Automatic image segmentation system and method
By dividing medical image data into batches and updating weights, using the block coordinate descent optimization solution, the problem that deep neural networks are difficult to utilize external data in medical image segmentation is solved, and more efficient image segmentation effect and computing performance are achieved.
Patent Information
- Application Number
- CN202310609875.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2023-05-26
- Publication Date
- 2024-12-03
AI Technical Summary
When training deep neural networks, existing medical image segmentation methods are difficult to effectively utilize external data, resulting in the model tending to be easy to learn modes and ignore useful information, or are affected by outliers, and the calculation cost is high.
By receiving training data from a trusted source and a second corpus that receives training data from an alternative source, initial weight is assigned to each data item in the second corpus, the data is divided into batches, and the output is generated using the learning network , calculate the loss value and difference of each batch, update the weight to minimize the loss value, adopt the block coordinate descent optimization scheme to weight the importance weight and network parameters, avoid calculating the second-order gradient of the network.
Effectively utilize external data, improve the accuracy and robustness of medical image segmentation, reduce calculation costs, reduce data collection burden, and suppress the impact of irrelevant patterns.
Smart Images

Figure CN119068180A_ABST
Abstract
Description
Technical Field
[0001] The present disclosure relates to an image segmentation system and method, which is particularly suitable for analyzing medical images. Background Art
[0002] Image segmentation is a method of dividing a digital image into segments that can be further processed or analyzed. Computer-assisted volumetric image segmentation of medical images is a particularly challenging type of image segmentation. Typically, computer-assisted image segmentation supports the daily work of radiologists who visually analyze anatomical structures in medical images and discern subtle changes in size, shape, and structure that may indicate a disease state in the subject.
[0003] A typical approach to training machine learning-based (deep neural network) image segmentation of medical images is to use large internal image datasets (e.g., computed tomography or CT images) with voxel-level annotations of target structures performed by domain experts (such as radiologists). It should be recognized that the preparation of this internal training data is laborious and time-consuming; and may even require multiple rounds of consultations between experts to reach a consensus.
[0004] It is possible to supplement such internal data with external data, including CT images from websites or other institutions. However, this is complicated by differences in image quality and relevance, and the difficulty in distinguishing useful data from harmful data from such images. One solution is to discard such external data with large loss values during training, but this distorts machine learning to favor patterns that are easy to learn; thereby ignoring informative data with useful difficult patterns that can make deep models more accurate and robust. At the same time, it should be recognized that if external data with large loss values are not discarded, the relevant learning may be unduly affected by outliers of difficult patterns.
[0005] Various methods have been used to learn weights from external data, including detection via gradient descent using mini-batches of data; however, such methods are computationally expensive because they require computing the second-order gradients of the network.
[0006] The object of the present disclosure is to solve or at least partially ameliorate some of the above-mentioned problems of current approaches. Summary of the invention
[0007] The features and advantages of the present disclosure will be set forth in the following description, and in part will be apparent from the description, or may be learned by practicing the principles disclosed herein. The features and advantages of the present disclosure may be realized and obtained by the instruments and combinations particularly pointed out in the appended claims.
[0008] According to a first aspect of the present disclosure, there is provided a computer-implemented method for training a learning network to perform image segmentation, the method comprising: receiving a first corpus of training data from a trusted source; receiving a second corpus of training data from at least one alternative source; assigning an initial weight to each data item in the second corpus; dividing the training data into one or more batches, wherein each batch includes at least one sample from the first corpus and at least one sample from the second corpus; and completing a predefined number of training epochs, wherein, for each batch, each training epoch comprises: generating an output for each sample in the batch using the learning network; determining a weight for each sample in the batch by comparing the generated output to the true data for the sample. The method comprises the steps of: defining a difference for each second corpus sample in the batch as the difference between the loss value of the sample and the average loss of all first corpus samples, wherein the average loss of all first corpus samples is approximated as the average loss of all first corpus samples from the previous training epoch modified by the change in the average loss of the first corpus samples in the batch from the previous training epoch to the current training epoch; identifying a set of updated weights for the next batch that minimizes a weighted sum of the differences for each second corpus sample; and identifying updated network parameters for the learning network that minimize the average loss for each first corpus sample in the batch and the average weighted loss for each second corpus sample in the batch.
[0009] The method may include training the learning network using a first corpus of training data to obtain a first optimal network, wherein network parameters are identified for each batch subject to a first constraint that an average loss for each first corpus sample in the batch must be less than an average loss for each first corpus sample processed by the first optimal network.
[0010] The network parameters may be identified by gradient descent using the sum of the weighted losses for the batch and a first constraint term representing the first constraint.
[0011] The first constraint term can be determined by: comparing the loss value of each first corpus sample in the batch with the loss value of the sample when processed by the first optimal network; and for one or more loss values of the first corpus samples in the batch that are greater than the corresponding loss value of the sample when processed by the first optimal network, summing the differences between the compared loss values to generate the first constraint term.
[0012] This first constraint term can be modified by a penalty coefficient that starts out as a small value and is updated to increase with each batch.
[0013] The penalty coefficient may be updated to increase a certain value based on the first constraint item of the batch.
[0014] The penalty coefficient may be weighted to taper off after a predetermined number of epochs have been completed.
[0015] The network parameters may be identified for each batch subject to a second constraint that the sum of the weights is non-zero.
[0016] The identified set of updated weights may be combined with a factor based on the scalar length of the vector formed by the plurality of weights to minimize the weighted sum of the differences.
[0017] The difference of each second corpus sample may be further multiplied by the difference between the loss value of the sample in the current epoch and the loss value of the sample in the previous epoch.
[0018] Each sample may be a computed tomography (CT) scan.
[0019] In this first epoch, the average loss of all first corpus samples may be set to 1.
[0020] According to a second aspect of the present disclosure, there is provided a computer-implemented method for performing image segmentation, the method comprising: training a learning network to perform image segmentation by receiving a first corpus of training data from a trusted source; receiving a second corpus of training data from at least one alternative source; assigning an initial weight to each data item in the second corpus; dividing the training data into one or more batches, wherein each batch includes at least one sample from the first corpus and at least one sample from the second corpus; and completing a predefined number of training epochs, wherein, for each batch, each training epoch comprises: generating an output for each sample in the batch using the learning network; determining a loss value for each sample in the batch by comparing the generated output with the true data of the sample; and performing a weighting on each sample in the batch. The difference of the second corpus sample is defined as the difference between the loss value of the sample and the average loss of all first corpus samples, wherein the average loss of all first corpus samples is approximated as the average loss of all first corpus samples from the previous training epoch modified by the change in the average loss of the first corpus samples in the batch from the previous training epoch to the current training epoch; identifying a set of updated weights for the next batch that minimizes a weighted sum of the differences of each second corpus sample; identifying updated network parameters for the learning network that minimize the average loss of each first corpus sample in the batch and the average weighted loss of each second corpus sample in the batch; receiving one or more samples of image data; and generating an output of segmented image data for each received sample.
[0021] According to a third aspect of the present disclosure, a computer-readable medium is provided, the computer-readable medium comprising instructions that, when executed by a processor, cause the processor to train a learning network to perform image segmentation by: receiving a first corpus of training data from a trusted source; receiving a second corpus of training data from at least one alternative source; assigning an initial weight to each data item in the second corpus; dividing the training data into one or more batches, wherein each batch includes at least one sample from the first corpus and at least one sample from the second corpus; and completing a predefined number of training epochs, wherein, for each batch, each training epoch includes: generating an output for each sample in the batch using the learning network; and performing a training on the generated output by comparing the generated output to the real data of the sample. performing the comparison to determine a loss value for each sample in the batch; defining a difference for each second corpus sample in the batch as the difference between the loss value of the sample and an average loss for all first corpus samples, wherein the average loss for all first corpus samples is approximated as the average loss for all first corpus samples from a previous training epoch modified by a change in the average loss for the first corpus samples in the batch from the previous training epoch to the current training epoch; identifying a set of updated weights for a next batch that minimizes a weighted sum of the differences for each second corpus sample; identifying updated network parameters for the learning network that minimize the average loss for each first corpus sample in the batch and the average weighted loss for each second corpus sample in the batch.
[0022] According to a fourth aspect of the present disclosure, there is provided an image segmentation network, which is trained by: receiving a first corpus of training data from a trusted source; receiving a second corpus of training data from at least one alternative source; assigning an initial weight to each data item in the second corpus; dividing the training data into one or more batches, wherein each batch includes at least one sample from the first corpus and at least one sample from the second corpus; and completing a predefined number of training epochs, wherein, for each batch, each training epoch includes: generating an output for each sample in the batch using the learning network; determining each sample in the batch by comparing the generated output with the true data of the sample. loss value of the first corpus sample in the batch; defining the difference of each second corpus sample in the batch as the difference between the loss value of the sample and the average loss of all first corpus samples, wherein the average loss of all first corpus samples is approximated as the average loss of all first corpus samples from the previous training epoch modified by the change in the average loss of the first corpus samples in the batch from the previous training epoch to the current training epoch; identifying a set of updated weights for the next batch that minimizes a weighted sum of the differences of each second corpus sample; identifying updated network parameters for the learning network that minimize the average loss of each first corpus sample in the batch and the average weighted loss of each second corpus sample in the batch. BRIEF DESCRIPTION OF THE DRAWINGS
[0023] In order to describe the manner in which the above and other advantages and features of the present disclosure can be obtained, a more particular description of the principles briefly described above will be presented by reference to specific embodiments of the present disclosure shown in the accompanying drawings. It should be understood that these drawings depict only exemplary embodiments of the present disclosure and, therefore, should not be considered to limit the scope of the present disclosure. The principles herein will be described and explained with additional specificity and detail through the use of the accompanying drawings.
[0024] Preferred embodiments of the present disclosure will be further explained in detail below by way of examples and with reference to the accompanying drawings, in which:
[0025] Figure 1 An image processing apparatus according to an embodiment is shown.
[0026] Figure 2 A network training process according to an embodiment is shown.
[0027] Figure 3 An exemplary image segmentation output is shown.
[0028] Figure 4 An exemplary image segmentation output is shown.
[0029] Figure 5 An exemplary performance evaluation of an image processing apparatus is shown.
[0030] Figure 6 An exemplary performance evaluation of an image processing apparatus is shown.
[0031] Figure 7 A method of training an image segmentation network according to an embodiment is shown. DETAILED DESCRIPTION
[0032] Various embodiments of the present disclosure are discussed in detail below. Although specific embodiments are discussed, it should be understood that this is only for illustrative purposes. Those skilled in the relevant art will recognize that other components and configurations may be used without departing from the spirit and scope of the present disclosure.
[0033] The data-efficient training algorithm of deep neural networks achieves segmentation of medical images by effectively discriminating external data by assigning large importance weights to "good" external data and small importance weights to "bad" data ("good" and "bad" data refer to data that will improve and degrade learning performance during training, respectively).
[0034] By optimizing the weighted loss function, “good” external data contributes more to the network update and enables the network to mine more valuable knowledge, while at the same time suppressing the network from memorizing irrelevant patterns of “useless” or “harmful” data, thereby maximizing the performance gain of using external data.
[0035] The data discriminative mechanism is a constrained nonlinear programming problem, where the segmentation network is constrained to better exploit external data for learning, and the importance weights and networks are jointly learned to maximize the relevant complementary benefits.
[0036] The present disclosure teaches solving the discrimination problem by adopting a block coordinate descent optimization scheme that can alternatively perform importance weight estimation and constrain network updates. Importance weights are estimated based on the distribution difference between external and internal data, and the network is updated using hard constraints on weighted loss and transformation.
[0037] Figure 1 An image processing device according to an embodiment is shown. The image processing device includes an image segmentation network and a network training module.
[0038] The image segmentation network is configured to receive one or more image data samples. In some examples, each sample is a scan of a computed tomography (CT). In some examples, the image data may include a microscope image, an MRI scan, an X-ray image, a photograph, a video image frame, or any other image data.
[0039] The image segmentation network is configured to generate an output of segmented image data for each received sample. The image segmentation network may be an encoder-decoder network. The image segmentation network may have a UNet architecture, or SegNet, CANet, LadderNet, U-Net with attention mechanism, recursive and residual convolutional U-Net (R2-UNet), U-Net with residual blocks, or any other suitable network architecture.
[0040] The network training module is configured to train the image segmentation network. In some examples, the network training module can be configured to first train the learning network using a first corpus of training data to obtain a first optimal network.
[0041] Figure 2 A network training process according to an embodiment is shown.
[0042] The network training module is configured to receive a first corpus of training data from a trusted source. The first corpus of training data can be referred to as internal data. The first corpus includes a plurality of samples. Each sample can include pixel-level or voxel-level annotations. In some examples, the data from the trusted source can be manually annotated, and in some examples, consensus is reached among experts through multiple rounds of annotations.
[0043] The network training module is configured to receive a second corpus of training data from at least one alternative source. The second corpus of training data may be referred to as external data. The second corpus includes a plurality of samples. In some examples, the data from the alternative source may be shared on a website or by an academic institution. In some examples, the image quality and relevance and the completeness and accuracy of the annotations may be variable.
[0044] The network training module is configured to assign an initial weight to each data item in the second corpus.
[0045] The network training module can be configured to determine an importance weight for each external data sample as stated below, so that the weighted empirical expected value is the expected value In the above expression, K is the size of the training data, w k is a data pair (x k ,y k ), l represents the loss function of training, f represents the network, and D is the underlying data distribution that can accurately reflect the segmentation task. In some examples, the loss function can be based on Dice loss, average surface detection, or any other suitable metric for boundary detection and / or image segmentation.
[0046] The network training module is configured to divide the training data into one or more batches. Each batch includes at least one sample from the first corpus and at least one sample from the second corpus.
[0047] The network training module is configured to complete a predefined number of training epochs to jointly learn the importance weights and network parameters. Training can be formulated as a constrained nonlinear programming problem, letting and Denote the internal and external datasets with sample numbers M and N respectively. The problem can be expressed as:
[0048]
[0049] In the above expression, is the network f and For simplicity, the notation l i (f) can be expressed as e(f(x i ), y i ).
[0050] Each training epoch consists of the following steps performed by the network training module for each batch:
[0051] First, the network training module is configured to generate outputs for each sample in the batch using the learned network.
[0052] The network training module is then configured to determine a loss value for each sample in the batch by comparing the generated output with the true data of the sample.
[0053] Next, the network training module is configured to define the difference of each second corpus sample in the batch as the difference between the loss value of the sample and the average loss of all first corpus samples.
[0054] That is, for the external data (x n ,y n ) Data distribution for sampling and the distribution of internal datasets The network training module is configured to measure the and The difference between Defined as in, Represents l on the internal dataset m (f (t) ). In this way, when the network is exposed to external data (x n ,y n) and the internal datasets, the difference d is small, while the difference is large when the network performs differently.
[0055] The network training module is configured to approximate the average loss of all first corpus samples to the average loss of all first corpus samples from the previous training epoch modified by the change in the average loss of the first corpus samples in the batch from the previous training epoch to the current training epoch. That is, the network training module is configured to approximate the average loss of all first corpus samples from the previous training epoch to the current training epoch. Estimated in, is the average of the loss values of the internal data during the last training epoch, and M b Represents the batch of data used for estimation in the current training iteration.
[0056] In some examples, the network training module can be configured to set the average loss of all first corpus samples to 1 in the first epoch.
[0057] In some examples, the network training module can be configured to multiply the difference of each second corpus sample by the difference between the loss value of the sample in the current epoch and the loss value of the sample in the previous epoch. That is, the network training module can calculate the difference by multiplying the difference of each second corpus sample by the difference between the loss value of the sample in the current epoch and the loss value of the sample in the previous epoch. Corrected to To mitigate the randomness in the differences.
[0058] In this way, the network training module can take into account the loss trajectory of external data from the previous period to the present to correct the estimation bias. n It is difficult to split, then the large l n (f (t) ) will not produce a large d, because in this case, l n (f′)-l n (f (t) ) are usually small and vice versa.
[0059] Next, the network training module is configured to identify a set of updated weights for the next batch that minimizes the weighted sum of the differences for each second corpus sample.
[0060] The network training module can determine the importance weight of each external data by solving the following problem:
[0061]
[0062] In the above expressions, t represents the iteration step, and represents the difference. If the difference value is small, the first term aims to assign a large importance weight w to the data n .
[0063] In some examples, the identified set of updated weights is combined with a factor based on the scalar length of the vector formed by the plurality of weights to minimize the weighted sum of the differences. That is, the second term w in the above expression T w. In this way, the network training module can regularize the weights, aiming to assign similar weights to all external data to encourage the network to learn from as much external data as possible. The hyperparameter λ>0 can be used to balance the contribution of these two terms to the estimation.
[0064] Finally, the network training module is configured to identify updated network parameters for the learning network that minimize the average loss of each first corpus sample in the batch and the average weighted loss of each second corpus sample in the batch. That is, the network training module is configured to solve the following problem:
[0065]
[0066] Advantageously, a block coordinate descent optimization solution may alternatively estimate the importance weights of the external data and update the network.
[0067] The method effectively reduces the data collection burden of using deep neural networks in medical image segmentation tasks by leveraging previously little-used external data; but at the same time is not unduly influenced by this external data. Advantageously, the proposed method is computationally efficient because it uses learning trajectories to estimate importance weights. The method does not require second-order gradients of the network, which significantly reduces computational cost and speeds up training.
[0068] More specifically, the network training module can use the learning trajectory to solve this problem, thereby using the data batch M b Instead of evaluating This is computationally expensive, or simply using l n (f (t) ), which will make the randomness stronger. The network training module can use data batch M b A trajectory vector of size M is added for evaluation, which greatly reduces the computational cost.
[0069] In some examples, the network training module can be configured to identify network parameters for each batch subject to a first constraint that an average loss for each first corpus sample in the batch must be less than an average loss for each first corpus sample processed by the first best network.
[0070] The first constraint can be expressed as in, represents the best network trained on internal data only. In this way, the network training module can ensure that the image segmentation network learns better using external data than using only internal data.
[0071] In some examples, the network training module can be configured to identify network parameters for each batch subject to a second constraint that the sum of the weights is non-zero.
[0072] The second constraint can be expressed as 1 T w>0. In this way, the network training module can prevent the degenerate solution of w (eg, w=0).
[0073] In some examples, the network training module can be configured to identify network parameters by gradient descent using the sum of the weighted losses of the batches and a first constraint term representing the first constraint. That is, the network training module can be configured to update the segmentation network f by calculating (t+1) :
[0074]
[0075] In the above expression, γ (t) and b represent the learning rate and batch size, respectively. Internal data and external data can be randomly mixed in each batch, and thus for internal data In the last term, the first constraint is given by Represents, where the penalty coefficient is ξ (t) .
[0076] In some examples, the network training module can be configured to determine the first constraint term by comparing the loss value of each first corpus sample in the batch with the loss value of the sample when processed by the first optimal network. For one or more loss values of the first corpus samples in the batch that are greater than the corresponding loss value of the sample when processed by the first optimal network, the network training module can be configured to sum the differences between the compared loss values to generate the first constraint term.
[0077] That is, the network training module can be configured to Defined as When each The value of this expression is 0 when , otherwise it is a positive value. When , this expression will not affect the network update, and when there is any violation, it will force the network to mainly eliminate the violation. Therefore, this expression makes the network more Smaller l m (f (t) ), which will ensure the first constraint. To see this, note that Gives the result:
[0078] In some examples, the network training module can be configured to modify the first constraint term by a penalty coefficient. The network training module can be configured to start the penalty coefficient with a small value and update the penalty coefficient to increase with each batch.
[0079] In some examples, the network training module can be configured to update the penalty coefficient to increase the first constraint term based on the batch by a certain value.
[0080] For example, the network training module can be configured to update the penalty coefficient to in, represents the update rate. This update rule can be obtained by using sub-gradients, because In some examples, the network training module can be configured to weight the penalty coefficient to gradually decrease after a predetermined number of epochs have been completed.
[0081] For example, the network training module can be configured to factor Weighted In the e a After epochs, anneal its contribution to the network update, where e i and e is the eth i individual periods and total periods.
[0082] By annealing the constraint penalty in the network update in the late stages, the network training module can avoid violating the constraint in the late stages of training. becomes larger, which may cause the network to have to sacrifice its segmentation accuracy to eliminate these violations. In this way, the network training module can allow the network to focus on performance improvement rather than just eliminating violations.
[0083] Figure 3 An exemplary image segmentation output for multi-organ segmentation based on an input CT scan is shown. Segmentation results for seven organs are shown for three examples. These results are generated using all external data. The color information in the figure reflects the distance from the boundary pixel of the segmentation result to the nearest boundary pixel of the real data. The distance is measured by the L1 norm of the pixel coordinates that have been converted to the original image size.
[0084] Figure 4 An exemplary image segmentation output of cervical cell segmentation based on an input microscope image is shown. The results of segmenting the cytoplasm and nucleus are shown for seven examples. These results were generated using all external data. The results of the algorithm output are shown on the right, while the real data is shown on the left.
[0085] The method of the embodiment of the proposed invention is compared with four methods for the multi-organ segmentation task. These methods include a method based on data selection (DS), two methods based on data weighting (RW and DD), and preliminary versions of these methods (SL). DS selects k data with the minimum loss value from the batch to update the network at each iteration; k is set to 5 here. RW assigns data weights by minimizing the weighted loss on the validation set; for this, 50% of the given internal training data are used. Note that both DD and our method use distribution differences to assign importance weights for external data. However, DD assumes that all external data in the same dataset are sampled from the same distribution, and therefore assigns the same importance weight to all external data in the same dataset, while our method finds a suitable importance weight for each external data because the external data in the same dataset can be sampled from different distributions. Our method is implemented by using the same learning environment as ours to reproduce its results.
[0086] Figure 5 An exemplary performance evaluation of an image processing apparatus for a multi-organ segmentation task is shown. Figure 3 Segmentation results of the 7 organs shown. Figure 5 (a) The calculated Deyss loss coefficient (DSC) relative to real data is shown in percent. Figure 5 (b) shows the average surface distance (ASD) metric in pixels relative to the true data. Both sets of results are generated for the five amounts of external data used: 10%, 25%, 50%, 75%, and 100%. UNet is used as the backbone network and trained using the Diess loss to generate the loss values. It can be seen that our algorithm produces more accurate results in all cases and is the only method that produces better results when more external data is given. These findings indicate that our algorithm is more effective in discriminating external data than the comparison methods and show that its discriminative ability is less affected by the amount of external data.
[0087] The methods for cervical cell segmentation were also compared. The same experimental environment was used to reproduce the results. The methods were compared for segmenting the cytoplasm and nucleus using 5 external data amounts: 10%, 25%, 50%, 75%, and 100%.
[0088] Figure 6 An exemplary performance evaluation of an image processing apparatus for the task of cervical cell segmentation is shown. Figure 6 (a) The calculated Deyss loss coefficient (DSC) relative to real data is shown in percent. Figure 6(b) shows the average surface distance (ASD) metric in pixels relative to the real data. Both sets of results are generated for the 5 external data amounts used: 10%, 25%, 50%, 75%, and 100%. These results are generated by UNet with Dice loss. It can be seen that our algorithm produces more accurate results than the comparison methods in all scenarios, confirming that our algorithm can work well in different types of medical image segmentation tasks and can effectively distinguish different types of medical images.
[0089] Figure 7 A method for training an image segmentation network according to an embodiment is shown. The method starts at step S01.
[0090] At step S02 , a first corpus of training data is received from a trusted source.
[0091] At step S03 , a second corpus of training data is received from at least one alternative source.
[0092] At step S04, an initial weight is assigned to each data item in the second corpus.
[0093] At step S05 , the training data is divided into one or more batches, wherein each batch includes at least one sample from the first corpus and at least one sample from the second corpus.
[0094] At step S06 , starting from the first batch, the learning network is used to generate outputs for each sample in the batch.
[0095] At step S07, the loss value of each sample in the batch is determined by comparing the generated output with the true data of the sample.
[0096] At step S08, the difference of each second corpus sample in the batch is defined as the difference between the loss value of the sample and the average loss of all first corpus samples, where the average loss of all first corpus samples is approximated as the average loss of all first corpus samples from the previous training epoch modified by the change in the average loss of the first corpus samples in the batch from the previous training epoch to the current training epoch.
[0097] At step S09 , a set of updated weights is identified for the next batch, the set of updated weights minimizing the weighted sum of the differences for each second corpus sample.
[0098] At step S10, updated network parameters are identified for the learning network, the updated network parameters minimizing the average loss of each first corpus sample in the batch and the average weighted loss of each second corpus sample in the batch.
[0099] At step S11, it is determined whether the batch is the final batch of one or more batches generated at step S05. If so, the method proceeds to step S12. Otherwise, the method returns to step S06 with the next batch.
[0100] At step S12, it is determined whether the batch is the final period of a predetermined number of periods. If so, the method proceeds to step S13. Otherwise, the method returns to step S06 with the first batch of the next period.
[0101] The method ends in step S13.
[0102] The above embodiments are described by way of example only. Many variations are possible without departing from the scope of the present disclosure as defined in the appended claims.
[0103] For clarity of explanation, in some instances, the technology may be presented as including various functional blocks including functional blocks including devices, device components, steps or routines in methods implemented in software, or a combination of hardware and software.
[0104] The method according to the above example can be implemented using computer executable instructions, which are stored in a computer readable medium or can be obtained from a computer readable medium in other ways. These instructions may include, for example, instructions and data, which enable or otherwise configure a general-purpose computer, a special-purpose computer or a special-purpose processing device to perform a specific function or function group. The part of the computer resources used can be accessed through a network. Computer executable instructions can be, for example, binary files, intermediate format instructions, such as assembly language, firmware or source code. The example of a computer readable medium that can be used to store instructions, information used and / or information created in the method process according to the described example includes a disk or optical disk, a flash memory, a universal serial bus (USB) device provided with a non-volatile memory, a network storage device, etc.
[0105] The device implementing the method according to these disclosures may include hardware, firmware and / or software, and may take any of a variety of form factors. Typical examples of these form factors include laptop computers, smart phones, small personal computers, personal digital assistants, etc. The functions described herein may also be implemented in peripheral devices or add-on cards. As a further example, such functions may also be implemented on circuit boards of different chips, or may be implemented on different processes executed in a single device.
[0106] Instructions, media for transmitting such instructions, computing resources for executing such instructions, and other structures for supporting such computing resources are means for providing the functionality described in these disclosures.
[0107] Although a variety of examples and other information are used to explain aspects within the scope of the appended claims, limitations to the claims should not be implied based on specific features or arrangements in such examples, because a person of ordinary skill in the art will be able to derive a variety of implementations using these examples. Further, and although a subject matter may have been described in language specific to examples of structural features and / or method steps, it should be understood that the subject matter defined in the appended claims is not necessarily limited to these described features or actions. For example, such functionality may be distributed differently or performed in components other than those identified herein. Instead, the described features and steps are disclosed as examples of components of systems and methods within the scope of the appended claims.
Claims
1. A computer-implemented method for training a learning network to perform image segmentation, the method comprising: receiving a first corpus of training data from a trusted source; receiving a second corpus of training data from at least one alternative source; assigning an initial weight to each data item in the second corpus; Dividing the training data into one or more batches, wherein each batch includes at least one sample from the first corpus and at least one sample from the second corpus; and Complete a predefined number of training epochs, where, for each batch, each training epoch consists of: Using the learned network to generate an output for each sample in the batch; determining a loss value for each sample in the batch by comparing the generated output with true data for the sample; defining the variance of each second corpus sample in the batch as the difference between the loss value of the sample and the average loss of all first corpus samples, wherein the average loss of all first corpus samples is approximated as the average loss of all first corpus samples from a previous training epoch modified by the change in the average loss of the first corpus samples in the batch from the previous training epoch to the current training epoch; identifying an updated set of weights for the next batch that minimizes a weighted sum of the differences for each second corpus sample; Updated network parameters are identified for the learning network, the updated network parameters minimizing an average loss for each first corpus sample in the batch and an average weighted loss for each second corpus sample in the batch.
2. The computer-implemented method of claim 1 , further comprising training the learning network using a first corpus of the training data to obtain a first optimal network, wherein: For each batch, network parameters are identified subject to a first constraint that an average loss for each first corpus sample in the batch must be less than an average loss for each first corpus sample processed by the first optimal network.
3. The computer-implemented method of claim 2, wherein: The network parameters are identified by gradient descent using the sum of the weighted losses of the batches and a first constraint term representing the first constraint.
4. The computer-implemented method of claim 3, wherein: The first constraint is determined by: comparing the loss value of each first corpus sample in the batch with the loss value of the sample when processed by the first best network; as well as For one or more loss values of first corpus samples in the batch that are greater than corresponding loss values of the samples when processed by the first optimal network, differences between the compared loss values are summed to generate the first constraint term.
5. The computer-implemented method of claim 4, wherein: The first constraint term is modified by a penalty coefficient that starts out as a small value and is updated to increase with each batch.
6. The computer-implemented method of claim 5, wherein: The penalty coefficient is updated to increase by a certain value based on the first constraint item of the batch.
7. The computer-implemented method of claim 5, wherein: The penalty coefficient is weighted to gradually decrease after a predetermined number of epochs have been completed.
8. The computer-implemented method of claim 1, wherein: For each batch, network parameters are identified subject to a second constraint that the sum of the weights is non-zero.
9. The computer-implemented method of claim 1, wherein: The identified set of updated weights is combined with a factor based on the scalar length of the vector formed by the plurality of weights to minimize a weighted sum of the differences.
10. The computer-implemented method of claim 1, wherein: The difference of each second corpus sample is further multiplied by the difference between the loss value of the sample in the current epoch and the loss value of the sample in the previous epoch.
11. The computer-implemented method of claim 1 , wherein: Each sample is a computed tomography (CT) scan.
12. The computer-implemented method of claim 1, wherein: In the first epoch, the average loss of all first corpus samples is set to 1.
13. The computer-implemented method of claim 1 , further comprising performing image segmentation by: receiving one or more samples of image data; and An output of segmented image data is generated for each received sample.
14. A computer readable medium comprising instructions which, when executed by a processor, cause the processor to perform the method of claim 1.
15. An image segmentation network, wherein the image segmentation network is trained according to the method of claim 1.
