SPECT thyroid gland automatic analysis method and device and computer program

Abstract

The invention provides an SPECT thyroid gland automatic analysis method. The method comprises the following steps: segmenting a thyroid gland and a thyroid gland nodule in an SPECT image; classifying the properties of the thyroid nodules obtained by segmentation; measuring the size of the thyroid obtained by segmentation; classifying the intake functions of the thyroid gland obtained by segmentation; and generating an analysis report according to the property classification result, the ingestion function classification result and the measurement result. According to the automatic thyroid analysis method provided by the invention, thyroid nodule function analysis, thyroid nodule detection, thyroid nodule position and classification, thyroid overall function analysis, automatic left leaf and right leaf thyroid size measurement and automatic ingestion function judgment can be provided. The method can basically meet the requirement of the whole process of SPECT thyroid analysis clinically, the function of intelligent auxiliary diagnosis is effectively achieved, the manual operation cost of doctors is greatly reduced, and good segmentation and classification performance is achieved.

Description

Technical Field

[0001] This specification relates to the field of medical image recognition technology, and in particular to a SPECT thyroid method, device, and program. Background Art

[0002] Single-photon emission computed tomography (SPECT), a key imaging technique in nuclear medicine, plays a vital role in the diagnosis of thyroid nodules. SPECT images categorize thyroid nodules as hot, warm, cold, and cool based on differences in their radioactivity, providing a powerful basis for clinical assessment of nodule characteristics. Machine learning techniques can assist physicians in making judgments based on SPECT thyroid images.

[0003] The existing technology with application number CN202110978848 provides a method for classifying the properties of thyroid nodules based on SPECT images. This method adopts difficult sample mining technology. Although it only uses very small sample data, reducing the cost of obtaining sample data, the recall rate of its classification results is not ideal, and the scheme can only classify the entire thyroid SPECT image. If there are multiple types of nodules in the same thyroid, it cannot be handled. The scheme also lacks quantitative and qualitative analysis of the thyroid gland, and is still highly dependent on the doctor's experience and operation.

[0004] Based on this, how to make full use of SECT thyroid images with very few training samples to automatically detect thyroid nodules and perform quantitative and qualitative analysis is an urgent problem to be solved. Summary of the Invention

[0005] In view of the shortcomings of the existing technology in the detection of thyroid nodules, the present invention proposes a SPECT thyroid automatic analysis method, device and computer program.

[0006] To this end, one aspect of the present invention provides a SPECT thyroid automatic analysis method, comprising:

[0007] Segmentation of the thyroid gland and thyroid nodules in SPECT images;

[0008] Classify the properties of the segmented thyroid nodules;

[0009] Measure the size of the segmented thyroid gland;

[0010] Classify the uptake function of the segmented thyroid gland;

[0011] Generate an analysis report based on the property classification results, uptake function classification results and measurement results.

[0012] Another aspect of the present invention provides a SPECT thyroid automatic analysis device, comprising:

[0013] The target segmentation module is used to segment the thyroid gland and thyroid nodules in SPECT images;

[0014] Property classification module, used to classify the properties of the segmented thyroid nodules;

[0015] A size measurement module, used to measure the size of the segmented thyroid gland;

[0016] A functional classification module is used to classify the uptake function of the segmented thyroid gland;

[0017] The report output module is used to generate analysis reports based on property classification results, uptake function classification results and measurement results.

[0018] Another aspect of the present invention further provides a computer program for automatic SPECT thyroid analysis, which, when executed by a processor, enables the processor to implement the various methods described above.

[0019] The automated thyroid analysis method based on SPECT images provided by this invention employs a segmentation model to segment the thyroid gland and thyroid nodules. The pre-trained CLIP model is incorporated into the segmentation model to overcome the limitations of SPECT thyroid data, which makes it difficult to train a generalizable model, thereby improving segmentation performance. This method can automatically classify thyroid uptake function. In the classification module, it innovatively compares and analyzes thyroid nodule segmentation results with normal thyroid regions and background regions, rather than analyzing the entire thyroid image. This significantly simplifies the task complexity. This method can also automatically measure thyroid size, effectively avoiding the labor costs and manual errors associated with manual operation.

[0020] Experimental data verify that the performance of each module provided by the present invention is excellent: the DICE index for left lobe thyroid segmentation is 0.9322, the DICE index for right lobe thyroid segmentation is 0.9289, and the DICE index for thyroid nodule segmentation is 0.8636; the DICE index for hot nodule segmentation is 0.9278, the detection sensitivity is 0.9231, and the specificity is 0.9958; the DICE index for warm nodule segmentation is 0.8091, the detection sensitivity is 0.800, and the specificity is 1.000; the DICE index for cool nodule segmentation is 0.8582, the detection sensitivity is 0.9385, and the specificity is 0.9952; the DICE index for cold nodule segmentation is 0.8679, the detection sensitivity is 0.9375, and the specificity is 0.9978. The thyroid uptake function classification model had a precision of 0.94 and a recall of 0.97 for the "increased uptake" classification indicator; a precision of 0.93 and a recall of 0.95 for the "normal uptake" classification indicator; and a precision of 0.95 and a recall of 1.00 for the "decreased uptake" classification indicator. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Other features, objects and advantages of the present invention will become more apparent from the following detailed description of non-limiting embodiments taken in conjunction with the accompanying drawings.

[0022] Figure 1 A schematic diagram of a system architecture for applying the method of the embodiment of the present disclosure;

[0023] Figure 2 A flow chart of the SPECT thyroid automatic analysis method provided in an embodiment of the present disclosure;

[0024] Figure 3 Schematic diagram of the model structure used in the automatic thyroid analysis method according to the embodiment of the present disclosure;

[0025] Figure 4 This is an example of labeling the contours of samples for training the segmentation model in the embodiment of the present disclosure;

[0026] Figure 5 Schematic diagram of the structure of the segmentation model in the embodiment of the present disclosure;

[0027] Figure 6 is a schematic diagram of measuring the size of the thyroid gland in an embodiment of the present disclosure;

[0028] Figure 7 This is a schematic diagram of determining thyroid uptake function in an embodiment of the present disclosure;

[0029] Figure 8 A schematic diagram of a SPECT thyroid automatic analysis device provided in an embodiment of the present disclosure;

[0030] Figure 9The network structure of the thyroid nodule property classification model provided by the existing technology. DETAILED DESCRIPTION

[0031] In the specification of the present disclosure, it should be understood that terms such as "include" or "have" are intended to indicate the presence of features, numbers, steps, behaviors, components, parts, or a combination thereof disclosed in the specification, and are not intended to exclude the possibility that one or more other features, numbers, steps, behaviors, components, parts, or a combination thereof exist or are added.

[0032] Before describing the embodiments of the technical solution disclosed herein, a brief analysis of the solution of the prior art (CN202110978848) is performed.

[0033] The network structure of the thyroid nodule property classification model provided by the prior art is as follows: Figure 9 As shown in the figure, the proposed method includes a classification convolutional neural network and a twin neural network. The difficult sample mining technique is used to train the thyroid nodule classification model. The trained model can output thyroid nodule classification results. The main shortcomings of this solution are as follows.

[0034] a) Classifying the entire thyroid SPECT image. This design cannot handle the situation where multiple types of nodules exist in the same thyroid gland. For example, if a thyroid gland contains both cold and hot nodules, the model cannot correctly classify them.

[0035] b) Only thyroid glands with known nodules were classified without considering the distinction between thyroid glands without nodules;

[0036] c) The method of using a classification convolutional neural network to implement nodule classification may lead to incorrect nodule classification due to different proportions of nodule types. In Example 2 of its specification, the distribution of the training data used is: 30 cold nodules, 199 cool nodules, 10 warm nodules, and 34 hot nodules. Although it uses a difficult sample mining method, the recall rates in Table 1 of the cold nodule experimental results and Table 4 of the hot nodule experimental results are 54.17% and 51.67% respectively, which are still much lower than the recall rate of 98.75% in Table 2 of the cool nodule experimental results with more data samples, indicating that this solution fails to solve the nodule classification problem well and is still limited by the sample distribution of the training data.

[0037] Based on this, the present disclosure proposes an improved automated SPECT thyroid analysis solution, using deep learning methods to perform instance-based segmentation of SPECT thyroid and thyroid nodules. Through a unique network structure, a qualified classification model is trained with minimal training samples, saving costs, improving efficiency, and achieving higher-accuracy automated classification. It should be understood that the above-mentioned prior art is merely illustrative, and the present disclosure's method also offers corresponding improvements over other similar prior art.

[0038] Hereinafter, exemplary embodiments of the present disclosure will be described in detail with reference to the accompanying drawings to facilitate implementation by those skilled in the art. Furthermore, for clarity, portions not relevant to the description of the exemplary embodiments are omitted from the accompanying drawings. It should be noted that the embodiments and features of the embodiments of the present disclosure may be combined with one another unless they conflict.

[0039] Figure 1 A schematic diagram of a system architecture for applying the method of an embodiment of the present disclosure is shown schematically.

[0040] like Figure 1 As shown, the system architecture may include terminal devices 101, 102, 103, a network 104, and a server 105. The network 104 is used to provide a medium for communication links between the terminal devices 101, 102, 103 and the server 105. The network 104 may include various connection types, such as wired or wireless communication links or fiber optic cables.

[0041] Terminal devices 101, 102, and 103 interact with server 105 via network 104 to receive or send messages, etc. Various client applications may be installed on terminal devices 101, 102, and 103, such as dedicated applications with functions such as medical image processing, uploading and displaying diagnostic results.

[0042] Terminal devices 101, 102, and 103 can be either hardware or software. When hardware is used, terminal devices 101, 102, and 103 can be various specialized or general-purpose electronic devices, including but not limited to CT machines, smart terminals, portable computers, and desktop computers. When software is used, terminal devices 101, 102, and 103 can be installed in the electronic devices listed above. They can be implemented as multiple software programs or software modules (e.g., multiple software programs or software modules used to provide distributed services), or as a single software program or software module.

[0043] The server 105 may be a server that provides various intelligent services, such as a backend server that provides services for client applications installed on the terminal devices 101, 102, and 103. For example, the server may train and deploy a SPECT bone scan image recognition model to support an intelligent diagnostic system so that visualization results can be displayed on the terminal devices 101, 102, and 103.

[0044] The server 105 may be hardware or software. When the server 105 is hardware, it may be implemented as a distributed server cluster consisting of multiple servers, or as a single server. When the server 105 is software, it may be implemented as multiple software programs or software modules (e.g., multiple software programs or software modules for providing distributed services), or as a single software program or software module.

[0045] The various methods provided in the embodiments of the present disclosure may be executed by the server 105 or by the terminal devices 101, 102, 103. Alternatively, the various methods provided in the embodiments of the present disclosure may be partially executed by the terminal devices 101, 102, 103 and the rest executed by the server 105.

[0046] It should be understood that Figure 1 The number of terminal devices, networks and servers in the embodiment is merely illustrative. Any number of terminal devices, networks and servers may be provided as required.

[0047] The following combination Figure 2 An embodiment of the SPECT thyroid automatic analysis method provided by the present disclosure is described.

[0048] Figure 2 This is a flow chart of the automated SPECT thyroid analysis method provided in an embodiment of the present disclosure. From a program perspective, the execution of the process can be a program hosted on an application server or application terminal. It is understood that this method can be executed by any device, equipment, platform, or device cluster with computing and processing capabilities.

[0049] like Figure 2 As shown, the automatic thyroid analysis method includes steps S1 to S5.

[0050] S1: Segment the thyroid gland and thyroid nodules in SPECT thyroid images.

[0051] S2: Classify the properties of the segmented thyroid nodules;

[0052] S3: Measure the size of the segmented thyroid gland;

[0053] S4: Classify the uptake function of the segmented thyroid gland;

[0054] S5: Output analysis report based on classification results and measurement results.

[0055] SPECT images are usually stored in 2D DICOM format. The original images can be pre-processed as necessary.

[0056] In one embodiment, operations S1-S4 can be performed by a pre-trained deep learning model, which includes four key modules: an automatic segmentation module for the thyroid gland and its nodules, a thyroid nodule property classification module, an automatic thyroid size measurement module, and an automatic thyroid uptake function classification module, which respectively perform the above-mentioned S1-S4. Figure 3 These key modules are described below.

[0057] Figure 3 Schematic diagram of the model structure used to perform the automatic thyroid analysis method in the embodiment of the present disclosure.

[0058] like Figure 3 As shown in the figure, after the SEPCT thyroid image to be analyzed is segmented by the thyroid and its nodule automatic segmentation module, it is processed by the thyroid nodule property classification module, the thyroid size automatic measurement module and the thyroid uptake function automatic classification module respectively. Finally, the analysis report is output based on the integration of the results of each module.

[0059] (1) SPECT thyroid and nodule automatic segmentation module

[0060] This module uses a deep learning-based segmentation model to segment the thyroid gland and thyroid nodules. The segmentation model has four end-to-end channels, corresponding to the background, left lobe thyroid, right lobe thyroid, and thyroid nodules in the thyroid image. That is, the segmentation types output by this module are 0-background, 1-left lobe thyroid, 2-right lobe thyroid, and 3-thyroid nodules.

[0061] When training the segmentation model, the training samples are first labeled. Specifically, the thyroid gland and possible thyroid nodules in the SPECT thyroid image are separately labeled with contour lines, i.e., the segmentation and labeling of the two categories of thyroid and thyroid nodules. If there is no thyroid nodule in the image, only the thyroid gland is labeled. The labeling format is as follows: Figure 4 The example given is shown. Figure 4 The yellow outline in the middle outer circle represents the outline of the thyroid organ, and the red outline in the inner circle represents the area of ​​the thyroid nodule. The image shows that the right lobe of the thyroid is a hot nodule, and the left lobe of the thyroid is a cold nodule, that is, there are two types of nodules in the thyroid gland at the same time.

[0062] When building the segmentation model, considering the relatively limited availability of SPECT thyroid imaging data and images containing nodules, we introduced CLIP (Contrastive Language-Image Pretraining) technology to improve model generalization. CLIP is a language-image pretraining method or model based on contrastive learning. Its training data consists of text-image pairs—an image and its corresponding text description. Through contrastive learning, the model learns the matching relationship between text-image pairs.

[0063] Figure 5 Schematic diagram of the structure of the segmentation model in the embodiment of the present disclosure.

[0064] like Figure 5 As shown in the figure, the segmentation model includes a text encoder, an image encoder and an image segmentation decoder, wherein both the text encoder and the image encoder adopt CLIP technology. This combination can effectively utilize text semantic information and the feature extraction capabilities of a large-scale trained model.

[0065] The text encoder is the CLIP text encoder, which encodes input text prompts into text features. Pre-trained on large-scale text data, the CLIP text encoder has learned rich semantic representation capabilities and can convert input text descriptions into high-dimensional semantic vectors. These vectors contain information about the concepts, attributes, and relationships expressed in the text. In thyroid and thyroid nodule segmentation, it is used to encode thyroid-related text, such as "normal thyroid," "normal right lobe thyroid," and "thyroid nodule," into semantic feature vectors, providing a semantic foundation for subsequent fusion with image features. The text encoder parameters remain fixed during training, and only text feature extraction is performed.

[0066] The image encoder includes a CLIP image encoder, which is used to encode the input SPECT thyroid image into image features. Because CLIP is trained on natural images, the natural image input is three-channel, while the SPECT thyroid image is single-channel. Considering that directly converting a single-channel image to a three-channel image may lose information, a learnable convolutional layer is added before the CLIP image encoder to achieve single-channel to three-channel conversion. That is, the image encoder includes a channel conversion layer and a CLIP image encoder. During training, the image encoder only updates the parameters of the newly added channel conversion layer, while the parameters of the CLIP image encoder remain unchanged. The reason why the CLIP image encoder parameters remain unchanged is that the amount of training data for this model (i.e., the CLIP image encoder) is far greater than the amount of SPECT thyroid data that can be collected. Fine-tuning with a small dataset will affect the model's generalization ability.

[0067] The image segmentation decoder can use the decoder in the UNet or UNETR architecture. It fuses the text features output by the text encoder with the image features output by the image encoder, performing feature upsampling and object segmentation. The input to the image decoder does not rely solely on the features of the image encoder; it is a fusion of text and image features. The image segmentation decoder requires parameter updates during training.

[0068] The training process for this segmentation model is as follows: a preset text prompt word is input into the CLIP text encoder to generate text features N*C. A SPECT thyroid image is input into the image encoder to generate image features H*W*C. The text features are multiplied with the image features to generate fused features H*W*N, where N represents the segmentation category, C is the number of feature channels, and H and W are the image feature sizes. The image segmentation decoder generates image segmentation predictions. The Dice Loss function is used to calculate the difference between the ground truth and the predictions. The image decoder parameters are updated through multiple rounds of optimization to ultimately obtain the segmentation model.

[0069] When using this segmentation model, the preset text prompt words and the SPECT thyroid image to be segmented are also input into the text encoder and image encoder respectively. After feature fusion, the segmentation result is obtained by the image segmentation encoder.

[0070] When dealing with SPECT thyroid and thyroid nodule segmentation, the biggest difficulty is that data is difficult to obtain and difficult to accumulate to the order of magnitude of existing two-dimensional image segmentation tasks. In addition, due to the variety of thyroid lesions, the complexity of the segmentation task is relatively high. In the embodiment disclosed herein, a CLIP model trained with large-scale data is introduced to fuse text information with image information, so that the segmentation model has high-quality text feature extraction and image feature extraction capabilities. This multimodal information fusion method and pre-trained CLIP model make up for the shortcomings of the small amount of SPECT thyroid data and the difficulty in training to obtain a model with good generalization. At the same time, the segmentation model only needs to learn the parameters of the image encoder channel conversion layer and the image segmentation decoder, which effectively reduces the learning complexity of the model. Experimental verification shows that the DICE index of the segmentation model can reach: the DICE index of the left lobe thyroid organ is 0.9322, the DICE index of the right lobe thyroid organ is 0.9289, and the DICE index of the thyroid nodule is 0.8636.

[0071] (2) Thyroid nodule classification module

[0072] In SPECT thyroid imaging, thyroid nodules can be divided into four categories based on the difference in radioactivity of the nodules: hot, warm, cold, and cool. The specific imaging manifestations of these nodules are as follows: thyroid nodules with radioactivity significantly higher than that of the surrounding normal thyroid tissue are considered hot nodules; thyroid nodules with radioactivity similar to that of the surrounding normal thyroid tissue are considered warm nodules; thyroid nodules with radioactivity lower than that of the surrounding normal thyroid tissue and higher than the normal background tissue are considered cool nodules; and thyroid nodules with radioactivity lower than the surrounding normal background tissue are considered cold nodules. Different types of nodules have different benign and malignant characteristics. Single hot nodules are mainly seen in functionally autonomous thyroid adenomas and generally do not require fine needle aspiration cytology examination. Multiple hot nodules can be seen in nodular goiters with different functions of each nodule; warm nodules are mainly seen in functionally normal thyroid adenomas, nodular goiters and chronic lymphocytic thyroiditis; cool nodules or cold nodules are mainly seen in thyroid cancer, thyroid adenoma, thyroid cysts, hemorrhage, calcification and focal subacute thyroiditis. The incidence of canceration of single cool nodules or cold nodules is higher, while the incidence of canceration of multiple cool nodules or cold nodules is lower.

[0073] According to the above principles of thyroid nodules, when classifying the properties of segmented thyroid nodules, the segmentation results obtained by S1 can be used to classify the properties of the segmented thyroid nodules by comparing the radioactivity of the segmented thyroid nodules, normal thyroid regions, and normal background regions. The classification results include hot nodules, warm nodules, cold nodules, and cool nodules. The thyroid nodules and normal thyroid regions are the segmented regions of the thyroid nodules and thyroid glands obtained by the segmentation model, the normal background region is the non-thyroid region, and the radioactivity is the radioactivity count after reading the image DICOM, similar to the pixel value in a natural image.

[0074] Since the single or multiple occurrences of various types of nodules have an impact on the diagnostic results, for example, single cold nodules or cold nodules are mostly malignant nodules, while multiple nodules have a lower incidence of cancer, it is equally important to classify whether thyroid nodules are single or multiple. The segmentation model in this embodiment can segment out thyroid nodules, and if there are multiple nodules, they will be automatically segmented out. Based on the segmentation results, the problem of whether the nodules are single or multiple can also be directly solved. Therefore, the segmentation results of the segmentation model in this embodiment are combined with the original SPECT thyroid image, and the classification of thyroid hot nodules, warm nodules, cold nodules and cool nodules, as well as single and multiple classifications, can be accurately determined by calculating the value.

[0075] The thyroid nodule classification model in the embodiment of the present disclosure was verified by experimental data, and the statistical results of the indicators of various nodules were obtained as follows:

[0076] a) The DICE for segmentation of hot nodules was 0.9278, with a sensitivity of 0.9231 and a specificity of 0.9958; b) the DICE for segmentation of warm nodules was 0.8091, with a sensitivity of 0.800 and a specificity of 1.000; c) the DICE for segmentation of cool nodules was 0.8582, with a sensitivity of 0.9385 and a specificity of 0.9952; d) the DICE for segmentation of cold nodules was 0.8679, with a sensitivity of 0.9375 and a specificity of 0.9978.

[0077] The thyroid nodule property classification module in the present embodiment makes full use of the SPECT imaging characteristics and the principle of thyroid nodules. By comparing the radioactivity of the thyroid gland and the thyroid nodules, the hot, warm, cold, and cool nodule classifications can be accurately obtained. The detection performance of various types of nodules is excellent. At the same time, this method is not affected by the imbalance of sample data between nodules. Since the existing technology has a large number of cold nodules and cool nodules in the SPECT thyroid data, the model obtained by training has a greater probability of misclassifying hot nodules or warm nodules as cold nodules or cool nodules. At the same time, the thyroid nodule property classification module can obtain single and multiple classifications of thyroid nodules at the same time, which is more valuable than the existing method. The method disclosed herein can also effectively distinguish cases without thyroid nodules and cases with thyroid nodules, thereby achieving the purpose of thyroid nodule screening and detection.

[0078] (3) Automatic thyroid size measurement module

[0079] SPECT thyroid examinations are not only used to diagnose thyroid nodules, but also to assess thyroid organ size, often used in the evaluation and review of drug or surgical treatments. Currently, thyroid size is often measured by manually outlining the thyroid's long and short diameters. This process is subject to significant human variability, increasing manual operation time and further complicating outlining if a thyroid nodule is present.

[0080] In the embodiment of the present disclosure, the segmented thyroid region can be automatically calculated based on the segmentation results of the segmentation model, and the long diameter and short diameter of the thyroid region can be directly obtained. The measurement diagram is as follows: Figure 6 Combined with the thyroid segmentation result obtained in step S1, the segmented thyroid gland is distinguished into left and right lobe regions, and the major and minor axes of the left and right lobe regions are measured respectively and / or the area and weight of the left and right lobe are measured respectively.

[0081] This measurement model automatically measures the size of the thyroid gland based on the segmentation results, reducing the time spent by doctors on manual outlining, solving the problem of poor repeatability of manual outlining, and providing more accurate quantitative information about the thyroid gland.

[0082] (4) Thyroid uptake function judgment module

[0083] A key function of SPECT thyroid imaging is the diagnosis of thyroid uptake function. Traditional methods rely on visual assessment of uptake patterns and lack quantitative metrics. When multiple thyroid nodules are present, it is difficult to accurately separate the uptake contributions of each nodule. Existing deep learning-based solutions are unable to distinguish uptake in the left and right lobes of the thyroid.

[0084] In the embodiment of the present disclosure, the segmented thyroid region and non-thyroid region are input into a classifier. The classifier may be a random forest model having three output categories, corresponding to increased, normal, and decreased thyroid uptake function, respectively. Figure 7 A schematic diagram for judging thyroid uptake function is given.

[0085] During training, the SPECT thyroid medical records describe thyroid uptake in the left and right lobes, and the classification is manually labeled into three categories: increased, normal, and decreased. Combined with the thyroid and nodule segmentation results obtained in step S1, the normal thyroid regions in the left and right lobes are obtained. These normal thyroid regions and the non-thyroid background region are input into the classifier. Through iterative optimization, the classification of thyroid uptake function in a single lobe can be achieved.

[0086] Experimental verification shows that the thyroid uptake function classification model can achieve the following results: the precision of the classification indicator for category 1 "increased uptake" is 0.94, and the recall rate is 0.97; the precision of the classification indicator for category 2 "normal uptake" is 0.93, and the recall rate is 0.95; the precision of the classification indicator for category 3 "decreased uptake" is 0.95, and the recall rate is 1.00.

[0087] (5) SPECT thyroid automatic analysis report output

[0088] The results of thyroid nodule classification, thyroid size measurement, and thyroid uptake function classification obtained in steps S2-S4, respectively, essentially complete the intelligent analysis capabilities of SPECT thyroid diagnosis. Combined with the thyroid and nodule segmentation results from step S1, thyroid nodule screening can be performed. Thyroid segmentation also provides information on the location of the left and right lobes, further clarifying the specific location of thyroid nodules, thyroid size, and thyroid uptake function. Based on this information, a corresponding analysis report is generated and output.

[0089] The automatic thyroid analysis method provided by the embodiment of the present disclosure can provide the following intelligent analysis results: (1) thyroid nodule function analysis, thyroid nodule detection and its location (left lobe, right lobe), classification (hot nodules, warm nodules, cold nodules, cool nodules and multiple nodules, single nodules); (2) thyroid overall function analysis, automatic measurement of the size of the left and right lobes of the thyroid gland and automatic judgment of the uptake function. This method can basically meet the clinical needs of the entire process of SPECT thyroid analysis, effectively play the role of intelligent auxiliary diagnosis, and greatly reduce the manual operation cost of doctors. At the same time, it has been verified by experiments that the accuracy and reliability of the automatic thyroid analysis method provided by the present invention are satisfactory.

[0090] Correspondingly, the specification of the present disclosure also provides a SPECT thyroid automatic analysis device 400 . Figure 8 Schematic diagram of an automatic analysis device 400 provided in an embodiment of the present disclosure.

[0091] like Figure 8 As shown, the apparatus 400 includes a target segmentation module 410 , a property classification module 420 , a size measurement module 430 , a function classification module 440 and a report output module 450 .

[0092] The object segmentation module 410 is used to segment the thyroid gland and thyroid nodules in the SPECT image;

[0093] A property classification module 420 is used to classify the properties of the segmented thyroid nodules;

[0094] a size measurement module 430 for measuring the size of the segmented thyroid gland;

[0095] a function classification module 440 for classifying the uptake function of the segmented thyroid gland;

[0096] The report output module 450 is used to generate an analysis report based on the property classification results, the ingestion function classification results and the measurement results.

[0097] The present disclosure also provides a computer program for automatic SPECT thyroid analysis, which, when executed by a processor, enables the processor to implement the various methods described above.

[0098] The disclosed automatic thyroid analysis device based on SPECT images employs a target segmentation module to segment the thyroid gland and thyroid nodules. The pre-trained CLIP model introduced into the segmentation module overcomes the limitations of SPECT thyroid data, which makes it difficult to train a generalizable model, thereby improving segmentation performance. The device can automatically classify thyroid uptake function. In the property classification module, a novel method is employed to analyze the left and right lobes of the thyroid gland separately, combining thyroid nodule segmentation results. This method analyzes only the thyroid region and background area, rather than the entire thyroid image, significantly simplifying the task complexity. The device can also automatically measure thyroid size, effectively avoiding the labor costs and manual errors associated with manual operation.

[0099] The foregoing description describes specific embodiments of the present disclosure. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims may be performed in an order different from that described in the embodiments and still achieve the desired results. Furthermore, the processes depicted in the accompanying drawings do not necessarily require the specific order shown or the sequential order to achieve the desired results. In certain embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0100] The various embodiments of this disclosure are described in a progressive manner. References to the common and similar parts of the various embodiments will be made to each other. Each embodiment focuses on the differences between the other embodiments. In particular, the device and computer program embodiments are generally similar to the method embodiments, so their descriptions are relatively simplified. For relevant details, references to the method embodiments will suffice.

[0101] The apparatus, computer program, and method provided in the embodiments of the present disclosure correspond to each other and therefore have similar beneficial technical effects as the corresponding methods. Since the beneficial technical effects of the methods have been described in detail above, they will not be repeated here.

[0102] The units or modules involved in the embodiments described in this disclosure may be implemented by software or programmable hardware. The units or modules described may also be provided in a processor, and the names of these units or modules do not, in certain circumstances, constitute limitations on the units or modules themselves.

[0103] The above description is merely a preferred embodiment of the present disclosure and an illustration of the technical principles employed. Those skilled in the art should understand that the scope of the invention herein is not limited to the technical solutions formed by the specific combination of the above-mentioned technical features, but also encompasses other technical solutions formed by any combination of the above-mentioned technical features or their equivalents without departing from the inventive concept. For example, a technical solution formed by replacing the above-mentioned features with (but not limited to) technical features with similar functions disclosed in this disclosure.

Claims

1. A SPECT thyroid automatic analysis method, characterized in that: include: Segmentation of the thyroid gland and thyroid nodules in SPECT images; Classify the properties of the segmented thyroid nodules; Measure the size of the segmented thyroid gland; Classify the uptake function of the segmented thyroid gland; Generate an analysis report based on the property classification results, uptake function classification results and measurement results.

2. The method according to claim 1, characterized in that Segmentation of the thyroid gland and thyroid nodules in SPECT thyroid images includes: The thyroid gland and thyroid nodules are segmented using a deep learning-based segmentation model, wherein the segmentation model has four end-to-end channels corresponding to the background, left lobe thyroid gland, right lobe thyroid gland, and thyroid nodules in the thyroid image, respectively.

3. The method according to claim 2, characterized in that The segmentation model includes a text encoder, an image encoder and an image segmentation decoder, wherein, The text encoder is a CLIP text encoder, which is used to encode the input text prompt words into text features; The image encoder comprises a CLIP image encoder, configured to encode the input SPECT thyroid image into image features; The image segmentation decoder adopts a decoder in a UNet structure or a UNETR structure, and is used to fuse the text features and the image features and then perform feature upsampling and target segmentation.

4. The method according to claim 3, characterized in that The image encoder further includes a channel conversion layer disposed before the CLIP image encoder, for converting the thyroid image into a three-channel image.

5. The method according to claim 3, characterized in that The segmentation model is obtained by training as follows: Input the preset text prompt word into the CLIP text encoder to obtain text features N*C, input the SPECT thyroid image into the image encoder to obtain image features H*W*C, and multiply the text features with the image features to obtain fusion features H*W*N, where N is the segmentation category, C is the number of feature channels, and H and W are image feature sizes; The Dice Loss function is used to calculate the difference between the true annotation and the predicted result, and the parameters of the image decoder are updated through multiple rounds of optimization.

6. The method according to claim 1, characterized in that The classifying of the properties of the segmented thyroid nodules includes: The radioactivity of the thyroid nodule, normal thyroid area and normal background area is obtained by comparison and segmentation, and the nature of the thyroid nodule is classified. The classification results include hot nodule, warm nodule, cold nodule and cool nodule.

7. The method according to claim 1, characterized in that The classification of the uptake function of the segmented thyroid gland includes: The segmented thyroid region and non-thyroid region are input into a classifier, which is a random forest model with three output categories, corresponding to increased, normal, and decreased thyroid uptake function.

8. The method according to claim 1, characterized in that Measurements of the size of the segmented thyroid gland include: The segmented thyroid gland is divided into a left lobe and a right lobe region, and the major axis and minor axis of the left lobe and the right lobe region are measured respectively, and / or the area and weight of the left lobe and the right lobe are measured respectively.

9. A SPECT thyroid automatic analysis device, characterized in that: include: The target segmentation module is used to segment the thyroid gland and thyroid nodules in SPECT images; Property classification module, used to classify the properties of the segmented thyroid nodules; A size measurement module, used to measure the size of the segmented thyroid gland; A functional classification module is used to classify the uptake function of the segmented thyroid gland; The report output module is used to generate analysis reports based on property classification results, uptake function classification results and measurement results.

10. A computer program for automatic SPECT thyroid analysis, which, when executed by a processor, enables the processor to implement the method according to any one of claims 1 to 8.

Images