Training method and training system for learning network for medical image analysis

Abstract

The present disclosure relates to a training method and system for a learning network for medical image analysis. The training method includes: setting a learning network with a predetermined structure and obtaining an original training data set; having a processor pre-train the learning network using the original training data set to obtain a pre-trained learning network; having the processor evaluate the pre-trained learning network to determine whether it has preset evaluation defects; having the processor perform data enhancement on the original training data set based on the evaluation defects if the pre-trained learning network has evaluation defects; and having the processor perform fine-tuning training on the learning network using the data-enhanced training data set based on the parameters of the pre-trained model. The present disclosure can evaluate and train the learning network in stages, thereby reducing the complexity of medical image processing and improving the efficiency and accuracy of medical image analysis.

Description

[0001] This application is a divisional application of the Chinese invention patent application with application number 202110525922.4, application date May 13, 2021, and invention name “Training method and training system for learning network for medical image analysis”. Technical Field

[0002] The present disclosure relates to the technical field of medical image analysis, and more particularly, to a training method and training system for a learning network for medical image analysis, a training device, and a computer-readable storage medium. Background Art

[0003] Medical images can quickly and non-invasively or minimally invasively capture structural and functional images of human internal tissues, providing a clear understanding of the physiological structure and metabolic status of various human organs. Currently, medical image analysis has become an indispensable aid in clinical diagnosis. Through image processing and computer technology, this information can be effectively processed for use in assisting diagnosis and surgical planning, offering significant social benefits and broad application prospects.

[0004] In recent years, deep learning (DL) has rapidly developed into a research hotspot in medical image analysis. It can automatically extract implicit disease diagnostic features from medical image big data, and therefore has gradually become an important technology in image screening and classification.

[0005] However, medical images require a long period of accumulation to reach a certain amount of data, and the deep learning algorithm's demand for high-quality data makes it more difficult to obtain data that can be used for training. Therefore, when using medical data for model training, data for some specific areas or containing specific lesions is obviously insufficient. In addition, the noise and artifacts of medical images not only affect the quality of medical images, but also affect the deep learning ability of the model. Therefore, when the existing technology analyzes medical images based on deep learning, due to the particularity of medical data and the limitations of deep learning, it is difficult to achieve clinical application-level training effects and obtain model algorithms based on deep learning, resulting in high complexity in the training and learning of medical images, low efficiency and accuracy in medical image analysis, and failure to meet the requirements of clinical applications for accurate analysis and diagnosis of diseases. Summary of the Invention

[0006] The present disclosure is provided to solve the above-mentioned problems existing in the prior art.

[0007] The present disclosure requires a training method and training system, training device, and computer-readable storage medium for a learning network for medical image analysis. The training method pre-trains the learning network using an original training data set to obtain a pre-trained learning network, evaluates the pre-trained learning network to determine whether it has preset evaluation defects, and if the pre-trained learning network has evaluation defects, performs data enhancement on the original training data set based on the existing evaluation defects, and fine-trains the learning network using the data-enhanced training data set based on the parameters of the pre-trained model. The learning network can be iteratively trained based on limited medical data, thereby quickly and accurately acquiring medical images relevant to analysis requirements, and performing more precise and targeted analysis on the acquired medical images. Therefore, the complexity of training and learning medical images is reduced, and the efficiency and accuracy of medical image analysis are improved.

[0008] According to a first aspect of the present disclosure, a method for training a learning network for medical image analysis is provided. The method comprises: providing a learning network having a predetermined structure and obtaining an original training dataset; pre-training the learning network using the original training dataset by a processor to obtain a pre-trained learning network; evaluating the pre-trained learning network to determine whether it has a preset evaluation defect; if the pre-trained learning network has an evaluation defect, performing data enhancement on the original training dataset to address the evaluation defect; and fine-tuning the learning network using the data-enhanced training dataset based on parameters of the pre-trained model by the processor.

[0009] According to a second aspect of the present disclosure, a training system for a learning network for medical image analysis is provided. The training system includes: a communication interface configured to obtain an original training dataset; and a processor configured to: pre-train a learning network having a predetermined structure using the original training dataset to obtain a pre-trained learning network; evaluate the pre-trained learning network to determine whether it has preset evaluation defects; if the pre-trained learning network has evaluation defects, perform data enhancement on the original training dataset to address the existing evaluation defects; and perform fine-tuning on the learning network using the data-enhanced training dataset based on parameters of the pre-trained model.

[0010] According to a third aspect of the present disclosure, a training device for a learning network for medical image analysis is provided, comprising a memory and a processor, wherein the memory is used to store one or more computer program instructions, wherein the one or more computer program instructions are executed by the processor to implement a training method for a learning network for medical image analysis.

[0011] According to a fourth aspect of the present disclosure, a computer-readable storage medium is provided, on which computer program instructions are stored, wherein the computer program instructions, when executed by a processor, implement a training method for a learning network for medical image analysis.

[0012] By utilizing the training method and training system, training device and computer-readable storage medium of a learning network for medical image analysis according to various embodiments of the present disclosure, the learning network is pre-trained using an original training data set to obtain a pre-trained learning network, and the pre-trained learning network is evaluated to determine whether it has preset evaluation defects. In the event that the pre-trained learning network has evaluation defects, data enhancement is performed on the original training data set based on the existing evaluation defects, and based on the parameters of the pre-trained model, the learning network is fine-trained using the data-enhanced training data set. The learning network can be iteratively trained based on limited medical data, thereby quickly and accurately acquiring medical images related to the analysis requirements, and performing more precise and targeted analysis on the acquired medical images. Therefore, the complexity of training and learning of medical images is reduced, and the efficiency and accuracy of medical image analysis are improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] In the drawings, which are not necessarily drawn to scale, the same reference numerals may describe similar components in different views. The same reference numerals with letter suffixes or different letter suffixes may represent different instances of similar components. The accompanying drawings generally illustrate various embodiments by way of example and not limitation, and together with the description and claims, serve to illustrate the disclosed embodiments. Where appropriate, the same reference numerals are used throughout the drawings to refer to the same or similar parts. Such embodiments are illustrative and are not intended to be exhaustive or exclusive of the embodiments of the present apparatus or method.

[0014] Figure 1 A flowchart illustrating a method for training a learning network for medical image analysis according to an embodiment of the present disclosure;

[0015] Figure 2(a) to Figure 2(c) A schematic diagram illustrating intravenous doping according to an embodiment of the present disclosure;

[0016] Figure 3(a) to Figure 3(c) A schematic diagram showing a vascular branch fracture according to an embodiment of the present disclosure;

[0017] 4( a ) and 4 ( b ) are schematic diagrams showing solid nodules and ground glass nodules according to an embodiment of the present disclosure;

[0018] Figure 5 A schematic diagram illustrating an iterative training process according to an embodiment of the present disclosure;

[0019] Figure 6 A flowchart illustrating a method for training a learning network for medical image analysis according to an embodiment of the present disclosure;

[0020] Figure 7 A block diagram illustrating a training system for a learning network for medical image analysis according to an embodiment of the present disclosure; and

[0021] Figure 8 A structural block diagram of a training device for a learning network for medical image analysis according to an embodiment of the present disclosure is shown. DETAILED DESCRIPTION

[0022] In order to enable those skilled in the art to better understand the technical solution of the present disclosure, the present disclosure is described in detail below in conjunction with the accompanying drawings and specific embodiments. The embodiments of the present disclosure are further described in detail below in conjunction with the accompanying drawings and specific embodiments, but they are not intended to limit the present disclosure. For the various steps described herein, if there is no necessity for a contextual relationship between each other, the order in which they are described as examples herein should not be regarded as a limitation. Those skilled in the art should know that they can be adjusted in order as long as the logic between them is not destroyed, resulting in the inability to implement the entire process.

[0023] Furthermore, persons of ordinary skill in the art will appreciate that the figures provided herein are for illustration purposes only and are not necessarily drawn to scale.

[0024] Unless the context clearly requires otherwise, throughout the specification and claims, the words "include," "comprising," and similar words should be construed in an inclusive sense rather than an exclusive or exhaustive sense; that is, in the sense of "including but not limited to."

[0025] In the description of the present disclosure, it should be understood that the terms "first", "second", etc. are used for descriptive purposes only and are not to be understood as indicating or implying relative importance. In addition, in the description of the present disclosure, unless otherwise specified, "plurality" means two or more.

[0026] A training method and a training system for a learning network for medical image analysis according to an embodiment of the present disclosure will be described in detail below with reference to the accompanying drawings.

[0027] Figure 1 FIG. 1 is a flow chart showing a method for training a learning network for medical image analysis according to an embodiment of the present disclosure. Figure 1As shown, the training method may include setting a learning network with a predetermined structure and obtaining an original training data set (step 102). Note that the technical term "original training data set" herein is relative and is intended to refer to a training data set that has not been subjected to various data enhancement processes according to the embodiments of the present disclosure; specifically, the "original training data set" may refer to a training data set that has been subjected to data enhancement processes other than those disclosed herein, or may refer to a training data set that has been subjected to various other pre-processing processes not disclosed herein.

[0028] Specifically, the training of a learning network for medical image analysis can pre-set the structure of the learning network (i.e., a deep learning network) as a fixed network model and set a loss function. Here, the training device can be a smartphone, tablet computer, laptop computer, desktop computer, smart speaker, smart watch, etc., or it can be integrated into an image acquisition device for acquiring medical images or an image workstation for analyzing various medical images, and the embodiments of the present disclosure are not limited to this. Deep learning is a type of machine learning (ML), and machine learning is the only way to achieve artificial intelligence (AI). Deep learning combines low-level features to form more abstract high-level representation attribute categories or features to discover distributed feature representations of data. The network model can be trained using supervised learning. The architecture of the network model can include a stack of different blocks and layers, each of which converts one or more inputs into one or more outputs. Examples of different layers can include one or more convolutional layers or fully convolutional layers, nonlinear operator layers, pooling layers or subsampling layers, fully connected layers and / or final loss layers. Each layer can be connected to an upstream layer and a downstream layer. The network model may include a Residual Network (ResNet) model, a Segmentation Network (UNet) model, an AlexNet model, a GoogLeNet model, a Visual Geometry Group (VGG) model, a Pyramid Scene Parsing Network (PSPNet) model, a DeepLabV3 network model, etc., and the embodiments of the present disclosure are not limited thereto. A loss function is a function that maps the value of a random event or its related random variables to a non-negative real number to represent the "risk" or "loss" of the random event.

[0029] Furthermore, the training device can acquire medical images from a medical imaging device in real time through a communication interface, and manually segment the acquired medical images (for example, manually annotating the medical images) to obtain the original training data set; or, the training device can acquire medical images from a medical imaging device in real time through a communication interface, segment the acquired medical images using an existing model, and manually correct the segmented medical images to obtain the original training data set; that is, the image segmentation in the original training data set usually requires manual intervention to ensure its accuracy. Here, the communication interface may include a network adapter, a cable connector, a serial connector, a USB connector, a parallel connector, a high-speed data transmission adapter (such as an optical fiber, USB 3.0, a Thunderbolt interface, etc.), a wireless network adapter (such as a WiFi adapter), a telecommunications adapter (such as 3G, 4G / LTE, etc.), etc. The medical imaging device may include an X-ray imaging device, a magnetic resonance imaging (MRI) imaging device, an ultrasound imaging device, a nuclear medicine imaging device, a thermal imaging device, a medical optical imaging device, etc., and the embodiments of the present disclosure are not limited to this.

[0030] Medical images are images acquired by medical imaging devices and may include X-ray images (e.g., computed tomography (CT) images), MRI images, ultrasound images, radionuclide images, etc., which are not limited in the present embodiment. Medical images may be two-dimensional (2D) or three-dimensional (3D) images, and a 3D image may include multiple 2D image slices. Furthermore, medical images may be images of the lungs, liver, kidneys, or other parts of the body, which are not limited in the present embodiment.

[0031] Image segmentation refers to dividing an image into several non-overlapping regions based on features such as grayscale, color, spatial texture, and geometric shape, so that these features show consistency or similarity within the same region, but show obvious differences between different regions. Image segmentation methods based on deep learning may include but are not limited to image segmentation methods based on feature encoding, image segmentation methods based on region selection, image segmentation methods based on recurrent neural networks (RNNs), image segmentation methods based on upsampling / deconvolution, image segmentation methods based on improving feature resolution, image segmentation methods based on feature enhancement, image segmentation methods using conditional random fields (CRFs) / Markov random fields (MRFs), etc.

[0032] It should be noted that the training device can also obtain medical images collected by the medical imaging device from the server, and the present embodiment does not limit this. Here, the server can be an independent physical server, such as a Picture Archiving and Communication Systems (PACS), or a server cluster or distributed system composed of multiple physical servers, or it can also be a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, CDN, and big data and artificial intelligence platforms. The present embodiment does not limit this. The training device and the server can be directly or indirectly connected through wired or wireless communication, and the present embodiment does not limit this.

[0033] In some embodiments, the training method may further include pre-training the learning network using the original training data set by the processor to obtain a pre-trained learning network (step 104).

[0034] Specifically, after obtaining the original training data set, the processor of the training device can use the original training data set to pre-train the learning network to obtain a pre-trained learning network. Here, the processor can be a processing device including one or more general-purpose processing devices, such as a microprocessor, a central processing unit (CPU), a graphics processing unit (GPU), etc. More specifically, the processor can be a complex instruction set computing (CISC) microprocessor, a reduced instruction set computing (RISC) microprocessor, a very long instruction word (VLIW) microprocessor, a processor that runs other instruction sets, or a processor that runs a combination of instruction sets. The processor can also be one or more special-purpose processing devices, such as an application-specific integrated circuit (ASIC), a field programmable gate array (FPGA), a digital signal processor (DSP), a system on a chip (SoC), etc.

[0035] "Pre-training" a learning network refers to determining one or more parameters of at least one layer in the learning network. The learning network may include one or a combination of convolutional neural networks (CNNs), recurrent neural networks (RNNs), and recursive neural networks (RNNs). Convolutional neural networks are a type of feedforward neural network (FNN) that incorporates convolutional computations and has a deep structure. Convolutional neural networks have representational learning capabilities and can perform translation-invariant classification of input information based on their hierarchical structure. The convolution layer of a convolutional neural network model may include at least one filter or kernel. One or more parameters of at least one filter (such as kernel weight, size, shape, and structure) can be determined through a training process based on backpropagation, for example. A recursive neural network is a type of recursive neural network that takes sequence data as input, performs recursion in the direction of sequence evolution, and has all nodes (recurrent units) connected in a chain-like manner. A recursive neural network is an artificial neural network (ANN) with a tree-like hierarchical structure, in which network nodes recursively respond to input information in the order in which they are connected.

[0036] In some embodiments, the training method may further include evaluating, by the processor, the pre-trained learning network to determine whether it has a preset evaluation defect (step 106).

[0037] Specifically, the processor of the training device can evaluate the pre-trained learning network to determine whether the pre-trained learning network has preset evaluation defects. Here, evaluation defects can include significant differences in the detection rate of target objects at different locations and / or different attributes. By sampling and enhancing the relevant training data of target objects with lower detection rates than other locations and / or other attributes and adding them to the existing training data set, a training data set for stage-by-stage fine-tuning training targeting evaluation defects is obtained. Pre-set evaluation defects refer to issues of clinical concern, such as whether there are ruptures in the vascular branches, whether the coronary stents are fully segmented, whether the plaque is accurately segmented, etc. in the coronary artery segmentation task; or whether there is tissue label intermixing, whether a smooth tissue boundary is obtained, etc. in the cardiac multi-label tissue segmentation task; or whether the detection rate of ground-glass pulmonary nodules (or non-independent pulmonary nodules) meets the requirements in the pulmonary nodule detection task. By evaluating the pre-trained learning network, it is possible to determine whether the current network model performs well. Evaluation indicators can be set and adjusted according to the clinical concerns of doctors, and the training data set can be trained based on the evaluation indicators for targeted data enhancement (see below). For doctors, they usually do not have an in-depth understanding of deep learning algorithms. Compared with determining the abstract algorithm indicators (such as ROC) that need to be paid attention to in the training phase and setting the training method, it is easier and more user-friendly to identify clinical issues of concern. Targeted data enhancement based on the clinical issues suggested by doctors can enable the learning network to solve clinical issues of concern in a targeted manner, thereby improving the controllability, traceability and repeatability of the learning network, significantly shortening the development cycle, and increasing the interpretability of the learning network (especially its training phase) to people who are not algorithm experts (such as doctors).

[0038] It should be noted that the process of evaluating the pre-trained learning network can be completed manually or by a computer program. For example, the overlap rate between the coronary segmentation point cloud and the annotation result (such as the ground truth centerline point cloud) can be calculated to evaluate whether the blood vessel branch is broken. The embodiments of the present disclosure do not limit this.

[0039] Exemplarily, medical images may include vascular-related images, and medical image analysis may include segmentation of target arteries. Preset evaluation defects may include broken vascular branches, venous doping in arteries, uneven segmentation at plaques, and low implant recognition rate. Medical images may also include lung-related images, and medical image analysis may include lung nodule detection. Preset evaluation defects may include a lower detection rate of ground glass nodules than that of solid nodules, and a lower detection rate of nodules close to other tissues than that of nodules far from other tissues.

[0040] In some embodiments, the training method may further include, by the processor, performing data enhancement on the original training dataset based on the existing evaluation defects in the case where the pre-trained learning network has evaluation defects (step 108).

[0041] Specifically, in the case where the pre-trained learning network has evaluation defects, the processor of the training device can perform data augmentation (Data Augmentation) on the basis of the original training data set to address the existing evaluation defects. Here, data augmentation is mainly used to prevent overfitting. Through data augmentation, the amount of data can be rapidly increased and the trained model can have a certain noise resistance. Data augmentation methods may include rotation / reflection transformation, flip (horizontal or vertical) transformation, scaling transformation, translation transformation, scale transformation, contrast transformation, noise perturbation, color change, etc., which are not limited in the embodiments of the present disclosure. In some embodiments, specific data augmentation methods can be adopted to address the existing evaluation defects. For example, in coronary CTA, the intermediate ramus is very important for cardiovascular clinical diagnosis, but the training data of CTA usually only shows very short intermediate ramus, resulting in poor performance of the pre-trained learning network in detecting the intermediate ramus. In step 108, samples with shorter intermediate ramus can be screened out, and the areas near the intermediate ramus in these samples can be enhanced. The enhanced training data can make it easier for the learning network to focus on the intermediate ramus. For example, when performing lung nodule detection, the pre-trained learning network has a poor detection rate for ground-glass lung nodules, but ground-glass lung nodules are very important for the clinical diagnosis of lung diseases. In this way, in step 108, samples labeled with ground-glass lung nodules can be screened out from the original training data set, and the samples can be sampled, enhanced, and expanded and then added to the training data set. The enhanced training data can enable the learning network to better learn the image features of ground-glass lung nodules.

[0042] For example, if a pre-trained learning network has both venous contamination and broken blood vessel branches, the training device's processor can perform data augmentation to address these issues. By augmenting the original training dataset to address evaluation flaws in the pre-trained learning network, the model's generalization and robustness can be improved.

[0043] In some embodiments, the training method may further include a processor, based on the parameters of the pre-trained model, using a data-enhanced training data set to perform fine training on the learning network (step 110). Note that the learning network may be a learning network with a predetermined structure set in step 102, and the data enhancement processing may be based on the original training data set (even if the original training data set has defects such as sample imbalance or insufficient sample quantity). In this way, the user does not have to spend time and energy on designing, analyzing and testing the structure and loss function of the new learning network, but can spend less time and energy to implement the structural preset of the learning network; the training steps 106 and 108 of the learning network clarify the defects of the clinical evaluation indicators (which can be given according to the needs of the doctor) targeted at each training stage and automatically perform targeted data enhancement, thereby more flexibly and thoughtfully meeting the needs of clinical applications, while further reducing the time and energy spent by the user to determine the details of the training process.

[0044] Specifically, the processor of the training device can perform fine training on the learning network based on the parameters of the pre-trained model using the data-enhanced training data set, i.e., fine-tuning or fine-tuning. Here, fine training is the process of modifying the structure of the pre-trained learning network, selectively loading the weights of the pre-trained learning network, and then retraining the learning network using the data-enhanced training data set. In the embodiment of the present disclosure, the same loss function is used for each stage of fine training.

[0045] For example, when the training data set after data augmentation is large and similar to the original training data set, or when the training data set after data augmentation is large and dissimilar to the original training data set, the learning network can be finely trained. By finely training the learning network, the learning network can be quickly trained using a relatively small amount of data to obtain good training results.

[0046] It should be noted that the training process from step 102 to step 110 is also referred to as an “iterative training process”, which will be described in detail later.

[0047] According to the training method of a learning network for medical image analysis of an embodiment of the present disclosure, the learning network is pre-trained by using an original training data set to obtain a pre-trained learning network, and the pre-trained learning network is evaluated to determine whether it has preset evaluation defects. In the case that the pre-trained learning network has evaluation defects, data enhancement is performed on the original training data set based on the existing evaluation defects, and based on the parameters of the pre-trained model, the learning network is finely trained using the data-enhanced training data set. The learning network can be iteratively trained based on limited medical data, so as to quickly and accurately obtain medical images related to the analysis requirements, and the obtained medical images can be analyzed more accurately and more targeted. Therefore, the complexity of training and learning of medical images is reduced, and the efficiency and accuracy of medical image analysis are improved.

[0048] In some embodiments, when a pre-trained learning network has at least two evaluation defects, data augmentation is performed on the original training dataset and the training of the learning network using the data-augmented training dataset is performed in stages for each evaluation defect, and the training dataset used for fine training in each stage is data augmented for the corresponding evaluation defect and then used for fine training in the next stage.

[0049] Specifically, a pre-trained learning network may have at least two evaluation defects, such as venous contamination and broken vascular branches, or venous contamination, broken vascular branches, and uneven plaque segmentation. If the pre-trained learning network has at least two evaluation defects, the processor of the training device can perform data augmentation on the original training dataset and use the augmented training dataset to perform phased fine training on the learning network for each evaluation defect. Here, the training dataset used for each phase of fine training is augmented for the corresponding evaluation defect and then used for the next phase of fine training.

[0050] In some embodiments, the Nth learning network obtained after the Nth stage of fine training eliminates all evaluation defects targeted by the first stage of fine training to the Nth stage of fine training, N is a positive integer, and the staged execution includes staged execution until all preset evaluation defects are eliminated.

[0051] Specifically, after the Nth stage of fine training, an Nth learning network can be obtained, which has eliminated all evaluation defects targeted by the first through Nth stages of fine training, where N is a positive integer, i.e., an integer greater than or equal to 1. By performing staged fine training on the learning network, all preset evaluation defects can be eliminated.

[0052] In some embodiments, the training method of the learning network for medical image analysis further includes: evaluating the Nth learning network obtained after the Nth stage of fine training to determine any remaining evaluation defects.

[0053] Specifically, after the Nth stage of fine training, the processor of the training device can evaluate the obtained Nth learning network to determine any evaluation defects that still exist.

[0054] In some embodiments, the phased execution further includes: performing data augmentation on the original training data set for a first evaluation defect among at least two evaluation defects to obtain a first training data set; performing first-stage fine training on the learning network using the first training data set and based on pre-trained model parameters to obtain a first learning network; evaluating the first learning network to determine whether it has the first evaluation defect and the second evaluation defect; when the first learning network does not have the first evaluation defect but has the second evaluation defect, performing data augmentation on the first training data set for the second evaluation defect to obtain a second training data set; and performing second-stage fine training on the first learning network using the second training data set and based on the first learning network model parameters to obtain a second learning network.

[0055] Specifically, the processor of the training device can perform data enhancement on the original training data set for the first evaluation defect (for example, venous doping) among at least two evaluation defects (for example, venous doping and vascular branch fracture) to obtain a first training data set; then, use the first training data set and perform first-stage fine training on the learning network based on the pre-training model parameters to obtain a first learning network; by evaluating the first learning network, it can be determined whether the first learning network has the first evaluation defect and the second evaluation defect, and when the first evaluation defect does not exist in the first learning network but the second evaluation defect exists, the first training data set is data enhanced for the second evaluation defect to obtain a second training data set; further, the processor of the training device can use the second training data set and perform second-stage fine training on the first learning network based on the first learning network model parameters to obtain a second learning network.

[0056] In some embodiments, when N is greater than 1, the Nth stage fine training further includes: performing data augmentation on the N-1th training dataset for the Nth evaluation defect to obtain the Nth training dataset; using the Nth training dataset and based on the N-1th learning network model parameters, performing the Nth stage fine training on the N-1th learning network to obtain the Nth learning network.

[0057] Specifically, when N is greater than 1, the processor of the training device can perform data enhancement on the N-1th training data set for the Nth evaluation defect to obtain the Nth training data set; further, the processor of the training device can use the Nth training data set and perform N-1th learning network fine training based on the N-1th learning network model parameters to obtain the Nth learning network.

[0058] It should be noted that the N+1th training data set may include the Nth training data set while adding the above-mentioned targeted new data cluster, or the number of the Mth training data set may be reduced first and then the above-mentioned targeted new data cluster may be added. The embodiments of the present disclosure do not limit this.

[0059] In some embodiments, the assessment defects targeted by each stage of fine-tuning training are set so that the importance of the assessment defects targeted in the previous stage is higher than the importance of the assessment defects targeted in the subsequent stage. In some embodiments, the priority of the importance of the assessment defects can be determined based on the specific clinical needs of the physician. For example, in the segmentation and reconstruction task of clinical CCTA images, physicians sometimes focus more on venous contamination in arteries than on vessel segmentation fractures caused by artifacts. In other words, the assessment defect of venous contamination in arteries is more important than the assessment defect of vessel segmentation fractures caused by artifacts. Therefore, the assessment defects targeted in the previous and subsequent stages can be set accordingly to better meet the physician's specific clinical needs. In some embodiments, the priority of the importance of the assessment defects can be further determined by considering the interactions between the various assessment defects. Still using the clinical CCTA image segmentation and reconstruction task as an example, by performing data augmentation on venous contamination in arteries in the previous stage, the training dataset after data augmentation significantly reduces the number of veins mis-segmented as arteries, thereby reducing the workload of vessel segmentation fractures in the subsequent stage. Conversely, performing data augmentation on vessel segmentation fractures in the previous stage increases the frequency of venous contamination, thereby increasing the overall training workload.

[0060] Specifically, the processor of the training device can pre-set the evaluation defects targeted by each stage of fine-tuning training, so that the importance of the evaluation defects targeted by the previous stage is higher than the importance of the evaluation defects targeted by the later stage. By determining the order of execution of each stage based on the evaluation defects, the evaluation defects with higher importance can be eliminated first.

[0061] In some embodiments, when there is a third evaluation defect that gives priority to the false positive rate over the missed detection rate and the missed detection rate among the preset evaluation defects, and a fourth evaluation defect that gives priority to the false positive rate over the false positive rate and the missed detection rate, the third evaluation defect is set so that the training stage for it is before the training stage for the fourth evaluation defect.

[0062] Specifically, when there are false positive rates and missed detection rates in the preset evaluation defects, if the missed detection rate is given priority, the priority of the third evaluation defect corresponding to the missed detection rate is set to be higher than the priority of the fourth evaluation defect corresponding to the false positive rate; further, based on the priorities of the third evaluation defect and the fourth evaluation defect, the execution order of the training stage for the third evaluation defect can be set before the execution order of the training stage for the fourth evaluation defect, that is, the training stage for the third evaluation defect is executed first.

[0063] In some embodiments, the training data set used for fine-tuning training in each stage is obtained by: based on the evaluation results of the fine-tuning in the previous stage, screening out the training data set used for fine-tuning training in the previous stage that has the evaluation defects targeted by fine-tuning training in the current stage; and performing more sampling and enhancement on the screened training data and adding it to the training data set used for fine-tuning training in the previous stage.

[0064] Specifically, the processor of the training device can filter out the training data that has evaluation defects targeted by the current stage of fine training in the training data set used for the previous stage of fine training based on the evaluation results of the previous stage of fine training; further, perform more sampling and enhancement on the filtered training data, and add the enhanced training data to the training data set used for the previous stage of fine training.

[0065] In some embodiments, obtaining the training data set used for fine training in each stage further includes: sampling and enhancing the filtered training data and adding it to the training data set used for fine training in the previous stage, and deleting a portion of other training data in the training data set used for fine training in the previous stage.

[0066] Specifically, after further sampling and enhancement of the filtered training data and adding the enhanced training data to the training data set used for the previous stage of fine training, the processor of the training device can delete part of the other training data in the training data set used for the previous stage of fine training to save storage space of the training device and reduce the amount of data computing.

[0067] In some embodiments, the training data set used for fine training in each stage is obtained by: collecting more original data based on the evaluation results of the fine training in the previous stage, and screening out training data that have evaluation defects targeted by the fine training in the current stage from the original data; and performing more sampling and enhancement on the screened training data and adding it to the training data set used for fine training in the previous stage.

[0068] Specifically, the processor of the training device can collect more raw data based on the evaluation results of the previous stage of fine training, and filter out training data for the evaluation defects targeted by the current stage of fine training from the raw data; further, the processor of the training device can perform more sampling and enhancement on the filtered training data, and add the enhanced training data to the training data set used for the previous stage of fine training.

[0069] The following describes in detail the training method of the learning network for medical image analysis through specific examples.

[0070] Figure 2(a) to Figure 2(c) A schematic diagram of vein doping according to an embodiment of the present disclosure is shown. Taking the input medical image as a coronary computed tomography angiography (CCTA) image as an example, as shown in FIG2(a), in the CCTA image, since the contrast intensity of the artery is close to the contrast intensity of the vein, the model (i.e., the learning network described above) cannot correctly distinguish between arteries and veins, or mistakenly identifies the vein as an artery. Therefore, vein doping exists in the final artery segmentation, as shown in FIG2(b). After the model is directional-adjusted using an iterative training strategy, the new model has the ability to correctly distinguish between arteries and veins, so that the doped veins can be identified and removed, as shown in FIG2(c). Therefore, the problem of vein doping is solved.

[0071] Figure 3(a) to Figure 3(c) A schematic diagram of a vascular branch fracture according to an embodiment of the present disclosure is shown. Still taking the CCTA image as an example, as shown in FIG3(a), in the CCTA image, due to the presence of motion artifacts, the model is unable to correctly segment the coronary vessels at the artifact location, further resulting in a coronary branch fracture in the 3D vascular reconstruction result, as shown in FIG3(b). After using an iterative training strategy to adjust the model in a targeted manner, the new model has the ability to correctly segment the coronary vessels at the artifact location, allowing the coronary vessels at the artifact location to be correctly reconstructed, as shown in FIG3(c), thereby solving the problem of vascular branch fracture.

[0072] Figures 4(a) and 4(b) show schematic diagrams of solid nodules and ground-glass nodules according to an embodiment of the present disclosure. Taking the input medical image as a lung computed tomography (CT) image, in the lung CT image, since solid nodules have a density similar to that of human soft tissue and are clearly presented in the lung CT image, while ground-glass nodules are translucent and fuzzy in the lung CT image, the model can correctly detect solid nodules but cannot detect ground-glass nodules, as shown in Figure 4(a); after using an iterative training strategy to adjust the model in a targeted manner, the new model has targetedly strengthened the learning of ground-glass nodules, and therefore can correctly detect solid nodules and ground-glass nodules, as shown in Figure 4(b).

[0073] Figure 5 Schematic diagram of the iterative training process according to an embodiment of the present disclosure is shown. Figure 5 As shown, at the beginning of iterative training, the existing training data is first named the first training dataset and input into the deep learning convolutional network. A 3D convolution kernel is used to extract the feature map of the input data, and the error is back-propagated by comparing it with the gold standard image to update the fixed model parameters. The model is pre-trained by repeating the above steps, and the pre-trained fixed model is named the first model. Next, the first model is evaluated to obtain an evaluation result. The evaluation result is compared with the ground truth to determine the evaluation defects of the first model. Furthermore, if the first model has N evaluation defects, the first training dataset is sampled and enhanced for the i-th evaluation defect among the N evaluation defects, and the sampled and enhanced data is added to the first training dataset to obtain a second training dataset. The above steps are repeated, the fixed model is used to load the N-th model parameters, and the fixed model is fine-tuned using the N+1-th training dataset until the training converges, that is, all evaluation defects are alleviated or eliminated.

[0074] Figure 6 FIG. 1 is a flow chart showing a method for training a learning network for medical image analysis according to an embodiment of the present disclosure. Figure 6As shown, the training method includes the following steps. The method begins with the step of setting up a learning network with a predetermined structure and obtaining an original training dataset (step 602). The method may also include the step of pre-training the learning network using the original training dataset to obtain a pre-trained learning network (step 604). The method may also include the step of evaluating the pre-trained learning network to determine whether it has a preset evaluation defect (step 606). The method may also include, if the pre-trained learning network has at least two evaluation defects, performing data augmentation on the original training dataset for a first evaluation defect of the at least two evaluation defects to obtain a first training dataset (step 608). The method may also include the step of performing a first-stage fine training on the learning network using the first training dataset and based on pre-trained model parameters to obtain the first learning network (step 610). The method may also include the step of evaluating the first learning network to determine whether it has the first evaluation defect and the second evaluation defect (step 612). The method may also include, if the first learning network does not have the first evaluation defect but has the second evaluation defect, performing data augmentation on the first training dataset for the second evaluation defect to obtain a second training dataset (step 614). The method may further include performing a second stage of fine training on the first learning network using a second training data set and based on the first learning network model parameters to obtain a second learning network (step 616).

[0075] According to the training method of a learning network for medical image analysis of an embodiment of the present disclosure, a learning network is pre-trained using an original training dataset to obtain a pre-trained learning network, the pre-trained learning network is evaluated to determine whether it has preset evaluation defects, and when the pre-trained learning network has at least two evaluation defects, the original training dataset is data-enhanced for a first evaluation defect of the at least two evaluation defects to obtain a first training dataset, the learning network is first-stage fine-trained using the first training dataset and based on pre-trained model parameters to obtain a first learning network, the first learning network is evaluated to determine whether it has the first evaluation defect and a second evaluation defect, and when the first learning network does not have the first evaluation defect but has the second evaluation defect, the first training dataset is data-enhanced for the second evaluation defect to obtain a second training dataset, and the first learning network is second-stage fine-trained using the second training dataset and based on the first learning network model parameters to obtain a second learning network. The learning network can be trained in stages, thereby quickly and accurately acquiring medical images relevant to analysis requirements and performing more precise and targeted analysis on the acquired medical images. Therefore, the complexity of medical image training and learning is reduced, and the efficiency and accuracy of medical image analysis are improved.

[0076] The following are device embodiments of the present disclosure, which can be used to implement the method embodiments of the present disclosure. For details not disclosed in the device embodiments of the present disclosure, please refer to the method embodiments of the present disclosure.

[0077] Figure 7 FIG. 5 is a block diagram of a training system for a learning network for medical image analysis according to an embodiment of the present disclosure. Figure 7 As shown, the training system may include a communication interface 702 and a processor 704 .

[0078] Specifically, the communication interface 702 is configured to obtain an original training data set, and the processor 704 is configured to use the original training data set to pre-train a learning network with a predetermined structure to obtain a pre-trained learning network; evaluate the pre-trained learning network to determine whether it has preset evaluation defects; if the pre-trained learning network has evaluation defects, perform data enhancement on the original training data set based on the existing evaluation defects; and based on the parameters of the pre-trained model, use the data-enhanced training data set to fine-tune the learning network. In some embodiments, the processor 704 can also be configured to present a user interface to prompt the user to input settings for the structure of the learning network. For example, a list or menu of various structures of the learning network can be presented on the user interface for the user to select, and the learning network with the predetermined structure selected by the user is used for the subsequent training process.

[0079] According to the training system of the learning network for medical image analysis of the embodiment of the present disclosure, the learning network is pre-trained by using the original training data set to obtain a pre-trained learning network, and the pre-trained learning network is evaluated to determine whether it has preset evaluation defects. In the case that the pre-trained learning network has evaluation defects, data enhancement is performed on the original training data set based on the existing evaluation defects, and the learning network is finely trained using the data-enhanced training data set based on the parameters of the pre-trained model. The learning network can be iteratively trained based on limited medical data, so as to quickly and accurately obtain medical images related to the analysis requirements, and the obtained medical images can be analyzed more accurately and more targeted. Therefore, the complexity of training and learning of medical images is reduced, and the efficiency and accuracy of medical image analysis are improved.

[0080] In some embodiments, the processor 704 is further configured to, when a pre-trained learning network has at least two evaluation defects, perform data augmentation on the original training data set and use the data-augmented training data set to train the learning network in stages for each evaluation defect, and the training data set used for the fine training in each stage is data augmented for the corresponding evaluation defect and then used for the fine training in the next stage.

[0081] In some embodiments, the Nth learning network obtained after the Nth stage of fine training eliminates all evaluation defects targeted by the first stage of fine training to the Nth stage of fine training, N is a positive integer, and the staged execution includes staged execution until all preset evaluation defects are eliminated.

[0082] In some embodiments, the processor 704 is further configured to: evaluate the Nth learning network obtained after the Nth stage of fine training to determine the evaluation defects that still exist therein.

[0083] In some embodiments, the processor 704 is further configured to: perform data enhancement on the original training data set for a first evaluation defect among at least two evaluation defects to obtain a first training data set; perform first-stage fine training on the learning network using the first training data set and based on pre-trained model parameters to obtain a first learning network; evaluate the first learning network to determine whether it has the first evaluation defect and the second evaluation defect; if the first evaluation defect does not exist in the first learning network but the second evaluation defect exists, perform data enhancement on the first training data set for the second evaluation defect to obtain a second training data set; and perform second-stage fine training on the first learning network using the second training data set and based on the first learning network model parameters to obtain a second learning network.

[0084] In some embodiments, when N is greater than 1, the processor 704 is further configured to: perform data augmentation on the N-1th training dataset for the Nth evaluation defect to obtain the Nth training dataset; and perform N-stage fine training on the N-1th learning network based on the N-1th learning network model parameters using the Nth training dataset to obtain the Nth learning network.

[0085] In some embodiments, the processor 704 is further configured to set the evaluation defects targeted by the fine training in each stage so that the importance of the evaluation defects targeted by the previous stage is higher than the importance of the evaluation defects targeted by the next stage.

[0086] In some embodiments, the processor 704 is further configured to, when there is a third evaluation defect that gives priority to the missed detection rate over the false positive rate and the missed detection rate among the preset evaluation defects, and a fourth evaluation defect that gives priority to the false positive rate over the false positive rate and the missed detection rate, set the third evaluation defect so that the training stage for it is before the training stage for the fourth evaluation defect.

[0087] In some embodiments, the defect assessment includes significant differences in detection rates of target objects at different locations and / or with different attributes. The relevant training data of target objects with lower detection rates than other locations and / or other attributes are sampled and enhanced more and added to the existing training data set to obtain a training data set for stage-by-stage fine-tuning training of defect assessment.

[0088] In some embodiments, the processor 704 is further configured to: based on the evaluation results of the previous stage of fine training, filter out the training data used in the previous stage of fine training that has the evaluation defects targeted by the current stage of fine training; and perform more sampling and enhancement on the filtered training data and add it to the training data set used in the previous stage of fine training.

[0089] In some embodiments, the processor 704 is further configured to: perform more sampling and enhancement on the filtered training data and add it to the training data set used for the previous stage of fine training, and delete part of other training data in the training data set used for the previous stage of fine training.

[0090] In some embodiments, the processor 704 is further configured to: collect more raw data based on the evaluation results of the previous stage of fine training, and filter out training data that have evaluation defects targeted by the current stage of fine training from the raw data; and perform more sampling and enhancement on the filtered training data and add it to the training data set used for the previous stage of fine training.

[0091] In some embodiments, each stage of fine-tuning uses the same loss function.

[0092] In some embodiments, the medical image includes a blood vessel-related image, the medical image analysis includes segmentation of the target artery, and the preset evaluation defects include broken blood vessel branches, venous doping in the artery, uneven segmentation at the plaque, and low implant recognition rate.

[0093] In some embodiments, the medical image includes a lung-related image, the medical image analysis includes lung nodule detection, and the preset evaluation defects include that the detection rate of ground glass nodules is lower than the detection rate of solid nodules and the detection rate of nodules close to other tissues is lower than the detection rate of nodules far from other tissues.

[0094] Figure 8 FIG. 1 shows a structural block diagram of a training device according to an embodiment of the present disclosure. Figure 8As shown, the training device is a general-purpose data processing device, including a general-purpose computer hardware structure, and the training device includes at least a processor 802 and a memory 804. The processor 802 and the memory 804 are connected via a bus 806. The memory 804 is suitable for storing instructions or programs executable by the processor 802. The processor 802 can be an independent microprocessor or a collection of one or more microprocessors. Thus, the processor 802 executes the commands stored in the memory 804, thereby executing the method flow of the embodiment of the present disclosure as described above to process data and control other devices. The bus 806 connects the above-mentioned multiple components together, and at the same time connects the above-mentioned components to the display controller 808 and the display device and the input / output (I / O) device 810. The input / output (I / O) device 810 can be a mouse, keyboard, modem, network interface, touch input device, somatosensory input device, printer, and other devices known in the art. Typically, the input / output (I / O) device 810 is connected to the system via an input / output (I / O) controller 812.

[0095] The memory 804 may store software components, such as an operating system, a communication module, an interaction module, and an application program. Each of the above modules and applications corresponds to a set of executable program instructions that implement one or more functions and methods described in the embodiments of the invention.

[0096] In some embodiments, the training device may be located in one place, distributed in multiple places, or may be a distributed training device, such as set up in the cloud, and the embodiments of the present disclosure are not limited to this.

[0097] The flowcharts and / or block diagrams of the methods, systems, and computer program products according to the embodiments of the present disclosure have been described above to provide various aspects of the present disclosure. It should be understood that each block of the flowchart and / or block diagram and the combination of the blocks in the flowchart legend and / or block diagram can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing device to produce a machine so that instructions (executed by a processor of a computer or other programmable data processing device) are created to implement the function / action specified in the flowchart and / or block diagram block or block.

[0098] At the same time, as will be appreciated by those skilled in the art, various aspects of the embodiments of the present disclosure may be implemented as systems, methods, or computer program products. Thus, various aspects of the embodiments of the present disclosure may take the form of a complete hardware implementation, a complete software implementation (including firmware, resident software, microcode, etc.), or an implementation that combines software aspects with hardware aspects, which may all be generally referred to herein as a "circuit," "module," or "system." Furthermore, aspects of the present disclosure may take the form of a computer program product implemented in one or more computer-readable media having computer-readable program code implemented thereon.

[0099] Any combination of one or more computer-readable media can be utilized. The computer-readable medium can be a computer-readable signal medium or a computer-readable storage medium. The computer-readable storage medium can be, for example, but not limited to, an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, device, or apparatus, or any appropriate combination thereof. More specific examples (non-exhaustive enumeration) of computer-readable storage media will include the following: an electrical connection with one or more wires, a portable computer floppy disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disc read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any appropriate combination thereof. In the context of the embodiments of the present disclosure, a computer-readable storage medium can be any tangible medium that can contain or store a program used by an instruction execution system, device, or apparatus, or a program used in conjunction with an instruction execution system, device, or apparatus.

[0100] A computer-readable signal medium may include a propagated data signal having computer-readable program code embodied therein, such as in baseband or as part of a carrier wave. Such a propagated signal may take any of a variety of forms, including but not limited to electromagnetic, optical, or any suitable combination thereof. A computer-readable signal medium may be any computer-readable medium that is not a computer-readable storage medium and that can communicate, propagate, or transmit a program for use by or in conjunction with an instruction execution system, device, or apparatus.

[0101] Program code embodied on a computer readable medium may be transmitted using any appropriate medium, including but not limited to wireless, wireline, optical fiber cable, RF, etc., or any suitable combination of the foregoing.

[0102] The computer program code for performing operations for various aspects of the present disclosure can be written in any combination of one or more programming languages, including object-oriented programming languages ​​such as Java, Smalltalk, C++, PHP, Python, and the like, and conventional procedural programming languages ​​such as the "C" programming language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a stand-alone software package; partially on the user's computer and partially on a remote computer; or entirely on a remote computer or server. In the latter case, the remote computer can be connected to the user's computer via any type of network, including a local area network (LAN) or a wide area network (WAN), or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0103] These computer program instructions may also be stored in a computer-readable medium that can direct a computer, other programmable data processing device, or other apparatus to operate in a particular manner, so that the instructions stored in the computer-readable medium produce an article of manufacture that includes instructions for implementing the functions / actions specified in the flowchart and / or block diagram blocks or blocks.

[0104] The computer program instructions may also be loaded onto a computer, other programmable data processing device or other apparatus to cause a series of operable steps to be performed on the computer, other programmable device or other apparatus to produce a computer-implemented process, such that the instructions executed on the computer or other programmable device provide a process for implementing the functions / actions specified in the flowchart and / or block diagram block or blocks.

[0105] Furthermore, although exemplary embodiments have been described herein, the scope includes any and all embodiments based on the present disclosure with equivalent elements, modifications, omissions, combinations (e.g., solutions that intersect various embodiments), adaptations, or changes. The elements in the claims are to be interpreted broadly based on the language employed in the claims and are not limited to the examples described in this specification or during the prosecution of this application, which examples are to be interpreted as non-exclusive. Therefore, this specification and examples are intended to be considered as examples only, with the true scope and spirit being indicated by the following claims and the full scope of their equivalents.

[0106] The above description is intended to be illustrative and not restrictive. For example, the above examples (or one or more schemes thereof) can be used in combination with each other. For example, a person of ordinary skill in the art may use other embodiments when reading the above description. In addition, in the above-mentioned specific embodiments, various features can be grouped together to simplify the present disclosure. This should not be interpreted as an intention that a disclosed feature that is not required to be protected is necessary for any claim. On the contrary, the subject matter of the present invention may be less than all the features of a specific disclosed embodiment. Thus, the following claims are incorporated into the specific embodiments as examples or embodiments, wherein each claim is independently a separate embodiment, and it is considered that these embodiments can be combined with each other in various combinations or arrangements. The scope of the present invention should be determined with reference to the appended claims and the full scope of equivalents to which these claims are entitled.

Claims

1. A training method for a learning network for medical image analysis, characterized in that: The training method comprises: Setting a learning network with a predetermined structure and obtaining an original training data set, wherein the learning network includes one or a combination of a convolutional neural network, a recurrent neural network, and a recursive neural network; The processor pre-trains the learning network using the original training data set to obtain a pre-trained learning network; The processor evaluates the pre-trained learning network to determine whether it has a preset evaluation defect; The processor performs data enhancement on the original training data set to address the evaluation defects when the pre-trained learning network has evaluation defects. The processor sets each assessment defect or the priority of each assessment defect according to the clinical needs of the doctor; The processor performs, based on parameters of the pre-trained model and the set priorities of the assessment defects, stage-by-stage fine training of the learning network using the data-augmented training data set, such that a training stage for high-priority assessment defects is performed before a training stage for low-priority assessment defects; In the case where the pre-trained learning network has at least two evaluation defects, data augmentation is performed on the original training data set and the training of the learning network is performed in stages using the data-augmented training data set for each evaluation defect, and the training data set used for the fine training in each stage is data augmented for the corresponding evaluation defect and then used for the fine training in the next stage; When there is a third evaluation defect in the preset evaluation defects that gives priority to the missed detection rate over the false positive rate and the missed detection rate, and a fourth evaluation defect that gives priority to the false positive rate over the false positive rate and the missed detection rate, the third evaluation defect is set so that the training stage for it is before the training stage for the fourth evaluation defect.

2. The training method according to claim 1, characterized in that The Nth learning network obtained after the Nth stage of fine training eliminates all evaluation defects targeted by the first stage of fine training to the Nth stage of fine training, N is a positive integer, and the staged execution includes staged execution until all preset evaluation defects are eliminated.

3. The training method according to claim 2, characterized in that Also includes: The Nth learning network obtained through the Nth stage of fine training is evaluated to determine any remaining evaluation defects.

4. The training method according to claim 1, characterized in that The phased execution further includes: performing data enhancement on the original training data set for a first evaluation defect among the at least two evaluation defects to obtain a first training data set; Using the first training data set and based on pre-trained model parameters, the learning network is subjected to a first stage of fine training to obtain a first learning network; Evaluating the first learning network to determine whether it has the first evaluation defect and the second evaluation defect; When the first learning network does not have the first evaluation defect but has the second evaluation defect, performing data enhancement on the first training data set for the second evaluation defect to obtain a second training data set; and The second training data set is used to perform second-stage fine training on the first learning network based on the first learning network model parameters to obtain a second learning network.

5. The training method according to claim 2, characterized in that: When N is greater than 1, the Nth stage of fine training further includes: Perform data augmentation on the N-1th training dataset for the Nth evaluation defect to obtain the Nth training dataset; The Nth stage fine training is performed on the N-1th learning network using the Nth training data set and based on the N-1th learning network model parameters to obtain the Nth learning network.

6. The training method according to any one of claims 1 to 5, characterized in that: The evaluation defects targeted by the detailed training in each stage are set so that the importance of the evaluation defects targeted by the previous stage is higher than the importance of the evaluation defects targeted by the next stage.

7. The training method according to any one of claims 1 to 5, characterized in that: The evaluation defect includes significant differences in detection rates of target objects at different locations and / or with different attributes. The relevant training data of target objects with lower detection rates than other locations and / or other attributes are sampled and enhanced and added to the existing training data set to obtain a training data set for stage-by-stage fine training of the evaluation defect.

8. The training method according to claim 3, characterized in that: The training datasets used for the detailed training at each stage are obtained as follows: Based on the evaluation results of the previous stage of fine training, screening out the training data set used for the previous stage of fine training that has the evaluation defects targeted by the current stage of fine training; as well as The filtered training data are further sampled and enhanced and added to the training data set used for the previous stage of fine training.

9. The training method according to claim 8, characterized in that Obtaining the training data set used for the fine training in each stage further includes: performing more sampling and enhancement on the screened training data and adding it to the training data set used for the fine training in the previous stage, and deleting a portion of other training data in the training data set used for the fine training in the previous stage.

10. The training method according to claim 3, characterized in that: The training datasets used for the detailed training at each stage are obtained as follows: Based on the evaluation results of the previous stage of refined training, more raw data is collected, and training data with evaluation defects targeted by the current stage of refined training is screened from the raw data; as well as The filtered training data are further sampled and enhanced and added to the training data set used for the previous stage of fine training.

11. The training method according to any one of claims 1 to 5, characterized in that: The same loss function is used in the fine-tuning training of each stage.

12. The training method according to any one of claims 1 to 5, characterized in that: The medical image includes a blood vessel-related image, the medical image analysis includes segmentation of a target artery, and preset evaluation defects include vascular branch breakage, venous contamination in the artery, uneven segmentation at the plaque, and low implant recognition rate.

13. The training method according to any one of claims 1 to 5, characterized in that: The medical image includes a lung-related image, and the medical image analysis includes lung nodule detection. Preset evaluation defects include that the detection rate of ground glass nodules is lower than that of solid nodules, and that the detection rate of nodules close to other tissues is lower than that of nodules far from other tissues.

14. A training system for a learning network for medical image analysis, characterized in that: The training system includes a communication interface and a processor, wherein: The communication interface is configured to obtain an original training data set; The processor is configured to execute the training method for a learning network for medical image analysis according to any one of claims 1 to 13.

15. A training device for a learning network for medical image analysis, comprising a memory and a processor, characterized in that: The memory is used to store one or more computer program instructions, wherein the one or more computer program instructions are executed by the processor to implement the operations performed by the training method for a learning network for medical image analysis according to any one of claims 1 to 13.

16. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the operations performed by the training method for a learning network for medical image analysis according to any one of claims 1 to 13 are implemented.

Images