A method and system for detecting lesions using positron emission tomography (PET)

Abstract

This specification provides a method and system for detecting lesions using positron emission tomography (PET) imaging. The method includes: classifying elements in a PET image of a scanned object to generate a first reference image; determining a target segmentation result based on the PET image, a CT image of the scanned object, and the first reference image using a segmentation model, wherein the segmentation model is a machine learning model; and determining a lesion segmentation result based on the target segmentation result.

Description

Technical Field

[0001] This specification relates to the field of medical image processing, and in particular to a positron emission tomography (PET) lesion detection method and system. Background Technology

[0002] Positron emission tomography (PET) is an examination method that combines PET technology with computed tomography (CT). PET provides detailed molecular information about the function and metabolism of lesions, while CT provides precise anatomical localization of the lesions. Because the metabolism and function of tumors differ from normal substances, PET can safely, conveniently, and effectively diagnose tumors at an early stage, and can detect very small lesions (such as those larger than 5 mm in diameter). Therefore, PET examination is of great significance for the early differential diagnosis of tumors, identifying tumor recurrence, staging and restaging tumors, locating primary and metastatic lesions, guiding and determining tumor treatment plans, and evaluating treatment efficacy. Accurately and effectively detecting lesions in PET images is also a current research hotspot.

[0003] Therefore, this invention proposes a method and system for lesion detection in PET images to improve the effectiveness of lesion detection in PET images. Summary of the Invention

[0004] One embodiment of this specification provides a positron emission tomography (PET) lesion detection method, the method comprising: classifying elements in a PET image of a scanned object to generate a first reference image; determining a target segmentation result based on the PET image, a CT image of the scanned object, and the first reference image, wherein the segmentation model is a machine learning model; and determining a lesion segmentation result based on the target segmentation result.

[0005] One embodiment of this specification provides a positron emission tomography (PET) lesion detection system. The system includes: a generation module for classifying elements in a PET image of a scanned object to generate a first reference image; a first determination module for determining a target segmentation result based on the PET image, a CT image of the scanned object, and the first reference image, using a segmentation model, wherein the segmentation model is a machine learning model; and a second determination module for determining a lesion segmentation result based on the target segmentation result.

[0006] One embodiment of this specification provides a positron emission tomography (PET) lesion detection device, including at least one processor and at least one storage device. The storage device is used to store instructions, and when the at least one processor executes the instructions, it implements the PET lesion detection method disclosed in this specification.

[0007] The positron emission tomography (PET) lesion detection method provided in this specification detects lesions by training a cascaded segmentation model. The output of the previous segmentation model is used as the input of the next, narrowing the detection range of the segmentation model, enhancing its learning of false positives, reducing false positives in PET lesion detection, and improving precision. Simultaneously, some embodiments of this specification utilize CT images and a first reference image to assist in PET lesion detection. The first reference image may contain information about high-metabolic physical points of the scanned object. Introducing the first reference image improves attention to high-metabolic physical points during lesion detection, thus improving lesion recall. Introducing CT images utilizes the more anatomical information contained in CT images (e.g., clearer lesion boundaries), enabling better differentiation between pathological and physiological regions during lesion detection, thereby improving the accuracy of lesion detection. On the other hand, during data preprocessing, the element values ​​in the PET images were not truncated, preserving their clinical significance and enhancing the stability of the segmentation model, thus improving lesion recall and precision. During model training, by setting the weight of the loss term for pathological labels to be greater than that for non-pathological labels, the segmentation model's learning of pathology was enhanced, improving its precision.

[0008] In some embodiments of this specification, positron emission tomography (PET) lesion detection is also referred to as PET lesion detection. Attached Figure Description

[0009] This specification will be further described by way of exemplary embodiments, which will be described in detail with reference to the accompanying drawings. These embodiments are not limiting; in these embodiments, the same reference numerals denote the same structures, wherein:

[0010] Figure 1a These are schematic diagrams illustrating application scenarios of the PET lesion detection system according to some embodiments of this specification;

[0011] Figure 1b This is a schematic diagram of the structure of a computer that can perform all or part of the functions of the lesion detection device 140 according to some embodiments of this specification.

[0012] Figure 2 These are exemplary block diagrams of a PET lesion detection system according to some embodiments of this specification;

[0013] Figure 3 This is an exemplary flowchart of a PET lesion detection method according to some embodiments of this specification;

[0014] Figure 4This is an exemplary schematic diagram illustrating the determination of target segmentation results based on a segmentation model according to some embodiments of this specification;

[0015] Figure 5 These are exemplary schematic diagrams illustrating the generation of segmentation models according to some embodiments of this specification;

[0016] Figure 6a This is an exemplary schematic diagram showing the determination of a first reference figure according to some embodiments of this specification;

[0017] Figure 6b This is an exemplary schematic diagram illustrating the generation of first training labels according to some embodiments of this specification;

[0018] Figure 6c This is an exemplary schematic diagram illustrating the generation of a second training label according to some embodiments of this specification. Detailed Implementation

[0019] To more clearly illustrate the technical solutions of the embodiments in this specification, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are merely some examples or embodiments of this specification. For those skilled in the art, these drawings can be applied to other similar scenarios without creative effort. Unless obvious from the context or otherwise specified, the same reference numerals in the drawings represent the same structures or operations.

[0020] It should be understood that the terms “system,” “device,” “unit,” and / or “module” used herein are one way to distinguish different components, elements, parts, sections, or assemblies at different levels. However, if other terms can achieve the same purpose, they may be replaced by other expressions.

[0021] As indicated in this specification and claims, unless the context clearly indicates otherwise, the words "a," "an," "an," and / or "the" do not specifically refer to the singular and may also include the plural. Generally speaking, the terms "comprising" and "including" only indicate the inclusion of expressly identified steps and elements, which do not constitute an exclusive list, and the method or apparatus may also include other steps or elements.

[0022] Flowcharts are used in this specification to illustrate the operations performed by the system according to embodiments of this specification. It should be understood that the preceding or following operations are not necessarily performed in exact order. Instead, the steps can be processed in reverse order or simultaneously. Furthermore, other operations can be added to these processes, or one or more steps can be removed from them.

[0023] Some descriptions of the present invention are provided in conjunction with positron emission tomography-computed tomography (PET-CT) images. It should be understood that this is for illustrative purposes and not intended to limit the scope of the invention. The methods and systems of this invention can be used to process images or image data from other imaging modalities, such as images or image data generated by a positron emission tomography-magnetic resonance imaging (PET-MR) system.

[0024] Figure 1a This is a schematic diagram illustrating the application scenarios of the PET lesion detection system according to some embodiments of this specification.

[0025] like Figure 1a As shown, the application scenario 100 of the PET lesion detection system may include an imaging device 110, a scanning object 120, a storage device 130, a lesion detection device 140, and a terminal device 150.

[0026] Imaging device 110 can generate an image by scanning object 120 or a portion thereof. The image may be a medical image, such as a PET image, CT image, etc. In some embodiments, the image generated by imaging device 110 can be transmitted to lesion detection device 140 for processing or stored in storage device 130. In some embodiments, the image may be in two-dimensional (2D) or three-dimensional (3D) form, and this specification is not limited thereto.

[0027] In some embodiments, the imaging device 110 may include one or more combinations of computed tomography (CT) devices, positron emission tomography (PET) devices, PET-CT devices, magnetic resonance imaging (MRI) devices, PET-MRI devices, etc.

[0028] The scanned object 120 refers to patients, phantoms, animals, etc. that need to be scanned by imaging equipment for medical testing.

[0029] Storage device 130 can store images and / or image-related information. Image-related information may include one or more combinations of algorithms and computer instructions for processing images, models for lesion detection, lesion segmentation results, etc. Storage device 130 can be one or more combinations of hierarchical databases, network databases, relational databases, etc. Storage device 130 can also store operating parameters related to the PET lesion detection system. Storage device 130 can be local or remote. In some embodiments, storage device 130 can be a storage device that stores information electrically, such as one or more combinations of random access memory (RAM), read-only memory (ROM), etc. In some embodiments, storage device 130 can be part of or independent of lesion detection device 140. In some embodiments, storage device 130 can be connected to other components in the PET lesion detection system via a network.

[0030] The lesion detection device 140 can acquire images generated by the imaging device 110 or images stored in the storage device 130. The lesion detection device 140 can perform lesion detection based on the acquired images and transmit the lesion segmentation results to the terminal device 150. In some embodiments, the lesion detection device 140 may also be referred to as a processing device. In some embodiments, the lesion detection device 140 may include one or more combinations of a processor, a processor core, memory, etc. For example, the lesion detection device 140 may include one or more combinations of a central processing unit (CPU), an application-specific integrated circuit (ASIC), an application-specific instruction set processor (ASIP), a graphics processing unit (GPU), a physical processor (PPU), a digital signal processor (DSP), a field-programmable gate array (FPGA), a programmable logic device (PLD), a controller, a microcontroller unit, a processor, a microprocessor unit, an advanced RISC processor (ARM), etc.

[0031] Terminal device 150 can communicate with imaging device 110, lesion detection device 140, and / or storage device 130. In some embodiments, terminal device 150 can obtain lesion segmentation results from lesion detection device 140. Terminal device 150 may include one or more combinations of mobile phone 151, tablet 152, computer 153, etc.

[0032] In some embodiments, the application scenario 100 of the PET lesion detection system may further include a network (not shown). The network may be a local area network (LAN), a wide area network (WAN), a public network, a private network, a dedicated network, a public switched telephone network (PSTN), the Internet, a virtual network, a metropolitan area network, a telephone network, or a combination thereof. In some embodiments, the imaging device 110, the storage device 130, the lesion detection device 140, and the terminal device 150 may be connected to and / or communicate with each other via a network (e.g., through wired connection, wireless connection, or a combination thereof).

[0033] In some embodiments, the lesion detection device 140 and / or storage device 130 can implement the functions of the PET lesion detection system through a cloud computing platform. The cloud computing platform may include a storage-based cloud platform, a data-based cloud platform, and an integrated cloud platform. The cloud platform configured in the PET lesion detection system can be one or more combinations of public cloud, private cloud, hybrid cloud, etc.

[0034] Figure 1b This is a schematic diagram of a computer that can perform all or part of the functions of the lesion detection device 140 according to some embodiments of this specification. The functions of the lesion detection device 140 or a portion thereof can be implemented by a computer via its hardware, software programs, firmware, or a combination thereof. Although only one computer is shown for convenience, the computer functions related to the lesion detection device 140 as described herein can be implemented in a distributed manner on multiple similar platforms to distribute the processing load. In some embodiments, the computer can be a general-purpose computer or a purpose-specific computer.

[0035] like Figure 1b As shown, the lesion detection device 140 may include a COM port 147, which can be connected to or from a network to facilitate data communication. The lesion detection device 140 may also include a central processing unit (CPU) 142, consisting of one or more processors, for executing program instructions. The computer platform may include an internal communication bus 141 and data storage (e.g., a disk 145, a read-only memory (ROM) 143, a random access memory (RAM) 144). The data storage is used to store various data files processed and / or transmitted by the computer, different forms of programs, program instructions that may be executed by the CPU 142, etc. The lesion detection device 140 may also include an input / output (I / O) port 146 for supporting input / output flows between the lesion detection device 140 and other components (e.g., terminal device 150) in the PET lesion detection system.

[0036] The above description is illustrative and does not limit the scope of this disclosure. Many alternatives, modifications, and variations will be apparent to those skilled in the art. The features, structures, methods, and other characteristics described in some embodiments of this specification can be combined in various ways to obtain other and / or alternative exemplary embodiments. For example, while implementations of various components described in some embodiments of this specification can be characterized on a hardware device, they can also be implemented as a software-only solution, such as installed on an existing server. Furthermore, the lesion detection apparatus disclosed in some embodiments of this specification can be implemented as firmware, a firmware / software combination, a firmware / hardware combination, or a hardware / firmware / software combination. In addition, a PET lesion detection system may include one or more other components, or one or more of the components described above may be omitted or combined. However, these changes and modifications do not depart from the scope of this disclosure.

[0037] Figure 2 These are exemplary block diagrams of a PET lesion detection system according to some embodiments of this specification. Figure 2 As shown, the PET lesion detection system 200 may include a generation module 210, a first determination module 220, and a second determination module 230. In some embodiments, the PET lesion detection system may be implemented on a lesion detection device 140 (e.g., the central processing unit 142 of the lesion detection device 140).

[0038] The generation module 210 is used to classify elements in the PET image of the scanned object to generate a first reference image. For example, the generation module 210 can obtain a normalized uptake value threshold and classify the normalized uptake values ​​of elements in the PET image based on the normalized uptake value threshold to determine the first reference image.

[0039] The first determining module 220 is used to determine the target segmentation result based on the PET image, the CT image of the scanned object, and the first reference image, through a segmentation model, which is a machine learning model.

[0040] The second determining module 230 is used to determine the lesion segmentation result based on the target segmentation result. In some embodiments, the second determining module 230 can be further used to perform quantitative analysis on the lesion segmentation result.

[0041] Further descriptions of the generation module 210, the first determining module 220, and the second determining module 230 can be found elsewhere in this specification (e.g., Figure 3 (and related descriptions).

[0042] It should be noted that the above description of the PET lesion detection system and its modules is for convenience only and should not be construed as limiting this specification to the scope of the illustrated embodiments. It is understood that those skilled in the art, after understanding the principles of the system, may arbitrarily combine the various modules or construct subsystems connected to other modules without departing from these principles. In some embodiments, Figure 2 The generation module 210, the first determining module 220, and the second determining module 230 disclosed herein can be different modules within a single system, or a single module can implement the functions of the two modules described above. For example, the modules can share a single storage module, or each module can have its own separate storage module. Such variations are all within the scope of protection of this specification.

[0043] Figure 3 This is an exemplary flowchart of a PET lesion detection method according to some embodiments of this specification. In some embodiments, process 300 may be performed by lesion detection device 140 and / or Figure 2 One or more modules are executed as shown. For example... Figure 3 As shown, process 300 may include the following steps:

[0044] Step 310: Classify the elements in the PET image of the scanned object to generate a first reference image. In some embodiments, step 310 may be performed by the generation module 210.

[0045] The object being scanned refers to the object that needs to be scanned using medical testing equipment. For example, the object being scanned could be the patient's entire body or the patient's head, neck, chest, abdomen, etc.

[0046] PET images are medical images obtained by scanning a target using a PET scanner. In some embodiments, the lesion detection device 140 can acquire PET images in various ways, and this specification does not limit this. For example, the lesion detection device 140 can acquire PET images directly from an imaging device or from stored PET images in a storage device. As another example, the lesion detection device 140 can reconstruct PET images from PET data.

[0047] An element is the smallest unit in a PET image. For example, an element can be a pixel or voxel in a PET image, and an element value can be a pixel value, voxel value, or standardized uptake value (SUV). The standardized uptake value (SUV) is a commonly used semi-quantitative indicator in PET for tumor diagnosis; it refers to the ratio of the radioactivity of the imaging agent taken up by a local tissue to the average injected activity throughout the body. SUV = radioactivity concentration of the lesion (kBq / ml) / (injected dose (MBq) / body weight (kg)). The element value of a particular element in a PET image reflects the uptake / metabolism of the corresponding physical point in the scanned object. For example, the higher the SUV value of an element, the higher the uptake of the corresponding physical point. High uptake of a physical point may be caused by normal physiological responses (such as brain activity) or by pathological responses (such as inflammation, tumors, etc.). Clinically, physical points with an SUV value greater than 2.5 are generally considered to be more likely to be malignant lesions or lesions caused by other inflammations, while physical points with an SUV value less than 2.0 are more likely to be normal or benign changes.

[0048] The first reference image can be obtained by classifying the elements in the PET image, and it can reflect the categories of the elements in the PET image. For example, the first reference image can show which elements in the PET image correspond to physical points with high metabolism (also known as high uptake).

[0049] In some embodiments, the lesion detection device 140 can classify elements based on element values ​​in a PET image to generate a first reference image. For example, elements can be classified based on a preset range or preset threshold. The preset range or preset threshold can be set based on actual medical diagnostic needs.

[0050] As an example only, the lesion detection device 140 can acquire a standardized uptake threshold and classify elements in a PET image based on the standardized uptake threshold to generate a first reference image. In this case, the first reference image can also be referred to as a PET threshold segmentation image.

[0051] A standardized uptake value threshold is a threshold used to classify standardized uptake values. In some embodiments, the standardized uptake value threshold can be determined based on the actual medical diagnosis. For example, in general medical diagnosis, an SUV value of 2.5 in a PET image is generally used as the dividing line between benign lesions and malignant tumors. Therefore, the standardized uptake value threshold can be determined to be 2.5 based on this.

[0052] In some embodiments, when classifying elements based on a standard uptake value threshold, elements with a standard uptake value not less than the threshold can be classified into a first category (corresponding to high uptake physical points), and elements with a standard uptake value less than the threshold can be classified into a second category (corresponding to low uptake physical points). In the first reference figure, the first category of elements can be presented in a specific manner. For example, as shown... Figure 6a As shown, the head and neck region can be identified in PET image 610-1, resulting in head and neck region image 610-2 (i.e., the portion in box B). Based on a standardized uptake value threshold of 2.5, elements in head and neck region image 610-2 (the portion in box B) are classified. Elements with an SUV value greater than or equal to 2.5 are set to 1, and elements with an SUV value less than 2.5 are set to 0, resulting in first reference image 610-3. First reference image 610-3 can be a binary image. Element values ​​(such as the grayscale value of pixels) of 0 correspond to black areas, reflecting low uptake physical points in PET image 610-1; element values ​​of 1 correspond to white areas, reflecting high uptake physical points in the PET image. Overlaying first reference image 610-3 onto PET image 610-1 yields schematic diagram 610-4. By overlaying the first reference image 610-3 onto the PET image 610-1 to form a schematic diagram 610-4, the areas of element values ​​of 0 and 1 in the first reference image can be mapped one-to-one with actual human body parts. Schematic diagram 610-4 allows for a more intuitive identification of the human body parts corresponding to high uptake physical points. It can be seen that in schematic diagram 610-4, elements with an SUV value greater than or equal to 2.5 in the head and neck region are presented in dark colors (such as the dark areas in boxes C, D, and E in schematic diagram 610-4), which can reflect the corresponding high-metabolic physical points in the PET image 610-1. It should be noted that the standardized uptake value threshold can also be any other value, and the first reference image can also be generated in other ways; this specification does not impose any restrictions on this.

[0053] In some embodiments of this specification, a first reference image is generated by classifying elements in a PET image based on SUV values ​​for subsequent lesion detection. The first reference image can reflect information about high-metabolic physical points of the scanned object, which can improve the attention of the segmentation network to these high-metabolic physical points during subsequent lesion detection, thereby improving the accuracy of lesion detection.

[0054] Step 320: Based on the PET image, the CT image of the scanned object, and the first reference image, the target segmentation result is determined using a segmentation model, which is a machine learning model. In some embodiments, step 320 may be performed by the first determining module 220.

[0055] CT images are medical images obtained by scanning an object using a CT scanner. For example, CT images can be generated by scanning the head, neck, abdomen, or other parts of a human body using a CT scanner. In some embodiments, CT images and PET images are acquired simultaneously based on the same scanned object.

[0056] In some embodiments, the lesion detection device 140 can acquire CT images in various ways, and this specification does not limit the embodiments thereto. For example, the lesion detection device 140 can acquire CT images directly from the imaging device, or it can acquire stored CT images from a storage device. As another example, the lesion detection device 140 can generate CT images based on the CT values ​​of the scanned object acquired by the imaging device.

[0057] A segmentation model is a model used to detect lesions in image data. In some embodiments, the segmentation model can be based on neural network models such as U-net, VB-net, fully convolutional network (FCN), and convolutional neural network (CNN). In some embodiments, the segmentation model can consist of multiple cascaded sub-networks. For example, the segmentation model can include a first segmentation model and a second segmentation model. The first segmentation model and the second segmentation model can be interconnected. In some embodiments, the output of the first segmentation model can be used as the input of the second segmentation model.

[0058] In some embodiments, the segmentation model can be obtained based on model training. In some embodiments, the training samples for the second segmentation model can be determined based on the first segmentation model, thereby generating the second segmentation model. For more information on segmentation models, please refer to [link to relevant documentation]. Figure 4 , Figure 5 And its related descriptions.

[0059] The target segmentation results include pathological segmentation results and physiological segmentation results. Pathological segmentation results can include lesion areas caused by pathological reactions (such as inflammation, tumors, etc.). In some embodiments, pathological segmentation results may also be referred to as lesion segmentation results. In some embodiments, lesion segmentation results can include information such as the location, size, outline, and shape of the lesion. In some embodiments, lesion segmentation results can include lesion segmentation images (e.g., lesion segmentation masks). Physiological segmentation results can include information related to high uptake areas caused by physiological reactions. For example, normal physiological reactions such as abundant blood flow, body movement, and activity in different brain functional areas lead to cellular uptake of imaging agents, which will appear as high uptake areas in PET images.

[0060] In some embodiments, the lesion detection device can determine the target segmentation result through a segmentation model. In some embodiments, the input to the segmentation model is a PET image, a CT image, and a first reference image, and the output is the target segmentation result. In some embodiments, the PET image, CT image, and first reference image can be input into the segmentation model via three separate channels. In some embodiments, the PET image, CT image, and first reference image can be concatenated before being input into the segmentation model. For example, assuming that each of the aforementioned images has 256×256 pixels, concatenating the three images will yield a 256×768 image.

[0061] In some embodiments, the pathological segmentation results and physiological segmentation results output by the segmentation model can be shown in the same segmentation graph or in different segmentation graphs, and this specification does not impose any restrictions on this.

[0062] Step 330: Based on the target segmentation result, determine the lesion segmentation result. In some embodiments, step 330 may be performed by the second determining module 230.

[0063] In some embodiments, the lesion detection device 140 can determine the lesion segmentation result by extracting the pathological segmentation result from the target segmentation result. For example, in the target segmentation result based on a PET image, assuming that the element corresponding to the pathological segmentation result is predicted as 1 and the element corresponding to the physiological segmentation result is predicted as 2, the corresponding lesion segmentation result can be generated by extracting the part with a predicted value of 1. As another example, the part corresponding to the pathological segmentation result can be extracted from the target segmentation result of the PET image to generate a lesion segmentation mask.

[0064] In some embodiments, before inputting the PET image, CT image, and first reference image into the segmentation model, the lesion detection device may preprocess at least one of the aforementioned three images. Preprocessing may include one or more of the following: mapping transformation, normalization, noise reduction, and resampling.

[0065] In some embodiments, before inputting the PET image into the segmentation model, the lesion detection device can perform a mapping transformation on the PET image to determine the mapping value corresponding to each element. In some embodiments, when an element is greater than a preset value, the mapping value of the element is positively correlated with the element's value. Mapping transformation refers to mapping the value of each element in the PET image to another numerical range based on a certain correspondence. For example, mapping transformation can be based on a certain functional relationship (e.g., a direct proportional function) to transform the element values ​​in the PET image into corresponding values. The mapping value refers to the value obtained by the element value in the PET image after the mapping transformation. The preset value refers to a preset threshold for the element values ​​in the PET image. For example, the preset value can be a pixel value threshold, an SUV value threshold, etc. In some embodiments, the preset value can be determined based on actual lesion detection requirements.

[0066] In some embodiments, the lesion detection device can perform a mapping transformation on the PET image based on a preset correspondence to determine the mapping value corresponding to each element. For example, the lesion detection device can perform a mapping transformation on the PET image through the following steps:

[0067] Step S1: Set the window level (wl) and window width (ww) for the input PET image. For example, the window level wl1 = 2.5 and the window width ww1 = 5 can be set. In some embodiments, the window level wl1 and window width ww1 of the PET image can be set in other ways, which are not limited in this specification.

[0068] Step S2: Perform a mapping transformation on the element values ​​in the PET image based on preset values. For example, the mapping transformation on the elements in the PET image can be performed based on the following formula (1):

[0069]

[0070] Where x1 represents the element value (e.g., SUV value), x1′ represents the element's mapping value, and wl1-ww1 / 2 represents the preset value. Formula (1) shows that when the element value x1 is less than or equal to xl1-ww1 / 2, the element's mapping value is wl1-ww1 / 2. When the element value x1 is greater than wl1-ww1 / 2 (i.e., the preset value), the element's mapping value is x1. In other words, when the element value x1 is greater than wl1-ww1 / 2, the element value itself can be retained without truncation. In some embodiments, when the element value x1 is greater than the preset value, i.e., x1 is greater than wl1-ww1 / 2, the element's mapping value can also be determined based on other positively correlated preset correspondences; this specification does not impose any restrictions on this. For example, when the element value x1 is greater than the preset value, the element's mapping value can be the logarithm, exponent, etc., of the element value.

[0071] Step S3: Based on the mapping value x′1 determined in step S2, calculate the element value y1 in the preprocessed PET image of the final input segmentation model using formula (2).

[0072]

[0073] In some embodiments, when the mapped value x1′ = wl1 - ww1 / 2, the value of y1 is -1, and when the mapped value x′1 = x1, the value of y1 may be a number greater than 1. For example, when the original SUV value is 7.5, its mapped value is also 7.5 according to formula (1), and then input into formula (2) to obtain the corresponding value of y1 as 2. In this way, the clinical significance of the element values ​​in the PET image can be preserved.

[0074] In conventional processing, the mapping transformation and normalization of PET images are performed in the same way as the mapping transformation and normalization of CT images. Referring to formula (3), the element values ​​in the PET image are first mapped using the same method as in formula (3). It can be seen that when the element value x1 in the PET image is greater than or equal to wl1+ww1 / 2 (i.e., the preset value), the mapped value x′1 of the element is wl1+ww1 / 2, that is, the part of the element value in the PET image that is greater than the preset value is truncated. Further, the truncated mapping transformation x′1 is normalized based on formula 2, so that the element values ​​in the PET image are normalized to [-1,1]. In this way, the element values ​​in the PET image that are greater than the preset value are truncated, and the clinical significance of the element values ​​greater than the preset value cannot be reflected in the lesion detection process.

[0075] In some embodiments, before inputting the segmentation model, the lesion detection device can perform mapping transformation and normalization processing on the CT image. For example, the window width ww2 = 350 and the window level wl2 = 50 can be set, and then the grayscale range of the CT image can be normalized to [-1, 1] based on the following method. First, the element values ​​in the CT image are mapped according to formula (3):

[0076]

[0077] Where x2 represents the gray value in the CT image, and x2′ represents the mapped value of the gray value in the CT image. It can be seen from formula (3) that the mapping transformation performed on the CT image truncates the element values ​​in the CT image that are greater than wl+ww / 2.

[0078] Then, for the CT image after mapping transformation, its gray value range can be normalized to [-1, 1] based on the following formula (4):

[0079]

[0080] Where y2 represents the normalized value of grayscale values ​​in the CT image.

[0081] In some embodiments, the grayscale value range of the first reference image can be normalized to [-1, 1] using the same mapping transformation and normalization method as that used for CT images. In some embodiments, when normalizing the grayscale value range of the first reference image, the window level wl3 = 0.5 and the window width ww3 = 1 can be set.

[0082] In some embodiments of this specification, the mapping transformation methods for PET images differ from those for CT images and the first reference image. Specifically, when mapping the element values ​​in a PET image, if the element value exceeds a corresponding preset value, its mapped value is proportional to the element value and will not be truncated. For example, if the SUV value of an element exceeds the preset value, its mapped value will be equal to the original SUV value. This is because a high SUV value is an important indicator for detecting tumors; by performing a non-truncated mapping transformation, the clinical significance of the SUV value in the PET image can be preserved, improving the accuracy of lesion detection.

[0083] In some embodiments, the lesion detection device can also perform quantitative analysis on the lesion segmentation results.

[0084] Quantitative analysis refers to representing lesion segmentation results using specific quantitative parameters and then analyzing and comparing them. For example, quantitative analysis may include calculating lesion volume, major and minor axes, contour, SUVmax, SUVmean, or performing 3D rendering and display.

[0085] In some embodiments, the lesion detection device 140 can quantitatively analyze the lesion segmentation results in various ways. In some embodiments, the lesion detection device can calculate the volume and major and minor axes of the lesion based on the lesion segmentation results through mathematical measurement, pixel calculation, etc. In some embodiments, the lesion detection device 140 can obtain the lesion contour based on the lesion segmentation results through edge extraction, mathematical morphology correction, etc. In some embodiments, the lesion detection device 140 can calculate the SUVmax and SUVmean in the PET image by statistically analyzing the SUV value of each pixel in the segmentation results. In some embodiments, the lesion detection device 140 can perform three-dimensional rendering display of the lesion segmentation results in various ways. For example, constructing a three-dimensional model of the lesion based on the contour of the lesion in the lesion segmentation results, displaying different colors based on the SUV value of each pixel in the lesion segmentation results, etc.

[0086] In some embodiments of this specification, by performing quantitative analysis on the lesion segmentation results, information such as the lesion's volume, major and minor axes, contour, maximum and minimum SUV values ​​can be obtained. Furthermore, the lesion can be displayed through three-dimensional rendering, enabling more intuitive observation, analysis, and diagnosis of the lesion, thereby improving the accuracy and reliability of medical diagnosis.

[0087] In some embodiments of this specification, the first reference image may include information about high-metabolic physical points of the scanned object. Introducing the first reference image increases attention to high-metabolic physical points during lesion detection, thereby improving lesion recall. Simultaneously, CT images contain more anatomical information (e.g., clearer lesion boundaries), enabling better differentiation between pathological and physiological regions during lesion detection, thus improving the accuracy of lesion detection.

[0088] In some embodiments, to verify the impact of introducing CT images and a first reference image on the recall and precision of lesions, a segmentation model was used to process PET images, PET images + CT images, PET images + a first reference image, and PET images + CT images + a first reference image, respectively. The results showed that when the segmentation model processed PET images + CT images + a first reference image, both the recall and precision of lesions were significantly improved. This indicates that in some embodiments of this specification, introducing CT images and a first reference image can improve the precision and recall of lesions.

[0089] In some embodiments of this specification, lesion detection is performed by training a cascaded segmentation network model. The output of the preceding segmentation model is used as the input to the following segmentation model, which narrows the detection range of the subsequent model. Simultaneously, training samples generated by the preceding segmentation model for training the subsequent segmentation network are used to enhance the subsequent model's learning of false positive detection results generated by the preceding model, reducing the false positive probability of lesion segmentation and improving lesion recall and precision.

[0090] In some embodiments of this specification, a non-truncated mapping transformation is performed on the element values ​​in the PET image before inputting them into the segmentation model, instead of a truncated mapping transformation. This results in a more stable recall and precision rate for lesions in the segmentation model, while simultaneously improving the recall and precision rates.

[0091] Figure 4 This is an exemplary schematic diagram illustrating the determination of target segmentation results based on a segmentation model according to some embodiments of this specification.

[0092] In some embodiments, the segmentation model may include a first segmentation model 420 and a second segmentation model 440.

[0093] like Figure 4As shown, the input to the first segmentation model 420 can be PET image 410-1, CT image 410-2, and first reference image 410-3, and the output can be an initial segmentation map 430. The initial segmentation map 430 includes initial pathological segmentation results and initial physiological segmentation results corresponding to the elements in PET image 410-1. The initial pathological segmentation result refers to the segmentation result of the first segmentation model for elements belonging to pathological hypermetabolism in PET image 410-1. The initial physiological segmentation result refers to the segmentation result of the first segmentation model for elements belonging to physiological hypermetabolism in PET image 410-1. For example, in the initial segmentation map, elements corresponding to pathological hypermetabolism are predicted as 1 (i.e., the initial pathological segmentation result), elements corresponding to physiological hypermetabolism are predicted as 2 (i.e., the initial physiological segmentation result), and other elements are predicted as 0.

[0094] Refer again Figure 4 The second segmentation model 440 can take PET image 410-1, CT image 410-2, and a second reference image 410-4 as inputs, and output a target segmentation result 450. The target segmentation result 450 can include the final pathological segmentation result and the physiological segmentation result. In some embodiments, the second reference image 410-4 can be determined based on the initial pathological segmentation result contained in the initial segmentation image 430. For example, the initial physiological segmentation result in the initial segmentation image can be removed, and the initial pathological segmentation result can be retained to generate the second reference image. As an example only, if the element corresponding to high pathological uptake in the initial segmentation image 430 is predicted as 1, the element corresponding to high physiological uptake is predicted as 2, and other elements are predicted as 0, then the prediction result of the element corresponding to high physiological uptake in the initial segmentation image 430 can also be modified to 0, thereby generating the second reference image 410-4.

[0095] In some embodiments, the pathological portion of the target segmentation result 450 can be extracted to obtain the lesion segmentation result, which is then post-processed. For example, segmentation results belonging to the brain region in the lesion segmentation results of the head and neck area can be removed to avoid misclassifying physiologically high-uptake areas on the brain as lesions. Another example is removing lesion connected regions with a maximum SUV value of less than 2.5.

[0096] It should be understood that Figure 4 The segmentation model and the generation process of the target segmentation result shown are for illustrative purposes only. In some embodiments, the segmentation model may include two or more cascaded sub-models. In some embodiments, the inputs and outputs of the first segmentation model 420 and the second segmentation model 440 can be set according to actual needs. For example, CT image 410-2 may not be used as input to the first segmentation model 420 and / or the second segmentation model 440. As another example, the initial segmentation map 430 output by the first segmentation model 420 may only include the initial pathological segmentation result.

[0097] In some embodiments, the parameters of the segmentation model can be determined through training, in which the weight of the loss term corresponding to the pathological label is greater than the weight of the loss term corresponding to the non-pathological label.

[0098] Pathological labels refer to the annotation results of elements in the training sample PET images that belong to pathological hypermetabolism. For example, a pathological label could be to mark elements belonging to pathological regions in the sample PET image as 1. Non-pathological labels refer to the annotation results of elements in non-pathological regions in the sample PET image. For example, non-pathological labels can include background labels, physiological labels, etc. Physiological labels refer to the annotation results of elements in the sample PET image that belong to physiological hypermetabolism. For example, background elements could be marked as 0, and physiological hypermetabolism elements as 2. In some embodiments, pathological and non-pathological labels can be manually annotated.

[0099] In some embodiments, when training a segmentation model (e.g., a first segmentation model), the weight of the loss term for the pathology label can be greater than the weight of the loss terms for the background label and the physiological label. For example, the weights of the loss terms for the pathology label, physiological label, and background label can be set to 3 / 5, 1 / 5, and 1 / 5, or 8 / 10, 1 / 10, and 1 / 10, or 18 / 20, 1 / 20, and 1 / 20; or 28 / 30, 1 / 30, and 1 / 30, etc., and this specification does not impose any limitations on these settings.

[0100] In some embodiments of this specification, by setting the weight of the loss term for pathological labels to be greater than the weight of the loss term for non-pathological labels, the model training can focus more on learning pathology, avoiding the bias towards physiological learning caused by setting the same weight for different labels in actual PET images where physiology is more prevalent than pathology.

[0101] For example, in the segmentation labels of head and neck PET images, there are many physiological labels (e.g., almost everyone has high uptake in the brain, oropharynx, etc.), while malignant tumors or lymph nodes in the head and neck are occasional lesions. This results in a situation where physiological labels predominate in the segmentation labels, while pathological labels are few. If the weights of the loss terms for different labels are set to 1 / 3, 1 / 3, 1 / 3 (i.e., the weights of the loss terms for background labels, pathological labels, and physiological labels are set to the same), the convergence direction of the segmentation model during training is prone to bias towards physiological labels. However, the purpose of training the segmentation model in this application is to learn from pathology so that the obtained segmentation model can predict the location of pathological conditions. Therefore, to avoid training bias in the segmentation model due to the scarcity of pathological labels in PET images, the weight of the loss term for pathological labels is increased during training.

[0102] In some embodiments, when training the first segmentation model, multiple different loss term weights for pathological labels can be set to test the recall and precision of the first segmentation model for lesions as the loss term weights for pathological labels gradually increase. For example, the loss term weights for pathological labels, physiological labels, and background labels can be set to (8 / 10, 1 / 10, 1 / 10), (18 / 20, 1 / 20, 1 / 20), (28 / 30, 1 / 30, 1 / 30), etc., to train the first segmentation model. Based on the lesion detection of the separately trained first segmentation models, it was found that as the loss term weights for pathological labels increase during training, the recall and precision of the trained first segmentation model for lesions gradually increase.

[0103] In some embodiments, when training the second segmentation model, multiple different loss term weights for pathological labels can be set. For example, the loss term weights for pathological labels, physiological labels, and background labels can be set to (1 / 3, 1 / 3, 1 / 3), (3 / 5, 1 / 5, 1 / 5), (8 / 10, 1 / 10, 1 / 10), (18 / 20, 1 / 20, 1 / 20), (28 / 30, 1 / 30, 1 / 30), etc., when training the second segmentation model with different loss term weights, and performing lesion detection, it was found that as the loss term weights for pathological labels increased during training, the recall and precision of the trained second segmentation model for lesions gradually increased.

[0104] Therefore, some embodiments of this specification can improve the recall and precision of the segmentation model for lesions by increasing the weight of the loss term for pathological labels during segmentation model training.

[0105] In some embodiments of this specification, a cascaded segmentation network is used to determine the target segmentation result, and the lesion segmentation result is determined based on the target segmentation result. The input to the second segmentation network is generated based on the pathological segmentation result in the output of the first segmentation network. This can improve the second segmentation network's attention to the false positive pathological segmentation results generated by the first segmentation network, reduce false positives in lesion detection, and improve the accuracy of lesion detection. Simultaneously, when training the segmentation model, setting the weight of the pathological label loss term to be greater than the weight of the non-pathological label loss term allows the model to focus more on pathological learning during training, resulting in a better recall and precision in lesion detection.

[0106] Figure 5 This is an exemplary schematic diagram of the generation of a segmentation model according to some embodiments of this specification.

[0107] In some embodiments, such as Figure 5 As shown, the first initial model 520 can be trained using the first training samples to generate the first segmentation model 540.

[0108] The first training sample may include multiple sets of medical image data and corresponding labels for training the first segmentation model. In some embodiments, the first training sample may include a sample PET image 510-1, a sample CT image 510-2, a sample first reference image 510-3, and a first training label 510-4.

[0109] Sample PET image 510-1 can be obtained by scanning the sample object using a PET device. Sample CT image 510-2 can be obtained by scanning the sample object using a CT device. In some embodiments, sample PET image 510-1 and sample CT image 510-2 can be acquired through a storage device or from the corresponding imaging device. In some embodiments, sample CT image 510-2 and sample PET image 510-1 in the same set of training samples are acquired simultaneously based on the same sample object.

[0110] The first reference image 510-3 can be obtained by classifying the elements in the sample PET image, reflecting the categories of elements in the sample PET image. For example, the first reference image can reflect which elements in the sample PET image correspond to physical points of high metabolism (also known as high uptake). In some embodiments, the first reference image 510-3 can be obtained in the same way as the first reference image, as detailed in [reference needed]. Figure 3 The details and related descriptions will not be repeated here.

[0111] The first training label 510-4 is used to train the first segmentation model. In some embodiments, the first training label may include one or more of the following: background label, pathological label, and physiological label annotated on the sample PET image 510-1. In some embodiments, the first training label may be obtained through image analysis or through manual annotation.

[0112] In some embodiments, the first training labels may include pathological labels and physiological labels. In some embodiments, pathological labels and physiological labels may be generated based on a first reference image of the sample and manually annotated pathological labels. Figure 6b As shown, the sample schematic diagram 620-1 of the first reference image includes multiple connected components that may belong to lesions (the black areas in the borders F, G, and H in the figure). Sample schematic diagram 620-1 refers to the schematic diagram obtained by overlaying the first reference image (such as a binary image) onto the sample PET image. Sample schematic diagram 620-1 can be presented to the user, allowing the user to label connected components that are pathological, thereby determining the pathological label. Other connected components in sample schematic diagram 620-1 that are not labeled as pathological by the user can be considered as physiological connected components, thus determining the physiological label. For example, assuming the user labels connected component G as a pathological connected component, a physiological label can be generated. Figure 6b The pathological labels (i.e., the dark area in box I of 620-2) are used to define physiological labels. In these labels, the element values ​​of elements within pathological connected components are marked as 1. Labeling all connected components except G as physiological connected components generates physiological labels (i.e., the light area in box J of 620-2). In these labels, the element values ​​of elements within physiological connected components are marked as 2. By determining the pathological and physiological labels, the first training label 620-2 is obtained; that is, the first training label 620-2 contains both pathological and physiological labels.

[0113] Refer again Figure 4 In some embodiments, the sample PET image 510-1, sample CT image 510-2, and sample first reference image 510-3 can be input into the first initial model 520 to obtain a reference segmentation image 530 output by the first initial model 520. The reference segmentation image 530 refers to the target segmentation result reference image corresponding to the sample obtained by processing the sample through the first initial model. The reference segmentation image 530 and... Figure 3 Similar to the initial segmentation map, the initial segmentation map may include pathological and physiological segmentation results corresponding to elements in the PET image, while the reference segmentation map 530 may include sample pathological and physiological segmentation results corresponding to elements in the sample PET image 510-1. The value of the loss function is calculated based on the reference segmentation map 530 and the first training label 510-4, where the value of the loss function is related to the weights of the loss terms for the pathological and physiological labels. For information on setting the weights of the loss terms for the pathological and physiological labels, please refer to [link to relevant documentation]. Figure 4 And related descriptions. Based on the value of the loss function, the parameters of the first initial model 520 are updated through gradient descent or other methods until a preset condition is met, at which point training is complete, resulting in the trained first segmentation model 540. The preset condition can be loss function convergence, or reaching the maximum number of iterations, etc.

[0114] In some embodiments, before inputting the first training sample into the first initial model 520, the sample PET image 510-1, sample CT image 510-2, and sample first reference image 510-3 need to be preprocessed. The preprocessing method is the same as that for the PET image, CT image, and first reference image; for details, please refer to [link to relevant documentation]. Figure 3 The details and related descriptions will not be repeated here.

[0115] In some embodiments, after the first segmentation model 540 is trained, its performance can be tested using a test set. For example, tests can be conducted at both the patient and lesion levels, using the first segmentation model 540 to process the test set data and statistically analyze the lesion detection results. Analysis of the test results reveals that the first segmentation model 540 has high sensitivity to lesions but is prone to producing false positives. Therefore, in some embodiments of this specification, a second segmentation model 570 cascaded with the first segmentation model 540 is trained to learn from the false positives produced by the first segmentation model 540, enabling the trained second segmentation model 570 to reduce false positives in lesion segmentation results and improve the accuracy of lesion detection.

[0116] After the first segmentation model is generated, a second training sample can be generated based on the first training sample and the first segmentation model. Furthermore, a second initial model 550 can be trained using the second training sample to generate a second segmentation model 570.

[0117] The second training samples may include multiple sets of image data and corresponding labels used to train the second segmentation model 570. In some embodiments, such as Figure 5 As shown, the second training sample may include sample PET image 510-1, sample CT image 510-2, sample second reference image 510-5, and second training label 510-6.

[0118] In some embodiments, a second training sample can be generated based on the first training sample and the first segmentation model. In some embodiments, a second reference map 510-5 and a second training label 510-6 in the second training sample can be generated based on the first training label 510-4 in the first training sample and the reference segmentation map generated by the first segmentation model.

[0119] In some embodiments, a first segmentation model 540 can be used to process the sample PET image 510-1, the sample CT image 510-2, and the sample first reference image 510-3 to determine a reference segmentation map. The reference segmentation map includes the sample pathological segmentation results and the sample physiological segmentation results of the elements in the sample PET image 510-1. Further, based on the sample pathological segmentation results in the reference segmentation map and the first training label 510-4, a sample second reference image 510-5 and a second training label 510-6 can be generated.

[0120] For example, based on the sample pathological segmentation results and sample physiological segmentation results in the reference segmentation map, the sample physiological segmentation result (e.g., the element corresponding to physiological hypermetabolism in the sample is predicted as 2) can be modified to 0, while the sample pathological segmentation result (e.g., the element corresponding to pathological hypermetabolism in the sample is predicted as 1) can be retained as the sample second reference map 510-5. As another example, a second training label 510-6 can be generated based on the sample pathological segmentation results in the reference segmentation map 530 and its corresponding first training label 510-4. In some embodiments, based on the pathological labels in the first training label 510-4, false positives in the sample pathological segmentation results in the reference segmentation map 530 can be identified, and the false positive parts in the sample pathological segmentation results can be labeled as physiological labels in the second training label 510-6, while the pathological parts in the first training label 510-4 can be labeled as pathological labels in the second training label 510-6. False positives refer to regions that are judged as pathological regions by the first segmentation model 540 but are actually physiological regions. Figure 6c The diagram shows the generation of the second training label. In the reference segmentation diagram 630-1, the dark area represents the pathological segmentation result of the sample. The second training label 630-2 consists of a physiological label and a pathological label. The physiological label is formed based on the pathological false positive part of the sample pathological segmentation result and is shown as the highlighted area (the bright area in box L) in the figure. The pathological label is formed based on the pathological part of the first training label and is shown as the dark area (the dark area in box K) in the figure.

[0121] Continue to refer to Figure 4 In some embodiments, the sample PET image 510-1, sample CT image 510-2, and sample second reference image 510-5 can be input into the second initial model 550 to obtain the sample target segmentation result 560 output by the second initial model 550. The sample target segmentation result 560 includes the pathological and physiological segmentation results of the sample. The value of the loss function is calculated based on the sample target segmentation result 560 and the second training label 510-6. The value of the loss function is related to the weights of the loss terms for the pathological and physiological labels. For information on setting the weights of the loss terms for the pathological and physiological labels, please refer to [reference needed]. Figure 4 And its related description. Based on the value of the loss function, the parameters of the second initial model 550 are updated through gradient descent or other methods until a preset condition is met, at which point training is complete, resulting in a trained second segmentation model 570. The preset condition can be loss function convergence, or reaching the maximum number of iterations during training, etc.

[0122] In some embodiments, before inputting the second training samples into the second initial model 550, the sample PET image 510-1, sample CT image 510-2, and sample second reference image 510-5 need to be preprocessed. The preprocessing method is the same as that for the PET image, CT image, and first reference image; for details, please refer to [link to relevant documentation]. Figure 3 The details and related descriptions will not be repeated here.

[0123] In some embodiments, the element values ​​in the sample PET images can be truncated with normalization and then transformed without truncation to generate two sets of control samples to train the segmentation model. The two trained segmentation models are then used for lesion detection. Analysis of the lesion detection results shows that, compared to truncating the element values ​​in the sample PET images before inputting them into the segmentation model, the embodiment of this specification uses a non-truncated transformation to process the sample PET images. The resulting segmentation model exhibits more stable recall and precision for lesions, and also achieves higher recall and precision compared to the truncated method.

[0124] In some embodiments, after the first segmentation model and the second segmentation model have been trained, training samples can be added to retrain the first and second segmentation models. In some embodiments, test samples can be added to test the trained first and second segmentation models, and the performance of the segmentation models can be analyzed based on the test results.

[0125] In some embodiments, testing the trained segmentation model may include: preprocessing the test PET image, test CT image, and test first reference image, then inputting them into the trained first segmentation model to obtain a test reference segmentation image; generating a test second reference image based on the test reference segmentation image; preprocessing the test PET image, test CT image, and test second reference image, then inputting them into the trained second segmentation model to output a target segmentation result, which includes pathological segmentation results and physiological segmentation results. In some embodiments, the lesion segmentation result can be determined by extracting the pathological segmentation result from the target segmentation result. More information on determining the lesion segmentation result can be found in [link to relevant documentation]. Figure 3 And related descriptions. The aforementioned samples are all test samples, which can be obtained in the same way as the training samples. Preprocessing here refers to performing a non-truncated mapping transformation on the test PET image, and normalizing and truncating the test CT image, the first test reference image, and the second test reference image. For specific preprocessing methods, please refer to [link to relevant documentation]. Figure 3 And its related descriptions.

[0126] Statistical analysis of the test results of the first segmentation model revealed that, while the recall rate of the second segmentation model was slightly lower than that of the first model at the lesion level, the precision rate was significantly improved. At the patient level, the recall rate of the second segmentation model was essentially the same as that of the first model, while the precision rate was significantly improved. These analyses demonstrate that, at both the lesion and patient levels, the second segmentation model can reduce false positives in lesion segmentation results and improve the accuracy of lesion detection.

[0127] Some embodiments in this specification demonstrate that by training a cascaded segmentation model for lesion detection, the attention range of the segmentation model can be narrowed, false positives in lesion segmentation results can be reduced, and the accuracy of lesion detection can be improved.

[0128] Some embodiments of this specification also provide a PET lesion detection device, including at least one processor and at least one memory, wherein the at least one memory is used to store computer instructions; and the at least one processor is used to execute at least some of the computer instructions to implement the PET lesion detection method disclosed in some embodiments of this specification.

[0129] The beneficial effects that the embodiments of this specification may bring include, but are not limited to: (1) by training the cascaded segmentation network model, the model's learning of false positives in lesion segmentation results is enhanced, which can reduce false positives in lesion detection and improve the accuracy of lesion detection; (2) using a first reference image and CT images to assist PET lesion detection. The first reference image may include information on high metabolic physical points of the scanned object. By introducing the first reference image, attention to high metabolic physical points during lesion detection is improved, which can improve the lesion recall rate. Introducing CT images can utilize the more anatomical information contained in CT images (e.g., clearer lesion boundaries), enabling better differentiation between pathological and physiological regions during lesion detection, thereby improving the accuracy of lesion detection; (3) During lesion detection, SUV value information exceeding the preset threshold in PET images is retained, preserving the clinical significance of SUV values ​​in medical diagnosis and improving the accuracy of lesion detection; (4) When training the model, by setting a greater weight for the loss term of pathological labels than that of non-pathological labels, the model can avoid bias towards physiological learning when there are more physiological labels than pathological labels in the training samples, thereby strengthening the model's learning of pathology, improving the performance of the lesion segmentation model, and further improving the recall and precision of lesion detection.

[0130] The basic concepts have been described above. Obviously, for those skilled in the art, the detailed disclosure above is merely illustrative and does not constitute a limitation of this specification. Although not explicitly stated herein, those skilled in the art may make various modifications, improvements, and corrections to this specification. Such modifications, improvements, and corrections are suggested in this specification and therefore remain within the spirit and scope of the exemplary embodiments described herein.

[0131] Furthermore, this specification uses specific terms to describe embodiments thereof. For example, "an embodiment," "one embodiment," and / or "some embodiments" refer to a particular feature, structure, or characteristic associated with at least one embodiment of this specification. Therefore, it should be emphasized and noted that references to "an embodiment," "one embodiment," or "an alternative embodiment" in different locations throughout this specification do not necessarily refer to the same embodiment. Moreover, certain features, structures, or characteristics in one or more embodiments of this specification can be appropriately combined.

[0132] Furthermore, unless expressly stated in the claims, the order of processing elements and sequences, the use of numbers and letters, or other names described in this specification are not intended to limit the order of the processes and methods described herein. Although various examples have been discussed in the foregoing disclosure of some embodiments of the invention that are currently considered useful, it should be understood that such details are for illustrative purposes only, and the appended claims are not limited to the disclosed embodiments; rather, the claims are intended to cover all modifications and equivalent combinations that conform to the spirit and scope of the embodiments described herein. For example, while the system components described above can be implemented using hardware devices, they can also be implemented solely using software solutions, such as installing the described system on existing servers or mobile devices.

[0133] Similarly, it should be noted that, in order to simplify the description disclosed herein and thus aid in the understanding of one or more embodiments of the invention, the foregoing description of embodiments in this specification may sometimes combine multiple features into a single embodiment, drawing, or description thereof. However, this method of disclosure does not imply that the subject matter of this specification requires more features than those mentioned in the claims. In fact, the embodiments contain fewer features than all the features of a single embodiment disclosed above.

[0134] In some embodiments, numbers describing the quantity of components and attributes are used. It should be understood that such numbers used in the description of embodiments are modified in some examples with the terms "approximately," "approximately," or "generally." Unless otherwise stated, "approximately," "approximately," or "generally" indicates that the numbers are allowed to vary by ±20%. Accordingly, in some embodiments, the numerical parameters used in the specification and claims are approximate values, which may be changed depending on the characteristics required by individual embodiments. In some embodiments, numerical parameters should take into account specified significant digits and employ a general method of digit reservation. Although the numerical ranges and parameters used to confirm their breadth of range in some embodiments of this specification are approximate values, in specific embodiments, such values ​​are set as precisely as feasible.

[0135] For each patent, patent application, patent application publication, and other material, such as articles, books, specifications, publications, and documents, referenced in this specification, the entire contents of which are incorporated herein by reference. This excludes historical application documents that are inconsistent with or conflict with the content of this specification, as well as documents that limit the broadest scope of the claims in this specification (currently or subsequently appended to this specification). It should be noted that in the event of any inconsistency or conflict between the descriptions, definitions, and / or terminology used in the supplementary materials to this specification and the content of this specification, the descriptions, definitions, and / or terminology used in this specification shall prevail.

[0136] Finally, it should be understood that the embodiments described in this specification are merely illustrative of the principles of the embodiments described herein. Other variations may also fall within the scope of this specification. Therefore, alternative configurations of the embodiments described herein are intended to be illustrative rather than limiting, and should be considered consistent with the teachings of this specification. Accordingly, the embodiments described herein are not limited to those explicitly introduced and described herein.

Claims

1. A method for detecting lesions using positron emission tomography (PET), characterized in that, The method includes: The elements in the PET image of the scanned object are classified to generate a first reference image; Based on the PET image, the CT image of the scanned object, and the first reference image, a target segmentation result is determined using a segmentation model, wherein the segmentation model is a machine learning model; and Based on the target segmentation result, the lesion segmentation result is determined, wherein the segmentation model includes a first segmentation model and a second segmentation model. The first segmentation model determines an initial segmentation map based on the PET image, the CT image, and the first reference image. The initial segmentation map includes the initial pathological segmentation result and the initial physiological segmentation result corresponding to the elements in the PET image. The initial pathological segmentation result is the segmentation result of the elements in the PET image that belong to pathological hypermetabolism, and the initial physiological segmentation result is the segmentation result of the elements in the PET image that belong to physiological hypermetabolism. Remove the initial physiological segmentation results from the initial segmentation image, and retain the initial pathological segmentation results to generate a second reference image; The second segmentation model determines the target segmentation result based on the PET image, the CT image, and the second reference image.

2. The method according to claim 1, characterized in that, The parameters of the segmentation model are determined through training using the following process: A first initial model is trained using the first training samples to generate the first segmentation model; Based on the first training sample and the first segmentation model, a second training sample is generated; as well as The second initial model is trained using the second training samples to generate the second segmentation model.

3. The method according to claim 2, characterized in that, The first training sample includes a sample PET image, a sample CT image, a sample first reference image, and a first training label; the second training sample includes the sample PET image, the sample CT image, a sample second reference image, and a second training label. The step of generating the second training sample based on the first training sample and the first segmentation model includes: The sample PET image, the sample CT image, and the sample first reference image are input into the first segmentation model to determine the reference segmentation image, which includes the sample pathological segmentation result and the sample physiological segmentation result of the elements in the sample PET image; Based on the pathological segmentation results of the sample in the reference segmentation image and the first training label, a second reference image of the sample and a second training label are generated.

4. The method according to claim 1, characterized in that, The step of classifying elements in the PET image of the scanned object to generate a first reference image includes: Obtain standardized intake thresholds; The elements in the PET image are classified based on a standardized uptake value threshold to determine the first reference image.

5. The method according to claim 1, characterized in that, Also includes: Before inputting the segmentation model, a mapping transformation is performed on the value of each element in the PET image to determine the corresponding mapping value for each element. When the value of an element is greater than a preset value, the mapping value of the element is positively correlated with the value of the element. When the value of the element is less than or equal to the preset value, the mapping value of the element is fixed to the preset value.

6. The method according to claim 5, characterized in that, The parameters of the segmentation model are determined through training, in which the weight of the loss term corresponding to the pathological label is greater than the weight of the loss term corresponding to the non-pathological label.

7. The method according to claim 1, characterized in that, Also includes: The lesion segmentation results were quantitatively analyzed.

8. The method according to claim 4, characterized in that, The elements in the PET image are classified based on a standardized uptake value threshold to determine that the first reference image includes: In response to an element’s normalized uptake value being not less than the normalized uptake value threshold, the element is classified into the first class corresponding to the high uptake physical point. In response to an element’s normalized uptake value being less than the normalized uptake value threshold, the element is classified into a second class corresponding to a low uptake physical point.

9. A positron emission tomography (PET) lesion detection system, characterized in that, The system includes: A generation module is used to classify elements in the PET image of the scanned object to generate a first reference image; and The first determining module is used to determine the target segmentation result based on the PET image, the CT image of the scanned object and the first reference image, by using a segmentation model, wherein the segmentation model is a machine learning model. The second determining module is used to determine the lesion segmentation result based on the target segmentation result, wherein the segmentation model includes a first segmentation model and a second segmentation model. The first segmentation model determines an initial segmentation map based on the PET image, the CT image, and the first reference image. The initial segmentation map includes the initial pathological segmentation result and the initial physiological segmentation result corresponding to the elements in the PET image. The initial pathological segmentation result is the segmentation result of the elements in the PET image that belong to pathological hypermetabolism, and the initial physiological segmentation result is the segmentation result of the elements in the PET image that belong to physiological hypermetabolism. Remove the initial physiological segmentation results from the initial segmentation image, and retain the initial pathological segmentation results to generate a second reference image; The second segmentation model determines the target segmentation result based on the PET image, the CT image, and the second reference image.

10. A positron emission tomography (PET) lesion detection device, characterized in that, The apparatus includes at least one processor and at least one storage device, the storage device being used to store instructions that, when executed by the at least one processor, implement the method as described in any one of claims 1 to 8.

Images