine the importance of each feature for target region recognition, comprehensively evaluate the matching degree between the region and the target region, and adapt to different image features to achieve accurate segmentation.

Description

Technical Field

[0001] This invention relates to the field of image segmentation technology, specifically to an intelligent segmentation system for tumor CT images based on multimodal and deep learning. Background Technology

[0002] In the field of tumor diagnosis, accurate CT image segmentation is crucial for tumor detection, localization, and treatment planning. With the development of medical imaging technology, multimodal data provides rich information for tumor analysis. At the same time, deep learning technology has shown great potential in image segmentation, as it can automatically learn image features and improve segmentation accuracy.

[0003] Currently, intelligent segmentation technology for tumor CT images that combines multimodal data fusion and deep learning has gradually become a research hotspot. However, in practical applications, the intelligent segmentation of most tumor regions is based on manual annotation by multiple experts. The annotations of different experts have certain differences. If only edge recognition is performed on the largest annotated area, the accuracy of image segmentation will inevitably be reduced. In addition, when extracting features from CT images, existing technologies often fail to fully consider the actual correlation between different features and different areas of the tumor. This results in some features that are crucial to tumor region identification not receiving the weight they deserve, while some irrelevant or weakly related features may occupy a high weight, affecting the accuracy of segmentation.

[0004] Secondly, most existing technologies rely on a single feature to assess whether an image region is a tumor-related region, which fails to fully reflect the degree of matching between the region and the tumor region, resulting in inaccurate segmentation results and a high risk of misjudgment and missed judgment.

[0005] Finally, existing technologies typically use a fixed segmentation threshold to distinguish between tumor and non-tumor regions without dynamically adjusting it based on the feature distribution of different regions in the image. This makes the segmentation threshold unable to adapt to the feature differences of different CT images, resulting in poor segmentation performance when faced with complex and varied tumor images. Summary of the Invention

[0006] The purpose of this invention is to provide an intelligent segmentation system for tumor CT images based on multimodal and deep learning, which solves the problems mentioned in the background art.

[0007] To achieve the above objectives, the present invention provides the following technical solution, including a data collection module, a data marking module, a data processing and adjustment module, a loop control module, and a database module;

[0008] The data collection module is used to collect CT images of glioma patients that meet the MNI criteria and have been converted to the MNI standard space, and transmit them to the data labeling module.

[0009] The data labeling module is used to acquire x manually labeled CT images of tumors at the pixel level for multimodal 3D images;

[0010] The pixel-level annotations include the background, edema area, necrotic area, and enhancement area;

[0011] Based on the diagnostic features preset by the system, the target segmented CT image containing target and non-target image regions is extracted, and the feature statistics and regional feature information in the target segmented CT image are obtained using the data collection module.

[0012] Based on the target segmented CT image, obtain the associated information;

[0013] The data processing and adjustment module is used to acquire the target segmented CT image, the correlation information, the regional feature information, and the feature statistical information;

[0014] Based on the association information and the feature statistics, the feature weights are obtained;

[0015] Based on the feature weights and the region feature information, the region feature score is obtained;

[0016] Based on the feature scores of all regions in the target segmented CT image, and the segmentation adjustment threshold and adjustment coefficient obtained from the database module, the adjusted segmentation threshold is obtained;

[0017] The loop control module is used to receive the adjusted segmentation threshold and the region feature score, and determine and intelligently execute the segmentation result based on the comparison result of the adjusted segmentation threshold and the region feature score.

[0018] Optionally, the data collection module may use equipment including CT scanners and clinical examination instruments;

[0019] The data labeling module uses equipment including image annotation software;

[0020] The data processing and adjustment module and the loop control module use devices including processors;

[0021] The database module uses devices including storage servers;

[0022] The database module stores and transmits data to the data collection module, the data tagging module, the data processing and adjustment module, and the loop control module through a data interface;

[0023] The purpose of database module design is not only to realize the storage and management of related data, but also to form a complete data model from the originally loose and independent data.

[0024] Optionally, the diagnostic features include inflammation diagnostic features and growth and metabolism diagnostic features;

[0025] The inflammatory diagnostic features are based on the inflammatory response of the tissue surrounding the tumor reflected in the edema area.

[0026] The growth and metabolism diagnostic features are based on the growth and metabolism of the tumor as reflected in the necrotic area.

[0027] Optionally, the feature statistics include the number of occurrences of a single feature and the total number of occurrences of the feature;

[0028] Based on the ratio of the occurrence frequency of the single feature to the total occurrence frequency of the features, the feature proportion of the single feature in the target segmented CT image is obtained;

[0029] In tumor CT image segmentation using multimodal data fusion, different features play different roles in distinguishing target and non-target image regions of the tumor. The feature weight of each feature, which can accurately quantify the importance of each feature in the identification of the target region, is obtained by multiplying the feature proportion obtained from the feature statistics with the association information.

[0030] Optionally, the regional feature information includes the average value of a single feature in the region and the feature weight of each feature;

[0031] Specifically as follows:

[0032] After matching the association information and feature statistics of each target with the same feature in the same region, the feature weight of each feature in each region is obtained.

[0033] Based on the segmented CT image of the target, obtain the regional feature quantity and multi-feature information of each target and non-target image region;

[0034] The multi-feature information within the region includes grayscale features, texture features, and shape features;

[0035] Based on the multi-feature information within the region, the value of the feature quantity within the region is 3;

[0036] After averaging the feature values ​​and multi-feature information within the region and matching each target and non-target image region, the average value of the single feature of each region is obtained.

[0037] In tumor CT images, whether a region is a tumor-related region requires comprehensive judgment based on multiple features. The average value of the single feature of the region, the feature quantity within the region, and the current feature weight of the region are obtained based on the feature weights and the region feature information. Finally, the region feature score under the multi-feature evaluation of each region is obtained to comprehensively evaluate the degree of matching between the region and the target region.

[0038] Optionally, based on the region feature scores of all regions obtained for each target and non-target image region, the maximum region feature score and the average region feature score are obtained.

[0039] The segmentation adjustment amount is obtained based on the maximum score of the region feature, the average score of the region feature, and the adjustment coefficient;

[0040] The segmentation adjustment threshold is added to the segmentation adjustment amount obtained based on the regional feature scores of all regions and the adjustment coefficient to finally obtain the adjusted segmentation threshold.

[0041] The adjustment coefficient ranges from 0 to 1.

[0042] The segmentation adjustment threshold is set to 0.5 during the initial segmentation, and after the initial segmentation, it is set to the previous adjusted segmentation threshold.

[0043] Optionally, the data processing and adjustment module includes a feature processing unit and a threshold adjustment unit;

[0044] The loop control module includes a segmentation determination unit, an image segmentation unit, and a loop update unit;

[0045] The feature processing unit is used to obtain the feature weights and the region feature scores.

[0046] The threshold adjustment unit is used to obtain the adjusted segmentation threshold;

[0047] The segmentation determination unit is used to obtain the comparison result and the segmentation result;

[0048] The image segmentation unit and the loop update unit are used to execute the segmentation results.

[0049] Optionally, the comparison results specifically include the following:

[0050] S1: The region feature score is greater than the adjusted segmentation threshold;

[0051] S2: The region feature score is less than or greater than the adjusted segmentation threshold;

[0052] The segmentation results based on S1 and S2 are as follows:

[0053] Based on S1, the current region is determined and segmented as the target image region;

[0054] Based on S2, the current region is determined and segmented as a non-target image region;

[0055] The loop control module stops looping and segmenting if the segmentation results are the same in two consecutive steps.

[0056] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0057] This invention fully considers the feature proportion in feature statistics and the correlation information with the target region. In practical applications, for common features in tumor CT images that are closely related to the tumor region, by statistically analyzing the occurrence frequency of single features and the total occurrence frequency of features, and combining the correlation information determined by machine learning algorithms, the feature weight of each feature in each region can be scientifically allocated. Moreover, it can more accurately highlight features that are important for tumor region identification and avoid interference from irrelevant or weakly related features, thereby improving the accuracy of image segmentation.

[0058] When processing tumor CT images, region feature scoring evaluates each region from multiple feature dimensions, including grayscale features, texture features, and shape features. This comprehensive evaluation method can more accurately reflect the degree of matching between the region and the tumor region, reduce misjudgment and omission, and make the segmentation results more consistent with the actual tumor region distribution.

[0059] Finally, the system dynamically adjusts the segmentation threshold based on the maximum and average regional feature scores of all regions and the adjustment coefficient. This allows the system to automatically adjust the segmentation threshold according to the distribution of regional scores in the image when faced with tumor CT images of different patients. This dynamic adjustment mechanism enables the system to adapt to various complex and variable tumor CT images, improving the stability and reliability of segmentation. Attached Figure Description

[0060] Figure 1 This is a flowchart of the overall modules of the intelligent segmentation system for tumor CT images.

[0061] Figure 2 This is a schematic diagram of the adjustment and acquisition of the data processing and adjustment module of the present invention. Detailed Implementation

[0062] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0063] This intelligent segmentation system for tumor CT images differs from existing intelligent segmentation systems for tumor CT images.

[0064] Existing intelligent segmentation systems for tumor CT images suffer from problems such as unreasonable feature extraction and weight allocation, single and inaccurate region assessment, and lack of dynamic adjustment of segmentation threshold settings. In contrast, this algorithm unit can perform scientific feature weight calculation, comprehensive and accurate region assessment, and dynamic adaptive segmentation threshold adjustment.

[0065] Example 1, please refer to Figures 1 to 2 This implementation provides an intelligent segmentation system for tumor CT images based on multimodal and deep learning, including a data collection module, a data labeling module, a data processing and adjustment module, a loop control module, and a database module;

[0066] The data collection module is used to collect CT images of glioma patients' tumors and transmit them to the data labeling module;

[0067] The data labeling module is used to acquire x manually labeled CT images of tumors at the pixel level;

[0068] The pixel-level annotations include the background, edema area, necrotic area, and enhancement area;

[0069] Based on the diagnostic settings preset by the system, the target segmented CT image containing the target and non-target image regions is extracted, and the feature statistics and regional feature information in the target segmented CT image are obtained by the data collection module.

[0070] Diagnostic features include inflammation diagnostic features and growth and metabolism diagnostic features;

[0071] The diagnostic features of inflammation are the inflammatory response of the tissue surrounding the tumor as reflected in the edematous area.

[0072] Growth and metabolism diagnostic features use the growth and metabolism of the tumor as reflected in the necrotic area as the target features for diagnosis.

[0073] Based on the target segmentation of the CT image, obtain the associated information;

[0074] The data processing and adjustment module is used to acquire target segmentation CT images, correlation information, regional feature information, and feature statistics.

[0075] Based on the correlation information and feature statistics, obtain the feature weights;

[0076] Based on feature weights and regional feature information, obtain regional feature scores;

[0077] Based on the feature scores of all regions in the target segmented CT image, and the segmentation adjustment threshold and adjustment coefficient obtained from the database module, the adjusted segmentation threshold is obtained;

[0078] The loop control module is used to receive the adjusted segmentation threshold and the region feature score, and to determine and intelligently execute the segmentation result based on the comparison result between the adjusted segmentation threshold and the region feature score.

[0079] The data processing and adjustment module includes a feature processing unit and a threshold adjustment unit;

[0080] The loop control module includes a segmentation determination unit, an image segmentation unit, and a loop update unit;

[0081] The feature processing unit is used to obtain feature weights and region feature scores.

[0082] The threshold adjustment unit is used to obtain the adjusted segmentation threshold;

[0083] The segmentation determination unit is used to obtain the comparison results and segmentation results;

[0084] The image segmentation unit and the loop update unit are used to execute the segmentation results.

[0085] The equipment used in the data collection module includes CT scanners and clinical examination instruments;

[0086] The data labeling module uses equipment including image annotation software;

[0087] The data processing and adjustment module and the loop control module use devices including processors;

[0088] The database module uses devices including storage servers;

[0089] The database module stores and transmits data to the data collection module, data tagging module, data processing and adjustment module, and loop control module through the data interface.

[0090] In this embodiment, for the annotation of multimodal 3D images, especially pixel-level annotation, the system needs to invite x, that is, more than 15 experts with more than three years of clinical experience to perform manual annotation. Manual annotation can use the experts' professional knowledge and clinical experience to accurately mark the category to which each pixel in the image belongs, which provides reliable basic data for subsequent intelligent processing.

[0091] Next, the overlapping areas of all manually annotated regions are extracted. Since there are certain differences in the annotations of different experts, the overlapping areas represent a relatively consistent judgment. This is used as the basis for intelligent processing, which can reduce the error caused by the difference in annotation.

[0092] Next, based on the target segmented CT image of the extracted overlapping region, the following are obtained in sequence: the feature weight that accurately reflects the importance of each feature in distinguishing the target region; the region feature score that makes the region score more realistically reflect the degree of matching between the region and the target region; and the region feature score that can better adapt to the image segmentation requirements based on the features of the overlapping region.

[0093] In summary, this approach of first manually labeling and then intelligently processing overlapping regions fully leverages the professional judgment of humans and the data processing advantages of intelligent algorithms. Manual labeling provides a professional and accurate data foundation, while intelligent processing utilizes algorithms to efficiently analyze and calculate large amounts of data, achieving automation and intelligence in image segmentation. Moreover, accurate image segmentation is a crucial step in developing intelligent diagnostic products for gliomas. By integrating manual and intelligent processing, the accuracy of image segmentation can be improved, providing high-quality data support for the subsequent development of intelligent diagnostic products.

[0094] Please see Figures 1 to 2 The feature statistics include the number of occurrences of a single feature and the total number of occurrences of the feature;

[0095] Based on the ratio of the occurrence frequency of a single feature to the total occurrence frequency of all features, the feature proportion of a single feature in the target segmentation CT image is obtained;

[0096] The feature weight of each feature is obtained by multiplying the feature proportion obtained from the feature statistics information with the correlation information.

[0097] In this embodiment, the formula for calculating the feature weights is as follows:

[0098] ;

[0099] in:

[0100] Q i For feature weights;

[0101] T i T represents the number of occurrences of the i-th feature in the total features. i This reflects the frequency of the feature in the target segmentation CT image. Features that appear more frequently occupy a more important position in the overall features and may also have a closer relationship with the target region.

[0102] ZT represents the total number of occurrences of the feature;

[0103] G i For related information, logistic regression models can be used to directly and intelligently obtain it;

[0104] The calculation result is the feature proportion. The higher the feature proportion value, the more common the feature is in the overall features.

[0105] It can effectively integrate feature information from different modalities of data, specifically features extracted from multi-dimensional information from clinical data, imaging data, and pathology data, all of which can be achieved through... Calculate its feature weights Q i This allows each feature to find a suitable location in the overall segmentation algorithm based on multimodal data fusion, laying the foundation for accurate segmentation based on multimodal data in the future.

[0106] Please see Figures 1 to 2 The regional feature information includes the average value of a single feature in the region and the feature weight of each feature;

[0107] Specifically as follows:

[0108] After matching the association information and feature statistics of each target with the same feature in the same region, the feature weight of each feature in each region is obtained.

[0109] Based on the segmented CT image of the target, the regional feature quantity and multi-feature information of each target and non-target image region are obtained;

[0110] The region contains multiple features, including grayscale features, texture features, and shape features;

[0111] Based on multiple feature information within the region, the value of the feature quantity within the region is 3;

[0112] After averaging the feature values ​​and multi-feature information within the region and matching each target and non-target image region, the average value of the single feature of each region is obtained.

[0113] Based on the average value of a single feature in the region, the number of features within the region, and the current feature weight of the region obtained from the feature weights and the region feature information, the region feature score of each region is finally obtained.

[0114] In this embodiment, the formula for calculating the region feature score is as follows:

[0115] ;

[0116] in:

[0117] QY i The regional feature score is calculated by summing the contribution values ​​of all features to the score of the j-th region.

[0118] n is a characteristic quantity within the region, Q ij H represents the feature weight of a single feature in the j-th region. ij Let be the average value of a single feature across j regions.

[0119] In tumor CT images, determining whether a region belongs to a tumor-related area cannot be based on a single feature alone; rather, it requires comprehensive consideration of multiple features. In this embodiment, not only grayscale features are considered, but also texture and shape features are combined, and the feature weights Q of each feature are used to determine this. i By weighted summation of multiple features, the matching degree between the region and the target region is comprehensively evaluated, which greatly improves the accuracy of the judgment compared with the evaluation method that relies on only a single feature.

[0120] Please see Figures 1 to 2 Based on the region feature scores of all regions obtained for each target and non-target image region, the maximum region feature score and the average region feature score are obtained.

[0121] The segmentation adjustment amount is obtained based on the maximum score of the regional feature, the average score of the regional feature, and the adjustment coefficient.

[0122] The segmentation adjustment threshold is obtained by adding the segmentation adjustment amount obtained based on the regional feature scores and adjustment coefficients of all regions.

[0123] The adjustment coefficient ranges from 0 to 1.

[0124] The threshold for segmentation adjustment is set to 0.5 during the initial segmentation, and after the initial segmentation, it is set to the previous adjusted segmentation threshold.

[0125] In this embodiment, the calculation formula for the post-segmentation threshold is first adjusted as follows:

[0126] ;

[0127] in:

[0128] FT new To adjust the segmentation threshold, FT old To adjust the segmentation based on the threshold, QY avg QY is the average score of regional features. max The maximum score for the regional feature is given by k, which is an adjustment coefficient.

[0129] The calculation result is the segmentation adjustment amount;

[0130] By adjusting the segmentation threshold FT new The system can obtain the average score QY based on the regional features of the current image. avg Maximum score QY of regional features max The segmentation threshold is automatically adjusted based on the scores of the regions formed.

[0131] For images with obvious tumor features and large differences in regional scores, the adjusted threshold can more accurately distinguish between tumor and non-tumor regions.

[0132] For images with more complex features and smaller differences in regional scores, the threshold can be adaptively adjusted to improve the accuracy and adaptability of segmentation.

[0133] It is worth noting that in intelligent segmentation of tumor CT images, the feature distribution of the data and the stability of the segmentation results are important considerations. A smaller adjustment coefficient k, i.e., k close to 0, will result in a smaller adjustment range for the segmentation threshold. This allows for the calculation of the average score QY based on regional features while ensuring relatively stable segmentation results. avg Maximum score QY of regional features max The scores of the regions formed are fine-tuned to avoid damaging the existing segmentation effect due to excessive adjustments;

[0134] A larger adjustment coefficient k, i.e. k close to 1, will increase the adjustment range, but too large a value will lead to excessive fluctuations in the segmentation results, which is not conducive to obtaining stable and accurate segmentation results. Therefore, a value of k between 0 and 1 can find a balance between the adaptability of data features and the stability of segmentation.

[0135] Example 2, please refer to Figures 1 to 2 The comparison results include the following:

[0136] S1: The region feature score is greater than the adjusted segmentation threshold;

[0137] S2: Region feature scores are less than or greater than the adjusted segmentation threshold;

[0138] The segmentation results based on S1 and S2 are as follows:

[0139] Based on S1, the current region is determined and segmented as the target image region;

[0140] Based on S2, the current region is determined and segmented as a non-target image region;

[0141] The loop control module stops looping and segmenting if the results of two consecutive segmentations are the same.

[0142] In this embodiment, tumors are highly heterogeneous, and the characteristics of tumors vary between different patients and even between different parts of the same patient. The cyclical influence of the loop control module enables the system to continuously adapt to this heterogeneity in multiple iterations. As the segmentation threshold is adjusted, the feature weights also change accordingly, thereby enabling the algorithm to better capture various complex features of tumors, improve the segmentation ability of different tumor types and morphologies, and enhance the generalization and adaptability of the algorithm.

[0143] This cyclical influence mechanism makes the entire segmentation system adaptive and intelligent. The system can dynamically adjust feature weights based on each segmentation result without manual intervention. When faced with a large number of different tumor CT images, the system can automatically learn the changing patterns of image features, continuously optimize its segmentation strategy, and improve segmentation efficiency and accuracy.

[0144] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A tumor CT image intelligent segmentation system based on multimodal and deep learning, characterized in that, It includes a data collection module, a data labeling module, a data processing and adjustment module, a loop control module, and a database module; The data collection module is used to collect CT images of glioma patients' tumors and transmit them to the data labeling module; The data labeling module is used to acquire x manually labeled CT images of tumors at the pixel level; The pixel-level annotations include the background, edema area, necrotic area, and enhancement area; Based on the diagnostic features preset by the system, the target segmented CT image containing target and non-target image regions is extracted, and the feature statistics and regional feature information in the target segmented CT image are obtained using the data collection module. Based on the target segmented CT image, obtain the associated information; The data processing and adjustment module is used to acquire the target segmented CT image, the correlation information, the regional feature information, and the feature statistical information; Based on the association information and the feature statistics, the feature weights are obtained; Based on the feature weights and the region feature information, the region feature score is obtained; Based on the feature scores of all regions in the target segmented CT image, and the segmentation adjustment threshold and adjustment coefficient obtained from the database module, the adjusted segmentation threshold is obtained; The loop control module is used to receive the adjusted segmentation threshold and the region feature score, and determine and intelligently execute the segmentation result based on the comparison result of the adjusted segmentation threshold and the region feature score. The diagnostic features include inflammation diagnostic features and growth and metabolism diagnostic features; The inflammatory diagnostic features are based on the edema area as the target feature for diagnosis. The growth and metabolic diagnostic features use the necrotic area as the target feature for diagnosis. The feature statistics include the number of occurrences of a single feature and the total number of occurrences of the feature; The feature percentage of a single feature is obtained based on the ratio between the number of occurrences of a single feature and the total number of occurrences of all features. The feature weight of each feature is obtained by multiplying the feature proportion obtained from the feature statistics with the association information. The regional feature information includes the average value of a single feature in the region and the feature weight of each feature; Specifically as follows: After matching the association information and feature statistics of each target with the same feature in the same region, the feature weight of each feature in each region is obtained. Based on the segmented CT image of the target, obtain the regional feature quantity and multi-feature information of each target and non-target image region; The multi-feature information within the region includes grayscale features, texture features, and shape features; Based on the multi-feature information within the region, the value of the feature quantity within the region is 3; After averaging the feature values ​​within the region and the region features matched for each target and non-target image region, the average value of the single feature of each region is obtained. Based on the average value of a single feature in the region, the number of features within the region, and the current feature weight of the region obtained from the feature weights and the region feature information, the region feature score of each region is finally obtained.

2. The intelligent segmentation system for tumor CT images based on multimodal and deep learning according to claim 1, characterized in that, The data collection module uses equipment including CT scanners and clinical examination instruments; The data labeling module uses equipment including image annotation software; The data processing and adjustment module and the loop control module use devices including processors; The database module uses devices including storage servers; The database module stores and transmits data to the data collection module, the data tagging module, the data processing and adjustment module, and the loop control module through a data interface.

3. The intelligent segmentation system for tumor CT images based on multimodal and deep learning according to claim 1, characterized in that: Based on the region feature scores of all regions obtained for each target and non-target image region, the maximum region feature score and the average region feature score are obtained. The segmentation adjustment amount is obtained based on the maximum score of the region feature, the average score of the region feature, and the adjustment coefficient; The segmentation adjustment threshold is added to the segmentation adjustment amount obtained based on the regional feature scores of all regions and the adjustment coefficient to finally obtain the adjusted segmentation threshold. The adjustment coefficient ranges from 0 to 1. The segmentation adjustment threshold is set to 0.5 during the initial segmentation, and after the initial segmentation, it is set to the previous adjusted segmentation threshold.

4. The intelligent segmentation system for tumor CT images based on multimodal and deep learning according to claim 3, characterized in that: The data processing and adjustment module includes a feature processing unit and a threshold adjustment unit; The loop control module includes a segmentation determination unit, an image segmentation unit, and a loop update unit; The feature processing unit is used to obtain the feature weights and the region feature scores. The threshold adjustment unit is used to obtain the adjusted segmentation threshold; The segmentation determination unit is used to obtain the comparison result and the segmentation result; The image segmentation unit and the loop update unit are used to execute the segmentation results.

5. The intelligent segmentation system for tumor CT images based on multimodal and deep learning according to claim 4, characterized in that: The comparison results specifically include the following: S1: The region feature score is greater than the adjusted segmentation threshold; S2: The region feature score is less than or greater than the adjusted segmentation threshold; The segmentation results based on S1 and S2 are as follows: Based on S1, the current region is determined and segmented as the target image region; Based on S2, the current region is determined and segmented as a non-target image region; The loop control module stops looping and segmenting if the segmentation results are the same in two consecutive steps.

Images