Medical data processing device, medical data processing method, medical image diagnostic device, and program

Abstract

To obtain a trained model with excellent accuracy.SOLUTION: A medical data processing device according to an embodiment includes: a candidate data obtaining unit; a selection unit; an adjustment unit; and a model training unit. The candidate data obtaining unit obtains candidate data. The selection unit, using a trained initial task model obtained using existing data, obtains characteristics of the candidate data, selects the candidate data that matches a condition based on the characteristics, and obtains a selection data set including the selected candidate data. The adjustment unit performs a characteristics relation analysis on the existing data set and the selected data set, and adjusts a training policy and a pre-processing plan based on the analysis result. The model training unit, using a training data set made up of the existing data set and the selected data set, performs model training to obtain a trained model based on the adjusted training policy and the adjusted pre-processing plan.SELECTED DRAWING: Figure 1

Description

[Technical field]

[0001] The embodiments disclosed in this specification and the drawings relate to a medical data processing device, a medical data processing method, a medical image diagnostic device, and a program. [Background technology]

[0002] Traditionally, in the field of medical data processing, including medical image processing, machine learning and deep learning technologies have been widely used to perform various tasks such as segmentation tasks, classification tasks, detection tasks, image generation tasks, etc., depending on the research objective. When performing the above-mentioned various tasks using machine learning and deep learning technologies, a large amount of medical data is required to maintain high accuracy of the trained model, and a large amount of high-quality annotation is required for medical images.

[0003] In reality, different research objectives require different tasks and different medical data to be collected, and it is difficult to collect a large amount of medical data for a certain task. If the medical data is insufficient, the accuracy of the trained model will be low, and the quality of the results obtained by performing the task will be reduced.

[0004] In order to ensure sufficient medical data when applying machine learning or deep learning technologies to perform a task, based on the current situation that similar applicable features may exist in medical data of different lesions, when performing a task using machine learning or deep learning technologies, it is conceivable to expand the training dataset for performing the task by selecting usable medical data containing similar features from medical data of different tasks.

[0005] Conventional technology involves adding medical data for a different task to an existing dataset for performing a certain task. However, using medical data from a different task for another task results in more usable medical data, but at the same time introduces a lot of unusable medical data, which affects the accuracy of a trained model for a certain task.

[0006] There is also a conventional technique of selecting medical data for different tasks and adding the selected data to the training dataset as is to expand the training dataset. However, even if the data is selected, there are still distribution differences between the existing data and the selected data, and adding data from different sources as is may affect the accuracy of the original trained model for a certain task.

[0007] In addition, when it comes to annotating medical image data, there are certain differences in the judgment criteria used by different doctors, so annotations of medical images can vary depending on the doctor's level of experience, and annotating large amounts of medical image data takes a very long time, affecting the real-time nature and practicality of the medical data processing process.

[0008] Therefore, there is a need for a medical data processing device that can overcome the problems of the above-mentioned conventional techniques, select appropriate medical data from medical data of different tasks or different lesions while taking into consideration the characteristics of the medical data and the annotation quality, secure sufficient medical data for each task, and ensure the accuracy of a model obtained by performing training based on a training data set including the selected appropriate medical data. In this way, it is desirable to obtain a trained model with good accuracy. [Prior art documents] [Patent documents]

[0009] [Patent Document 1] JP 2005-301840 A Summary of the Invention [Problem to be solved by the invention]

[0010] One of the problems to be solved by the embodiments disclosed in this specification and the drawings is to obtain a trained model with good accuracy. However, the problems to be solved by the embodiments disclosed in this specification and the drawings are not limited to the above problem. Problems corresponding to the effects of each configuration shown in the embodiments described later can also be positioned as other problems. [Means for solving the problem]

[0011] A medical data processing device according to an embodiment includes an initial task model training unit, a candidate data acquisition unit, a selection unit, an adjustment unit, and a model training unit. The initial task model training unit trains an initial task model using an existing data set to obtain a trained initial task model. The candidate data acquisition unit acquires candidate data. The selection unit acquires features of the candidate data using the trained initial task model, selects candidate data that meets a condition based on the features, and obtains a selection data set including the selected candidate data. The adjustment unit performs feature relationship analysis on the existing data set and the selection data set, and adjusts a training policy and a pre-processing plan based on the analysis result. The model training unit performs model training based on the adjusted training policy and the adjusted pre-processing plan using a training data set consisting of the existing data set and the selection data set, and obtains a trained model. [Brief description of the drawings]

[0012] [Figure 1] 1 is a diagram showing an example of the configuration of an X-ray computed tomography apparatus equipped with a medical data processing apparatus according to a first embodiment. [Diagram 2] 4 is a flowchart of an example of an operation of the medical data processing device according to the first embodiment. [Diagram 3]4 is a flowchart of an example of a selection function of the medical data processing device according to the first embodiment. [Figure 4] FIG. 2 is a diagram illustrating an example of a selection function of the medical data processing apparatus according to the first embodiment. [Diagram 5] 4 is a schematic diagram of another specific example of the selection function of the medical data processing device according to the first embodiment. FIG. [Figure 6] 4 is a flowchart of an example of an adjustment function of the medical data processing device according to the first embodiment. [Figure 7] 3 is a schematic diagram of a specific example of an adjustment function of the medical data processing device according to the first embodiment. FIG. [Figure 8] 5 is a schematic diagram of another specific example of the adjustment function of the medical data processing device according to the first embodiment. FIG. [Figure 9] 4 is a diagram showing a specific example of adjustment of a pre-processing plan by an adjustment function of the medical data processing device according to the first embodiment; FIG. [Figure 10] 4 is a schematic diagram of a specific example of adjustment of a training policy by an adjustment function of the medical data processing device according to the first embodiment. FIG. [Figure 11] 4 is a schematic diagram of a specific example of adjustment of a training policy by an adjustment function of the medical data processing device according to the first embodiment. FIG. [Figure 12] 10 is a flowchart of an example of an operation of the medical data processing device according to the second embodiment. [Figure 13] 1 is a schematic diagram comparing the results of an embodiment and a method of the prior art; DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0013] Hereinafter, a medical data processing device, a medical data processing method, a medical image diagnostic device, and a storage medium according to each embodiment will be described with reference to the drawings. In the following description, components having the same or substantially the same functions as those already described with reference to the previously mentioned drawings will be given the same reference numerals and will be described repeatedly only when necessary. Even when the same parts are shown, the dimensions and ratios of the components may be different depending on the drawing. In addition, for example, from the viewpoint of ensuring the visibility of the drawings, reference numerals may be given only to main or representative components in the description of each drawing, and reference numerals may not be given to components having the same or substantially the same functions.

[0014] In each embodiment described below, a case where the medical data processing device according to each embodiment is mounted on an X-ray computed tomography (CT (Computed Tomography) device will be exemplified. Note that the medical data processing device according to each embodiment is not limited to being mounted on an X-ray CT device, and may be realized as an independent device by a computer having a processor and memories such as a ROM (Read Only Memory) and a RAM (Random Access Memory) as hardware resources. In this case, the processor mounted on the computer can realize various functions according to each embodiment by executing a program read from the ROM or the like and loaded into the RAM.

[0015] Furthermore, the medical data processing device according to each embodiment may be implemented in a medical image diagnostic device other than an X-ray CT device. In this case, a processor installed in each medical image diagnostic device can implement the functions according to each embodiment by executing a program read from a ROM or the like and loaded into a RAM. As other medical image diagnostic devices, various medical image diagnostic devices may be used, such as an X-ray diagnostic device, an MRI (Magnetic Resonance Imaging) device, an ultrasound diagnostic device, a SPECT (Single Photon Emission Computed Tomography) device, a PET (Positron Emission computed Tomography) device, a SPECT-CT device in which a SPECT device and an X-ray CT device are integrated, and a PET-CT device in which a PET device and an X-ray CT device are integrated.

[0016] For example, there are various types of X-ray CT devices, such as third-generation CT and fourth-generation CT, and any of these types can be applied to each embodiment. Here, the third-generation CT is a rotate / rotate-type in which the X-ray tube and the detector rotate around the subject as a unit. The fourth-generation CT is a stationary / rotate-type in which a large number of X-ray detection elements arranged in a ring shape are fixed, and only the X-ray tube rotates around the subject.

[0017] (First embodiment) 1 is a diagram showing an example of the configuration of an X-ray CT apparatus 1 equipped with a medical data processing apparatus according to the first embodiment. The X-ray CT apparatus 1 irradiates an object P with X-rays from an X-ray tube 11, and detects the irradiated X-rays with an X-ray detector 12. The X-ray CT apparatus 1 generates a CT image of the object P based on an output from the X-ray detector 12.

[0018] As shown in FIG. 1, the X-ray CT apparatus 1 includes a gantry 10, a bed 30, and a console 40. For convenience of explanation, a plurality of gantry 10 are depicted in FIG. 1. The gantry 10 is a scanning apparatus having a configuration for performing X-ray CT imaging of a subject P. The bed 30 is a transport device for placing the subject P to be subjected to X-ray CT imaging and positioning the subject P. The console 40 is a computer for controlling the gantry 10. For example, the gantry 10 and the bed 30 are installed in a CT examination room, and the console 40 is installed in a control room adjacent to the CT examination room. The gantry 10, the bed 30, and the console 40 are connected to each other by wire or wirelessly so as to be able to communicate with each other.

[0019] The console 40 does not necessarily have to be installed in a control room. For example, the console 40 may be installed in the same room as the gantry 10 and the bed 30. Also, the console 40 may be incorporated in the gantry 10.

[0020] In this embodiment, the rotation axis of the rotating frame 13 in the non-tilted state or the longitudinal direction of the top plate 33 of the bed 30 is defined as the Z-axis direction, the axial direction that is perpendicular to the Z-axis direction and horizontal to the floor surface is defined as the X-axis direction, and the axial direction that is perpendicular to the Z-axis direction and perpendicular to the floor surface is defined as the Y-axis direction.

[0021] As shown in FIG. 1, the gantry 10 has an X-ray tube 11, an X-ray detector 12, a rotating frame 13, an X-ray high voltage device 14, a control device 15, a wedge 16, a collimator 17, and a data acquisition circuit (Data Acquisition System: DAS) 18.

[0022] The X-ray tube 11 is a vacuum tube having a cathode (filament) that generates thermoelectrons and an anode (target) that generates X-rays upon impact of the thermoelectrons. The X-ray tube 11 irradiates the subject P with X-rays by irradiating thermoelectrons from the cathode to the anode using a high voltage supplied from an X-ray high voltage device 14.

[0023] It should be noted that the hardware for generating X-rays is not limited to the X-ray tube 11. For example, instead of the X-ray tube 11, X-rays may be generated using a fifth generation system. The fifth generation system includes a focus coil for focusing the electron beam generated from the electron gun, a deflection coil for electromagnetic deflection, and a target ring that surrounds half of the subject P and generates X-rays when the deflected electron beam collides with the target ring.

[0024] The X-ray detector 12 detects X-rays irradiated from the X-ray tube 11 and passing through the subject P, and outputs an electric signal corresponding to the detected X-ray dose to the DAS 18. The X-ray detector 12 has, for example, an X-ray detection element row in which a plurality of X-ray detection elements are arranged in the channel direction along one arc with the focus of the X-ray tube 11 as the center. The X-ray detector 12 has, for example, a structure in which a plurality of X-ray detection elements are arranged in the slice direction (column direction, row direction) in the channel direction. The X-ray detector 12 is an indirect conversion type detector having, for example, a grid, a scintillator array, and a photosensor array. The scintillator array has a plurality of scintillators. The scintillator has a scintillator crystal that outputs a light amount according to the amount of incident X-ray. The grid is arranged on the surface of the scintillator array on the X-ray incidence side, and has an X-ray shielding plate that has a function of absorbing scattered X-rays. The grid may also be called a collimator (one-dimensional collimator or two-dimensional collimator). The photosensor array has a function of converting light from the scintillator into an electrical signal according to the amount of light. For example, a photomultiplier tube (PMT) is used as the photosensor. The X-ray detector 12 may be a direct conversion type detector having a semiconductor element that converts the incident X-rays into an electrical signal.

[0025] The rotating frame 13 is an annular frame that supports the X-ray tube 11 and the X-ray detector 12 facing each other and rotates the X-ray tube 11 and the X-ray detector 12 by a control device 15 described later. An image field of view (FOV) is set in an opening 19 of the rotating frame 13. For example, the rotating frame 13 is a casting made of aluminum. Note that the rotating frame 13 can further support an X-ray high voltage device 14, a wedge 16, a collimator 17, a DAS 18, and the like in addition to the X-ray tube 11 and the X-ray detector 12. The rotating frame 13 can also further support various components not shown in FIG. 1.

[0026] The X-ray high voltage device 14 has a high voltage generator and an X-ray control device. The high voltage generator has electric circuits such as a transformer and a rectifier, and generates a high voltage to be applied to the X-ray tube 11 and a filament current to be supplied to the X-ray tube 11. The X-ray control device controls the output voltage according to the X-rays irradiated by the X-ray tube 11. The high voltage generator may be of a transformer type or an inverter type. The X-ray high voltage device 14 may be provided on the rotating frame 13 in the gantry 10, or may be provided on a fixed frame (not shown) in the gantry 10. The fixed frame is a frame that rotatably supports the rotating frame 13.

[0027] The control device 15 includes a driving mechanism such as a motor and an actuator, and a processing circuit having a processor, memory, etc. for controlling the driving mechanism. The control device 15 receives input signals from the input interface 43 and an input interface provided on the gantry 10, and controls the operation of the gantry 10 and the bed 30. Examples of the operation control by the control device 15 include control to rotate the rotating frame 13, control to tilt the gantry 10, and control to operate the bed 30. The control to tilt the gantry 10 is realized by the control device 15 rotating the rotating frame 13 around an axis parallel to the X-axis direction based on inclination angle information inputted by an input interface attached to the gantry 10. The control device 15 may be provided on the gantry 10 or on the console 40.

[0028] The wedge 16 is a filter for adjusting the amount of X-rays irradiated from the X-ray tube 11. Specifically, the wedge 16 is a filter that transmits and attenuates the X-rays irradiated from the X-ray tube 11 so that the X-rays irradiated from the X-ray tube 11 to the subject P have a predetermined distribution. For example, the wedge 16 is a wedge filter or a bow-tie filter, and is formed by processing aluminum or the like to have a predetermined target angle and a predetermined thickness.

[0029] The collimator 17 limits the irradiation range of the X-rays that have passed through the wedge 16. The collimator 17 slidably supports multiple lead plates that block X-rays, and adjusts the shape of the slits formed by the multiple lead plates. The collimator 17 is sometimes called an X-ray aperture. The DAS 18 reads out an electrical signal from the X-ray detector 12 corresponding to the dose of X-rays detected by the X-ray detector 12. The DAS 18 amplifies the read out electrical signal and integrates (adds up) the electrical signal over a view period to collect detection data having a digital value corresponding to the dose of X-rays over the view period. The detection data is called projection data. The DAS 18 is realized, for example, by an Application Specific Integrated Circuit (ASIC) equipped with circuit elements capable of generating projection data. The projection data is transmitted to the console 40 via a non-contact data transmission device or the like.

[0030] The detection data generated by the DAS 18 is transmitted by optical communication from a transmitter having a light emitting diode (LED) provided on the rotating frame 13 to a receiver having a photodiode provided on a non-rotating portion of the gantry 10 (e.g., a fixed frame; not shown in FIG. 1), and then transferred to the console 40. The method of transmitting the detection data from the rotating frame 13 of the rotating portion to the non-rotating portion of the gantry 10 is not limited to the optical communication described above, and any method of non-contact data transfer may be used.

[0031] In this embodiment, an X-ray CT device 1 equipped with an integral type X-ray detector 12 is described as an example, but the technology according to this embodiment can also be realized as an X-ray CT device 1 equipped with a photon counting type X-ray detector.

[0032] The bed 30 is a device for placing and moving the subject P to be scanned. The bed 30 includes a base 31, a bed drive device 32, a top plate 33, and a support frame 34. The base 31 is a housing for supporting the support frame 34 so as to be movable in the vertical direction. The bed drive device 32 is a drive mechanism for moving the top plate 33 on which the subject P is placed in the longitudinal direction of the top plate 33. The bed drive device 32 includes a motor, an actuator, and the like. The top plate 33 is a plate on which the subject P is placed. The top plate 33 is provided on the upper surface of the support frame 34. The top plate 33 can protrude from the bed 30 toward the gantry 10 so that the whole body of the subject P can be imaged. The top plate 33 is formed of, for example, carbon fiber reinforced plastic (CFRP) having good X-ray transmittance and physical properties such as rigidity and strength. For example, the inside of the tabletop 33 is hollow. The support frame 34 supports the tabletop 33 so as to be movable in the longitudinal direction of the tabletop 33. The bed driving device 32 may move the support frame 34 in the longitudinal direction of the tabletop 33 in addition to the tabletop 33.

[0033] The console 40 has a memory 41, a display 42, an input interface 43, and a processing circuit 44. Data communication between the memory 41, the display 42, the input interface 43, and the processing circuit 44 is performed via a bus. Note that, although the console 40 will be described as being separate from the gantry 10, the gantry 10 may include the console 40 or some of the components of the console 40.

[0034] The memory 41 is realized by, for example, a semiconductor memory element such as a RAM or a flash memory, a hard disk, an optical disk, etc. For example, the memory 41 stores projection data and reconstructed image data. Also, for example, the memory 41 stores various programs. The storage area of ​​the memory 41 may be in the X-ray CT device 1 or in an external storage device connected via a network.

[0035] The display 42 displays various information. For example, the display 42 displays medical images (e.g., CT images) generated by the processing circuit 44, a GUI (Graphical User Interface) for receiving various operations from an operator, and the like. The information displayed on the display 42 includes a medical image including a region of interest, a stretched output image, a polar coordinate output image, a segmented output image, a ring structure, and a quantized numerical value according to the embodiment. As the display 42, various arbitrary displays can be used as appropriate. For example, as the display 42, a liquid crystal display (LCD), a cathode ray tube (CRT) display, an organic electro luminescence display (OLED), or a plasma display can be used.

[0036] The display 42 may be provided anywhere in the control room. The display 42 may be provided on the stand 10. The display 42 may be a desktop type, or may be configured as a tablet terminal or the like capable of wireless communication with the main body of the console 40. One or more projectors may be used as the display 42. Here, the display 42 is an example of a display unit.

[0037] The input interface 43 accepts various input operations from the operator, converts the accepted input operations into electrical signals, and outputs the electrical signals to the processing circuit 44. For example, the input interface 43 accepts from the operator acquisition conditions for acquiring projection data, reconstruction conditions for reconstructing CT images, image processing conditions for generating post-processed images from CT images, etc. In addition, for example, the input interface 43 accepts from the operator calculation conditions for calculating the distribution of lesion amounts relative to anatomical structures from CT image data, etc.

[0038] As the input interface 43, for example, a mouse, a keyboard, a trackball, a switch, a button, a joystick, a touchpad, a touch panel display, and the like can be used as appropriate. In this embodiment, the input interface 43 is not limited to having these physical operation parts. For example, an example of the input interface 43 includes an electrical signal processing circuit that receives an electrical signal corresponding to an input operation from an external input device provided separately from the device and outputs the electrical signal to the processing circuit 44. The input interface 43 may also be provided on the stand 10. The input interface 43 may also be configured with a tablet terminal or the like that can wirelessly communicate with the console 40 main body.

[0039] The processing circuitry 44 controls the overall operation of the X-ray CT apparatus 1. The processing circuitry 44 has a processor and memories such as ROM and RAM as hardware resources. The processing circuitry 44 executes a system control function 441, an image generation function 442, an image processing function 443, an initial task model training function 444, a candidate data acquisition function 445, a selection function 446, an adjustment function 447, a model training function 448, and the like, by the processor that executes a program expanded in the memory.

[0040] In the system control function 441, the processing circuitry 44 can control various functions of the processing circuitry 44 based on an input operation received from an operator via the input interface 43. For example, the processing circuitry 44 can control a CT scan performed on the gantry 10. The processing circuitry 44 can acquire volume data on the subject P using an image generating function 442 and an image processing function 443 described later based on detection data obtained by the CT scan. In this embodiment, as an example, a case will be described in which volume data including at least a lesion region, which is a region of interest, is acquired. Here, the region of interest is a lung region. The lesion may be a lung tumor region or a lung nodule region. The region of interest is not limited to a region of an organ such as the lung, and may be a region of another organ such as the brain or liver.

[0041] The processing circuitry 44 may acquire volume data related to the subject P from outside the X-ray CT apparatus 1. In the present embodiment, for example, a case where volume data obtained by the X-ray CT apparatus 1 is used is illustrated, but volume data obtained by other medical image diagnostic apparatuses may be used.

[0042] In the image generation function 442, the processing circuitry 44 can generate data by performing preprocessing such as logarithmic conversion processing, offset correction processing, inter-channel sensitivity correction processing, and beam hardening correction on the detection data output from the DAS 18. The processing circuitry 44 can store the generated data in the memory 41. Note that the data before preprocessing (detection data) and the data after preprocessing may be collectively referred to as projection data. The processing circuitry 44 can generate CT image data by performing reconstruction processing using a filtered back projection method, an iterative reconstruction method, machine learning, or the like on the generated projection data (projection data after preprocessing). The processing circuitry 44 can store the generated CT image data in the memory 41. In the image processing function 443, the processing circuitry 44 can convert the CT image data generated by the image generating function 442 into tomographic image data of an arbitrary cross section or three-dimensional image data by a known method based on an input operation received from an operator via the input interface 43. For example, the processing circuitry 44 can perform three-dimensional image processing such as volume rendering, surface rendering, image value projection processing, MPR (Multi-Planar Reconstruction) processing, and CPR (Curved MPR) processing on the CT image data to generate rendering image data of an arbitrary viewpoint direction. Note that the image generating function 442 may directly generate three-dimensional image data such as rendering image data of an arbitrary viewpoint direction, i.e., volume data. The processing circuitry 44 can store the tomographic image data or three-dimensional image data in the memory 41.

[0043] In the initial task model training function 444, the processing circuitry 44 can train the initial task model using an existing data set to obtain a trained initial task model. Here, the existing data set is, for example, a medical image including a region of interest based on CT image data generated by the image generation function 442. The existing data set may be, for example, stored in the memory 41 in advance, and in this case, the processing circuitry 44 acquires the existing data set from the memory 41. The existing data set may also include data source information such as a lesion site, task type information, and annotation information. The obtained trained initial task model may be stored in, for example, the memory 41. Here, the processing circuitry 44 that realizes the initial task model training function 444 may be an example of an initial task model training unit. As for the model training method, for example, a known method such as machine learning or deep learning is used, but a description thereof will be omitted here.

[0044] In the candidate data acquisition function 445, the processing circuitry 44 acquires medical images based on the CT image data generated by the image generation function 442 as candidate data. The candidate data may be stored in advance in, for example, the memory 41, and in this case, the processing circuitry 44 acquires the candidate data from the memory 41. Exceptionally, the candidate data may include data source information such as a lesion site, task type information, and annotation information, as in the case of an existing data set. In addition, the processing circuitry 44 may acquire candidate data based on at least one of, for example, a task and a lesion site. This makes it possible to acquire candidate data that is significantly related to the task, which is advantageous for improving the training accuracy of the model. Here, the processing circuitry 44 that realizes the candidate data acquisition function 445 may be an example of a candidate data acquisition unit.

[0045] In the selection function 446, the processing circuit 44 can infer the candidate data obtained by the candidate data acquisition function 445 using the trained initial task model obtained by the initial task model training function 444, and select the candidate data that meets the conditions as a selection data set. For example, the selection function 446 includes a candidate data inference function that infers the candidate data using the trained initial task model and acquires the characteristics of the candidate data, and a candidate data selection function that calculates a score of similarity between the candidate data and an existing data set based on the characteristics of the candidate data, and selects the candidate data that meets the conditions as data included in the selection data set based on the score. Here, the inference is, for example, inputting a medical image that is the candidate data into the trained initial task model and obtaining an output image that includes various characteristics. In addition, the similarity score is, for example, a numerical value (similarity) indicating the degree of similarity between the candidate data and the existing data set. Note that here, the candidate data inference function is an example of a candidate data inference unit, and the candidate data selection function is an example of a candidate data selection unit. In addition, the candidate data inference unit is also referred to as a feature acquisition unit.

[0046] The features here may vary depending on the task, and may be one or more features. In Table 1 below, available features for different tasks are listed. Note that the features listed in the table are merely examples and are not intended to be limiting.

[0047] [Table 1]

[0048] In addition, when calculating the score of similarity between the candidate data and the existing dataset, different indices can be selected depending on whether the data is labeled or not and the difference in the task. In Table 2, similarity indices that can be used for different tasks when the data is labeled and when it is unlabeled are listed, and the score of similarity between the candidate data and the existing dataset is calculated. Note that the ones listed in the table are merely examples and are not limited to these.

[0049] [Table 2]

[0050] Moreover, the processing circuitry 44 may store the obtained selected data set in the memory 41. Here, the processing circuitry 44 realizing the selection function 446 may be an example of a selection section. Details of how the processing circuitry 44 realizing the selection function 446 will be described later.

[0051] In the tuning function 447, the processing circuitry 44 can perform feature-relational analysis on the existing data set and the selected data set, and adjust the training policy and the pre-processing plan based on the analysis results, where the training policy includes at least one of feedback weights of the loss function and sampling probability of the samples, and the pre-processing plan includes at least one of image normalization parameters and data argumentation parameters for positive and negative samples.

[0052] Here, the features of the dataset that can be used for the feature relation analysis include numerical features and annotation features obtained by analyzing the dataset. The numerical features include at least one of the following: statistical characteristics of foreground pixels in the dataset, distribution features of positive and negative samples in the dataset, and distribution features of targets in the dataset. The annotation features include, for example, whether the annotation is fine and whether the annotation meets the needs of the current task. The annotation features also include at least one of the following: annotation edge fineness, annotation miss rate, annotation frame, and landmark point error rate.

[0053] In Table 3, we list the numerical features and annotation features that can be used for feature relationship analysis for different tasks.

[0054] [Table 3]

[0055] The details of how the processing circuitry 44 adjusts the training policy and the pre-processing plan based on the analysis results will be described later. Here, the processing circuitry 44 realizing the adjustment function 447 may be one example of an adjustment unit.

[0056] In the model training function 448, the processing circuitry 44 performs model training using a training data set consisting of an existing data set and a selected data set based on the adjusted training policy and preprocessing plan to obtain a trained model (a learned model). Here, the processing circuitry 44 may store the obtained trained model in the memory 41. Here, the processing circuitry 44 that realizes the model training function 448 may be an example of a model training unit.

[0057] Moreover, each of the functions 441 to 448 is not limited to being realized by a single processing circuit. The processing circuit 44 may be configured by combining a plurality of independent processors, and each of the processors may execute a respective program to realize each of the functions 441 to 448. Here, each of the functions 441 to 448 may be realized by being appropriately distributed or integrated into a single or a plurality of processing circuits.

[0058] Although the case where the console 40 executes multiple functions by a single console has been described, multiple functions may be executed by different consoles. For example, the functions of the processing circuit 44, such as the image generation function 442, the image processing function 443, the initial task model training function 444, the candidate data acquisition function 445, the selection function 446, the adjustment function 447, and the model training function 448, may be distributed.

[0059] In addition, a part or all of the processing circuitry 44 is not limited to being included in the console 40, but may be included in an integrated server that collectively processes detection data acquired by one or more medical image diagnostic devices.

[0060] In addition, the processing circuitry 44 is configured as a medical data processing device serving as an initial task model training unit, a candidate data acquisition unit, a selection unit, an adjustment unit and a model training unit by realizing an initial task model training function 444, a candidate data acquisition function 445, a selection function 446, an adjustment function 447 and a model training function 448.

[0061] Hereinafter, the operation, selection function, and adjustment function of the medical data processing device according to the first embodiment will be described in detail with reference to FIGS.

[0062] In the present embodiment, which will be described below with reference to the drawings, taking the segmentation (division) task of metastatic lung cancer as an example, the selected existing data is a group of lung CT images of patients with metastatic lung cancer and a metastatic lung cancer segmentation annotation dataset GT1, and the candidate data is a group of lung CT images of patients with pulmonary nodules and a pulmonary nodule segmentation annotation dataset GT2.

[0063] FIG. 2 is a flowchart of an example of the operation of the medical data processing apparatus according to the first embodiment.

[0064] 2, in step 100, an initial task model training unit of the medical data processing device trains an initial task model using an existing dataset to obtain a trained initial task model, and then proceeds to step 200. Here, the initial task model is trained using a group of lung CT images of metastatic lung cancer patients and a metastatic lung cancer segmentation annotation dataset GT1 to obtain a trained initial task model.

[0065] In step 200, the candidate data acquisition unit of the medical data processing device acquires candidate data, and then the process proceeds to step 300. Here, the candidate data acquired by the candidate data acquisition unit is a group of lung CT images of a pulmonary nodule patient and a pulmonary nodule segmentation annotation dataset GT2.

[0066] In step 300, the selection unit of the medical data processing device infers candidate data using the trained initial task model, and selects candidate data that meets the conditions as a selection data set, and then proceeds to step 400.

[0067] Here, the selection function of the medical data processing device will be described in detail with reference to Fig. 3 to Fig. 5. Fig. 3 is a flowchart of an example of the selection function of the medical data processing device according to the first embodiment. Fig. 4 is a diagram showing an example of the selection function of the medical data processing device according to the first embodiment. Fig. 5 is a schematic diagram of another specific example of the selection function of the medical data processing device according to the first embodiment.

[0068] 3, for example, when the processing circuitry 44 of the medical data processing device executes the selection function 446, for example, in step 310, the processing circuitry 44 first executes the candidate data inference function, and uses the trained initial task model to infer the candidate data and obtain the features of the candidate data, and then proceeds to step S320. Here, as shown in FIG. 4, the medical image sequence on the left side shows the candidate data, and the candidate data is input into the trained initial task model CM to obtain the feature CDF of the candidate data.

[0069] In step S320, the processing circuit 44 executes a candidate data selection function, specifically, calculates a similarity index value, which is a score of similarity between the candidate data and the existing data set, based on the features of the candidate data, and selects candidate data that meets the condition as a selection data set based on the score. Here, as shown in Fig. 4, the processing circuit 44 calculates a score of similarity between the candidate data and the existing data set based on the feature CDF of the candidate data, and selects candidate data that meets the condition TT as a selection data set SD based on the score.

[0070] The selection function 446 will be described in more detail below by taking a metastatic lung cancer segmentation task as a specific example. Here, the existing dataset is a group of lung CT images of a metastatic lung cancer patient and a metastatic lung cancer segmentation annotation dataset GT1, the trained initial task model obtained based on the existing data is CM, and the candidate data is lung nodule segmentation data including a group of lung CT images of a patient and a lung nodule segmentation annotation dataset GT2 shown on the left side of FIG. 5.

[0071] In this example, when the processing circuitry 44 executes the sorting function 446, the processing circuitry 44 inputs the pulmonary nodule segmentation data, which is the candidate data, into the trained initial task model CM to infer the candidate data and obtain the features of the candidate data. From the above, since this example is a segmentation task, according to Table 1, the extractable features for the segmentation task include a segmentation mask. Here, the segmentation mask is obtained as the feature of the candidate data.

[0072] Then, a similarity score between the candidate data and the existing dataset is calculated based on the segmentation mask, which is a feature of the candidate data. When calculating the similarity score, as described above, since this specific example is a segmentation task and the data has labels, as can be seen from Table 2 above, DICE can be selected as the similarity score between the candidate data and the existing dataset. Then, based on the score, candidate data that meets the condition is selected as a selection dataset. Here, the selection condition is that DICE is greater than 0.8. That is, the selection dataset is constructed by selecting candidate data with DICE greater than 0.8. The data selected from the pulmonary nodule segmentation data is shown on the far right side of FIG. 5. Note that, here, DICE is selected as the similarity index, and 0.8 is selected as the threshold used for selection, but this is not limited thereto, and other values ​​may be selected as the threshold used for selection according to the actual situation, and other features may be appropriately selected as the similarity index according to the task.

[0073] In this way, appropriate features are selected for the candidate data according to the type of task, a score of similarity between the candidate data and the existing dataset is calculated based on the features, and selection is performed based on the similarity score, and candidate data that meets the conditions is selected as a selection dataset, which is then combined with the existing dataset to form a new training dataset. In this way, instead of using the candidate data as is, candidate data that is correlated with the task is selected based on the task, and candidate data that is not related to the task is excluded from the model training set that completes the task. The exclusion of unrelated data leads to the accuracy of the model training not being reduced.

[0074] Next, returning to FIG. 2, in step S400, the adjustment unit of the medical data processing device performs feature relationship analysis on the existing data set and the selected data set, and adjusts the training policy and pre-processing plan based on the analysis result, and then proceeds to step S500.

[0075] Here, the details of the adjustment function of the medical data processing device will be described with reference to Figs. 6 to 11. Fig. 6 is a flowchart of an example of the adjustment function of the medical data processing device according to the first embodiment. Fig. 7 is a schematic diagram of a specific example of the adjustment function of the medical data processing device according to the first embodiment. Fig. 8 is a schematic diagram of another specific example of the adjustment function of the medical data processing device according to the first embodiment. Fig. 9 is a diagram showing a specific example of adjustment of a pre-processing plan by the adjustment function of the medical data processing device according to the first embodiment. Fig. 10 is a schematic diagram of a specific example of adjustment of a training policy by the adjustment function of the medical data processing device according to the first embodiment. Fig. 11 is a schematic diagram of a specific example of adjustment of a training policy by the adjustment function of the medical data processing device according to the first embodiment.

[0076] As shown in FIG. 6, for example, when the processing circuit 44 of the medical data processing device executes the adjustment function 447, for example, in step S410, first, a feature relationship analysis is performed on the existing data set and the selected data set, and then the process proceeds to step S420.

[0077] In step S420, processing circuitry 44 adjusts the training policy and pre-processing plan based on the analysis results.

[0078] Details of the feature relationship analysis and the adjustment of the training policy and pre-processing plan are described below with reference to a concrete example.

[0079] Regarding the feature relationship analysis, as described above, the features of the dataset on which the feature relationship analysis can be performed include at least one of the numerical features and annotation features obtained by analyzing the dataset, and the target used for the analysis is the existing dataset, or a set consisting of the existing dataset and the selected dataset.

[0080] Hereinafter, the details of the feature relationship analysis performed based on the data features for the existing dataset and the set consisting of the existing dataset and the selected dataset will be described with reference to FIG. 7. Here, the segmentation task of metastatic lung cancer will be continued as an example, and the results of each step so far will be used. Thus, as can be seen from Table 3 above, when performing the feature relationship analysis for the segmentation task, the numerical features that can be used may be the statistical characteristics of the foreground pixels.

[0081] The feature relationship analysis based on the data features is, for example, to analyze the statistical characteristics of the foreground pixels, for example, to calculate the statistics (analysis, calculation) of the foreground pixel range of the existing data set and the set consisting of the existing data set and the selected data set. Here, the lower limit of the foreground pixel range is the 0.5% quantile of all foreground pixels, and the upper limit of the foreground pixel range is the 99.5% quantile. For example, as a result of analyzing the statistical characteristics of the foreground pixels of the existing data set, the foreground pixel range of the existing data set is, for example, (-409, 376). On the other hand, as a result of analyzing the statistical characteristics of the foreground pixels of the set consisting of the existing data set and the selected data set, the foreground pixel range of the set consisting of the existing data set and the selected data set is, for example, (-768, 514). It can be seen that the foreground pixel range is expanded by introducing the selected data set. It is closer to the foreground pixel range (-1250, 250) that is generally used for image diagnosis.

[0082] In addition, the three diagrams from left to right in FIG. 7 respectively show the range of the lesion obtained after preprocessing the same image with the foreground pixel range of each of the existing data set, the set consisting of the existing data set and the selected data set, and the foreground pixel range generally used for image diagnosis. As a result, when the foreground pixel range of the set consisting of the existing data set and the selected data set is used, more useful information is obtained, a larger lesion area is obtained, and redundant information is reduced compared to when the foreground pixel range of the existing data set is used. As a result, it is possible to avoid increasing the load of calculation or judgment (determination) by processing using too much redundant information, which is advantageous for improving the real-time performance of the medical data processing device. That is, according to this embodiment, since processing is performed using a smaller amount of information, it is possible to suppress an increase in the load of calculation or judgment, and to improve the real-time performance of the medical data processing device.

[0083] Hereinafter, with reference to FIG. 8, the details of the feature relationship analysis performed based on the annotation features for the existing dataset and the set consisting of the existing dataset and the selected dataset will be described. Here, the metastatic lung cancer segmentation task will be continued as an example, and the results of each step so far will be used. Thus, as can be seen from Table 3 above, when performing the feature relationship analysis for the segmentation task, the annotation features available may include whether the annotation data edge is fine or not, and whether the annotation data target meets the current task needs.

[0084] Performing feature relationship analysis based on annotation data features specifically includes assigning annotation data quality scores for the current task to the selected data in the selected dataset based on the task. For example, for the segmentation task, as shown in Fig. 8, 4 points are assigned to the selected data whose annotation data is detailed and meets the annotation data criteria (OK) as shown in the upper right diagram of Fig. 8, 3 points are assigned to the selected data whose annotation data is detailed and does not meet the annotation data criteria (NG) as shown in the upper left diagram of Fig. 8, 2 points are assigned to the selected data whose annotation data is not detailed (CC) and meets the annotation data criteria (OK) as shown in the lower right diagram of Fig. 8, and 1 point is assigned to the selected data whose annotation data is not detailed (CC) and does not meet the annotation data criteria (NG) as shown in the lower left diagram of Fig. 8. Here, for a segmentation task, the annotation data is detailed if, for example, the mask edges are smooth and clear, there is no obvious over-segmentation or under-segmentation, and otherwise the annotation data is considered not detailed. Also, meeting the annotation data criteria means, for example, that the annotation does not include a ground-glass region, and otherwise the annotation data is considered not to meet the annotation data criteria.

[0085] The details of adjusting the training policy and pre-processing plan based on the result of the feature analysis in step S420 will be described below.

[0086] Adjusting the pre-processing plan based on the result of the feature analysis as described above is, for example, adjusting pre-processing regularization parameters using the foreground pixel range extracted in step S410.

[0087] As an example, it can be adjusted according to the following formula (1).

[0088]

number

[0089] where α∈(0,1), β∈(0,1), and I min is the lower bound of the foreground pixel range after adjusting the preprocessing plan, and I max is the upper limit of the foreground pixel range after adjusting the preprocessing plan, and I min_ori is the lower bound of the foreground pixel range of the existing dataset, and I min_new is the lower bound of the foreground pixel range of the set consisting of the existing dataset and the selected dataset, and I max_ori is the upper limit of the foreground pixel range of the existing dataset, and I max_new is the upper bound of the foreground pixel range of the set consisting of the existing dataset and the selected dataset.

[0090] α and β are adjustment parameters. The values ​​of α and β can be adjusted according to the similarity index value, thereby adjusting the influence of the selected data on the pixel range. As can be seen from formula (1), the larger the values ​​of α and β, the greater the weight of the selected data, and the more significant the influence of the selected data on the pixel range.

[0091] Fig. 9 shows an example of adjusting the pre-processing plan by adjusting the value of α. In the example of Fig. 9, the value of α is adjusted based on the value of DICE, which is the similarity index in step S320. In FIG. 9, the horizontal axis indicates the value of DICE, and the vertical axis indicates the value of α. As shown in FIG. 9, the value of DICE and the value of α are piecewise linear function related. Specifically, when the value of DICE is within the range of 0 to 0.4, the value of α is 0. At this time, a value of DICE less than 0.4 indicates that the similarity between the selected data and the existing data is low, and the introduction of such selected data indicates that the model accuracy is likely to decrease, and such selected data is excluded from the training data. Also, when the value of DICE is within the range of 0.4 to 0.8, the value of α increases linearly with an increase in the value of DICE. At this time, the proportion of selected data increases with an increase in the value of DICE. Also, when the value of DICE is within the range of 0.8 to 1, the value of α is 1. In this case, the similarity between the selected data and the existing data is high, and the selected data can be completely adopted.

[0092] As described above, in step S320, the selection condition is that DICE is greater than 0.8. Therefore, all selected data in the selected data set are data with DICE greater than 0.8. Thus, according to the example of FIG. 9, the value of α is 1 for the data set obtained in step S320.

[0093] Although an example of adjusting the value of α has been shown here, the adjustment of the value of β is similar to the adjustment of the value of α, and therefore the value of β is also 1.

[0094] Substituting the values ​​of α and β into equation (1), we obtain the following equation (2).

[0095]

number

[0096] That is, the selected data set selected in step S320 is fully adopted to extend the existing data set.

[0097] In the example of Fig. 9, the value of α is adjusted assuming that the value of the similarity index and the value of α have a piecewise linear function relationship, but the present invention is not limited to this and the value of α may be adjusted using an exponential function, a logarithmic function, or the like. In other words, the values ​​of α and β may be adjusted when the value of the similarity index and the values ​​of α and β have an exponential function or a logarithmic function relationship. In this way, the adjustment of the pre-processing plan may be realized by designing a fitting conversion function based on at least one of a linear function, an exponential function, and a logarithmic function.

[0098] Adjusting the training policy based on the above feature analysis result may include, for example, adjusting the training policy using the annotation feature scores obtained in step S410. For example, for the selected data with high annotation quality, the feedback weight in the model training or network training process may be increased.

[0099] As an example, the adjustment may be made according to the following formula (3).

[0100]

number

[0101] Here, Lw is the feedback weight, whose upper limit is 1 and whose lower limit is 0 or a set predetermined value, and Q is the annotation feature score.

[0102] As can be seen from equation (3), the higher the annotation feature score, the larger the feedback weight, and the greater the effect of the filtered data on model training. Fig. 10 shows an example of adjusting the training policy based on Equation 3. In the example of Fig. 10, the feedback weights Lw of the filtered data are adjusted based on the annotation feature scores in step S410.

[0103] 10, the horizontal axis indicates the value of the annotation feature score (Q-label quality), and the vertical axis indicates the value of the feedback weight Lw. The lower limit of the feedback weight Lw is set to 0.01.

[0104] 10, when the annotation feature scores are 1, 2, 3, and 4, the values ​​of the feedback weight Lw are 0.01, 0.046415888, 0.215443469, and 1, respectively. That is, as the annotation feature score increases, the feedback weight Lw increases exponentially.

[0105] Here, the feedback weights of the selected data are adjusted according to an exponential function based on the annotation feature score, but the present invention is not limited to this, and the feedback weights of the selected data may be adjusted using a linear function, a logarithmic function, etc. In other words, if the feedback weights increase according to a linear function or a logarithmic function as the annotation feature score increases, the feedback weights of the selected data may be adjusted according to a linear function or a logarithmic function based on the annotation feature score. In this way, the adjustment of the training policy may be realized by designing a fitting conversion function based on at least one of a linear function, an exponential function, and a logarithmic function.

[0106] A specific example will be given below to compare the embodiment with the prior art.

[0107] In this specific example, the annotation feature score of certain selected data is set to 1, and model training is performed based on the same existing data and the above-mentioned certain selected data by the present embodiment and the conventional technology, respectively, and certain selected data is inferred based on the obtained model. In Figure 11, certain selected data, the inference result of the conventional technology, the inference result of this embodiment, and the expected inference result that meets the technical standard are shown in order from left to right. In the conventional technology, the selected data is used as it is for model training without adjusting the training policy.

[0108] According to the present embodiment, the model training includes a training policy adjustment. Specifically, for the selected data, the annotation feature score is 1, so the feedback weight is adjusted to 0.01 in the model training according to the curve in FIG.

[0109] As can be seen from FIG. 11, the inference result of the present embodiment is closer to the expected inference result, while the inference result of the prior art contains more noise.

[0110] In this manner, in this embodiment, a feature relationship analysis is performed on the existing data set and the selected data set, and the training policy is adjusted based on the analysis result, so that data with high annotation feature scores are fully adopted in model training, and data with low annotation feature scores are made to have a small effect in model training. In this manner, the expanded training data improves the accuracy of model training by minimizing the decrease in model accuracy caused by the introduction of different task data.

[0111] Next, returning to FIG. 2, in step 500, model training is performed using a training dataset consisting of the above-mentioned existing dataset and the above-mentioned selected dataset based on the adjusted training policy and pre-processing plan, a trained model is obtained, and then the process is terminated.

[0112] As described above, the medical data processing device according to this embodiment includes an initial task model training unit that trains an initial task model using an existing dataset and obtains a trained initial task model; a candidate data acquisition unit that acquires candidate data; a selection unit that acquires features of the candidate data using the trained initial task model, selects candidate data that meets conditions based on the acquired features, and acquires a selection dataset including the selected candidate data; an adjustment unit that performs feature relationship analysis on the existing dataset and the selection dataset and adjusts a training policy and a preprocessing plan based on the analysis results; and a model training unit that performs model training using a training dataset consisting of the existing dataset and the selection dataset based on the adjusted training policy and the adjusted preprocessing plan to obtain a trained model.

[0113] In this way, instead of using the candidate data as is, the candidate data having a correlation with the task is selected, and the candidate data not related to the task is excluded from the model training set for completing the task. By not reducing the accuracy of the model training by introducing data, it is possible to avoid increasing the load of calculation or judgment by processing using too much redundant information, which is advantageous for improving the real-time performance of the medical data processing device.

[0114] On the other hand, by performing feature relationship analysis on the existing dataset and the selected dataset and adjusting the training policy based on the analysis results, data with high annotation feature scores are fully adopted in model training, and data with low annotation feature scores are made to have a small effect in model training. In this way, the expanded training data improves the accuracy of model training by minimizing the decrease in model accuracy caused by the introduction of different task data.

[0115] As described above, according to the first embodiment, a trained model with good accuracy can be obtained.

[0116] Second Embodiment Hereinafter, the medical data processing device according to the second embodiment will be described with reference to FIG. 12. FIG. 12 is a flowchart of an example of the operation of the medical data processing device according to the second embodiment. The configuration of the medical data processing device according to the second embodiment is different from the configuration of the medical data processing device according to the first embodiment only in that the processing circuit 44 executes a determination function for determining whether the accuracy of the trained model has improved or not, and a control function for repeatedly executing the selection function 446, the adjustment function 447, and the model training function 448 when the accuracy is determined to have improved by the determination function, until the determination function determines that the accuracy has not improved. The determination function is, for example, an example of a determination unit.

[0117] In FIG. 12, the operations of steps S100 to S500 are similar to those described in FIG. 2 relating to the first embodiment, so the description will be omitted here, and only the different step S600 will be described.

[0118] In step S600, it is determined whether the accuracy of the trained model obtained in step S500 has improved. If the accuracy of the trained model has improved (YES in step S600), the process returns to step S300. If the accuracy of the trained model has not improved (NO in step S600), the process ends.

[0119] According to the second embodiment, in addition to the effects of the first embodiment, the following effects are achieved. For example, in the second embodiment, the results of the selection of candidate data and the adjustment of the training policy and the pre-processing plan are judged, and if these processes can improve the model training accuracy, the selection of candidate data and the adjustment of the training policy and the pre-processing plan are repeated until the model training accuracy is no longer improved. As a result, according to the second embodiment, the model training accuracy can be maximized.

[0120] (Experimental Results) Experiments 1 and 2 were conducted for the above-described embodiment to verify the technical effects of the embodiment.

[0121] (Experiment 1) Experiment 1 was a metastatic lung cancer segmentation task. In experiment 1, the existing dataset was lung CT images of metastatic lung cancer patients and segmentation annotations of metastatic lung cancer, and the candidate data was lung CT images of lung nodule patients and segmentation annotations of lung nodules.

[0122] First, we analyze and compare the existing dataset and the selected dataset selected from the candidate data. Using the method of selecting the selected dataset from the candidate data described above, we performed three consecutive data selections, and performed lesion long diameter distribution statistics on the same image data group using the existing dataset, the first selected dataset, the second selected dataset, and the third selected dataset. The statistical results are shown in Table 4.

[0123] [Table 4]

[0124] As can be seen from Table 4, the first, second, and third screening datasets selected from the candidate data, lung CT images of lung nodule patients and lung nodule segmentation annotations, show high similarity to the existing dataset, metastatic lung cancer training set, in terms of lesion long axis distribution. Furthermore, the similarity gradually improves with each screening.

[0125] In addition, the existing data set and the selected data set selected from the candidate data are analyzed and compared. Based on the above method of selecting the selected data set from the candidate data, data selection is performed three times consecutively, and the existing data set, the first selected data set, the second selected data set, and the third selected data set are used to perform the texture distribution statistics of the lesions for the same image data group. The statistical results are shown in Table 5.

[0126] [Table 5]

[0127] As can be seen from Table 5, the lesion textures include solid, solid / Mixed, Part solid / Mixed, Non-solid / Mixed, and Non-solid / GGO.

[0128] Regarding the distribution of lesion texture, the first, second and third screening datasets selected from the candidate data, lung CT images of lung nodule patients and lung nodule segmentation annotations, have a large similarity to the existing dataset, metastatic lung cancer training set, but there are some differences. In view of this, the training policy is adjusted based on the above embodiment.

[0129] Based on the existing dataset and the selected dataset from Experiment 1, model training was performed based on the above analysis and adjustment, and the model accuracy was evaluated. Here, the average tumor-wise DICE was used as the evaluation index.

[0130] The results of comparing the segmentation results of metastatic lung cancer segmentation by the initial task model, the conventional method, and the method of this embodiment are shown in Table 6. In addition, in Fig. 13, these three types of segmentation results are shown from top to bottom, respectively.

[0131] [Table 6]

[0132] As can be seen from Table 6, when using the initial task model, the Average Tumor-wise DICE obtained is the lowest when no candidate data is added during model training, no data is selected, and no training policy and preprocessing plan are customized based on the selected data.

[0133] In addition, when using the known method, candidate data is added and data is selected during model training, but the training policy and preprocessing plan are not customized based on the selected data, and the obtained Average Tumor-wise DICE is lower than the value when the initial task model is used. This is thought to be because the introduction of candidate data introduces noise, which affects the training accuracy of the model.

[0134] In addition, when using the method of this embodiment, candidate data is added during model training, the candidate data is selected according to different tasks, and the training policy and pre-processing plan are customized according to the characteristics of the selected data, and the obtained Average Tumor-wise DICE is higher than that when using the initial task model and the method of the prior art. As can be seen, the image data processing device of this embodiment makes full use of the useful information of the candidate data, takes into account multiple evaluation requirements, and obviously improves the effect of completing the task.

[0135] (Experiment 2) Experiment 2 was the segmentation task of primary lung cancer. In experiment 2, the existing dataset was lung CT images of patients with primary lung cancer and primary lung cancer segmentation annotations, and the candidate data was lung CT images of patients with metastatic lung cancer and metastatic lung cancer segmentation annotations.

[0136] In experiment 2, similar to experiment 1, the existing dataset and the selected dataset selected from the candidate data are analyzed and compared. Based on the method of selecting the selected dataset from the candidate data described above, data selection is performed continuously, and lesion long axis distribution statistics are performed on the same image data group using the existing dataset and the selected dataset. The statistical results are shown in Table 7.

[0137] [Table 7]

[0138] As can be seen from Table 7, the metastatic lung cancer data itself objectively contains many small tumors, so when the selected metastatic lung cancer dataset is introduced, many tumors between 10 mm and 20 mm are selected. However, in terms of the distribution of the long axis of the lesions, there is a large similarity between the candidate data, the metastatic lung cancer dataset, and the existing dataset, the primary lung cancer dataset.

[0139] In addition, in experiment 2, the operation flow of the first embodiment is executed, specifically, an initial task training model based on an existing dataset called a primary lung cancer dataset is used to infer metastatic lung cancer data, which is candidate data, and then the DICE score, which is a similarity index, is calculated and analyzed, and data with a DICE score in the range of 0.7 to 0.9 is selected as a selected dataset, which is integrated with the existing dataset, and then the CT value distribution of the integrated dataset is analyzed to obtain a new pre-processed pixel range (-362, 296), and then the annotation feature score is performed based on metastatic lung cancer, and the feedback weight as the training weight of the selected data is adjusted. Here, since the annotation quality and needs of metastatic lung cancer are almost the same as those of primary lung cancer, the annotation feature score is set to 4, and the training weight is set to 1.

[0140] Based on the existing dataset and selected dataset from Experiment 2, model training was performed based on the above analysis and adjustment, and the model accuracy was evaluated. Here, the average tumor-wise DICE was used as the evaluation index.

[0141] The initial task model and the segmentation results of primary lung cancer segmentation using the method of this embodiment are compared, and the comparison results are shown in Table 8.

[0142] [Table 8]

[0143] As can be seen from Table 8, when using the initial task model, the average tumor-wise DICE obtained is at least about 0.77 when no candidate data is added during model training, no data is selected, and no training policy and preprocessing plan are customized based on the selected data.

[0144] In addition, when using the method of this embodiment, during model training, candidate data is added, the candidate data is selected according to different tasks, and the training policy and pre-processing plan are customized according to the characteristics of the selected data, and the obtained Average Tumor-wise DICE is 0.81, which is obviously higher than the value when using the initial task model. As can be seen, the medical data processing device of this embodiment makes full use of the useful information of the candidate data, takes into account multiple evaluation requirements, and obviously improves the effect of completing the task.

[0145] (Modification) In the above-mentioned embodiments, specific examples, and experimental results, and in the drawings, examples of medical data processing devices have been shown and explained, but the present invention is not limited to these, and various modifications are possible and may be combined as necessary.

[0146] For example, in the above description and drawings, the initial task model training function 444 and the model training function 448 are provided as independent functions, but this is not limited thereto. The initial task model training function 444 and the model training function 448 may be configured as a single module and may perform processing in response to different inputs.

[0147] Furthermore, although the above description and drawings show details of the medical data processing device and method, this realization form is not limited to the medical data processing device and method, and may be realized as a medical image diagnostic device including a medical data processing device, a program for causing a computer to execute the steps of the medical data processing method, an integrated circuit into which the steps for causing a computer to execute the medical data processing method are written, and a storage medium storing a program that non-temporarily stores the steps for causing a computer to execute the medical data processing method.

[0148] According to at least one of the embodiments described above, a trained model with good accuracy can be obtained.

[0149] Although some embodiments have been described, these embodiments are presented as examples and are not intended to limit the scope of the invention. These embodiments can be implemented in various other forms, and various omissions, substitutions, and modifications can be made without departing from the spirit of the invention. These embodiments and their modifications are included in the scope of the invention and its equivalents as described in the claims, as well as in the scope and spirit of the invention. [Explanation of symbols]

[0150] 1 X-ray CT device 444 Initial task model training function 445 Candidate data acquisition function 446 Sorting Function 447 Adjustment function 448 Model Training Function

Claims

1. an initial task model training unit that trains an initial task model using an existing dataset to obtain a trained initial task model; a candidate data acquisition unit that acquires candidate data; a selection unit that acquires features of the candidate data using the trained initial task model, selects the candidate data that meets a condition based on the features, and acquires a selection data set including the selected candidate data; an adjustment unit that performs a feature relationship analysis on the existing dataset and the selected dataset, and adjusts a training policy and a pre-processing plan based on the analysis result; A model training unit that performs model training using a training dataset consisting of the existing dataset and the selected dataset based on an adjusted training policy and an adjusted preprocessing plan to obtain a trained model; A medical data processing device comprising:

2. The sorting unit includes: a feature acquisition unit that acquires the features of the candidate data using the trained initial task model; a candidate data selection unit that calculates a score of similarity between the candidate data and the existing data set based on the characteristics of the candidate data, and selects the candidate data that meets a condition based on the score as data to be included in the selection data set. The medical data processing device according to claim 1 .

3. A determination unit that determines whether the accuracy of the trained model has improved; a control unit that repeatedly operates the selection unit, the adjustment unit, and the model training unit until the determination unit determines that the accuracy has not improved, when the determination unit determines that the accuracy has improved; The medical data processing device according to claim 1 , further comprising:

4. the candidate data acquisition unit acquires the candidate data based on at least one of a task and a lesion site; The medical data processing device according to claim 1 .

5. The medical data processing apparatus according to claim 1 , wherein the features include at least one of numerical features and annotation features obtained by analyzing the dataset.

6. The numerical features include at least one of a statistical characteristic of foreground pixels, a positive and negative sample distribution feature, and a target distribution feature; The annotation features include at least one of a definition of an annotation edge, a miss rate of an annotation, a frame of an annotation, and an error rate of a landmark point. The medical data processing device according to claim 5 .

7. The characteristics vary depending on the task, The feature is one or more features. The medical data processing device according to claim 1 .

8. The training policy includes at least one of feedback weights of a loss function and sampling probabilities of samples; the pre-processing plan includes at least one of image normalization parameters and data argumentation parameters for positive and negative samples; The medical data processing device according to claim 1 .

9. The adjustment of the training policy and the pre-processing plan is achieved by designing a fitting transformation function based on at least one of a linear function, an exponential function, and a logarithmic function; The medical data processing device according to claim 1 .

10. an initial task model training step of training an initial task model using an existing dataset to obtain a trained initial task model; a candidate data acquisition step of acquiring candidate data; a screening step of acquiring features of the candidate data using the trained initial task model, screening the candidate data that meets a condition based on the features, and acquiring a screening data set including the screened candidate data; performing a feature relationship analysis on the existing dataset and the selected dataset, and adjusting a training policy and a pre-processing plan based on the analysis result; A model training step of performing model training using a training dataset consisting of the existing dataset and the selected dataset based on the adjusted training policy and the adjusted pre-processing plan to obtain a trained model; A medical data processing method comprising:

11. To your computer training an initial task model using an existing dataset to obtain a trained initial task model; A process of obtaining candidate data; A process of acquiring features of the candidate data using the trained initial task model, selecting the candidate data that meets a condition based on the features, and acquiring a selection data set including the selected candidate data; performing a feature-relationship analysis on the existing dataset and the selected dataset, and adjusting a training policy and a pre-processing plan based on the analysis result; A process of performing model training using a training dataset consisting of the existing dataset and the selected dataset based on the adjusted training policy and the adjusted pre-processing plan to obtain a trained model; A program for executing the above.

12. A medical image diagnostic apparatus comprising the medical data processing device according to any one of claims 1 to 9.

Images