Generalized base model framework for medical imaging

Abstract

The present disclosure relates to a general-purpose base model framework for medical imaging. Systems and methods are provided for performing a medical imaging analysis task using a general-purpose base model. One or more input medical images are provided, each image being in a domain, and a domain code identifying the domain for each of the one or more input medical images is provided. For each respective one of the domain codes, one or more weights are determined based on the respective domain code. One or more parameters of a dynamic convolutional layer are updated based on the one or more weights. A first set of features is extracted from the one or more input medical images using the dynamic convolutional layer with the one or more updated parameters. The first set of features is encoded into a second set of features using a machine learning-based encoder. The medical imaging analysis task is performed based on the second set of features. A result of the medical imaging analysis task is output.

Description

Technical Field

[0001] This invention relates generally to AI / ML (artificial intelligence / machine learning) based medical imaging analysis, and particularly to a general basic model framework for medical imaging. Background Technology

[0002] In recent years, advancements in foundational models have garnered significant attention, with notable progress observed across various applications. For instance, progress in LLM (Large Language Models) has significantly improved performance in question-answering tasks, while diffusion models have demonstrated excellent performance in image synthesis. However, the inherent complexity and heterogeneity of medical imaging data present substantial challenges to the design of general-purpose foundational models.

[0003] Conventional base models typically accept a single image as input. However, medical diagnosis often requires multiple input medical images, such as multiple MR (magnetic resonance) images with different contrasts. Furthermore, the number and corresponding contrasts of input images acquired across different diagnostic tasks and clinical settings can vary. Conventional base models are generally unable to handle varying numbers of input images. Summary of the Invention

[0004] According to one or more embodiments, a system and method are provided for performing a medical imaging analysis task using a general base model. The system provides 1) one or more input medical images, each image in a domain, and 2) a domain code identifying the domain for each of the one or more input medical images. For each corresponding domain code, one or more weights are determined based on the corresponding domain code. Based on the one or more weights, one or more parameters of a dynamically convolutional layer are updated. A first set of features is extracted from the one or more input medical images using the dynamically convolutional layer with one or more updated parameters. The first set of features is encoded into a second set of features using a machine learning-based encoder. Based on the second set of features, a medical imaging analysis task is performed. The result of the medical imaging analysis task is output.

[0005] In one embodiment, the dynamic convolutional layer and the machine learning-based encoder are trained by receiving 1) one or more training medical images, each in a domain, and 2) training domain codes identifying the domain for each of the one or more training medical images. Patches of the one or more training medical images are masked. For each corresponding training domain code, one or more additional weights are determined based on the corresponding training domain code. One or more additional parameters of the dynamic convolutional layer are updated based on the one or more additional weights. A third set of features is extracted from the unmasked patches of the one or more training medical images using the dynamic convolutional layer with the one or more updated additional parameters. The third set of features is encoded into a fourth set of features using the machine learning-based encoder. Tokens representing the masked patches of the one or more training medical images are inserted into the fourth set of features. The one or more training medical images are reconstructed based on the fourth set of features with the inserted tokens using the machine learning-based decoder. The dynamic convolutional layer and the machine learning-based encoder are trained based on the one or more training medical images and the one or more reconstructed training medical images. The trained dynamic convolutional layer and the trained machine learning-based encoder are output.

[0006] In one embodiment, a linear projector is used to project the corresponding domain code onto one or more weights.

[0007] In one embodiment, the weight and bias parameters of the dynamic convolutional layer are updated. This is achieved by determining the dot product of one or more parameters of the dynamic convolutional layer with the corresponding weight.

[0008] In one embodiment, a first set of features is encoded using positional embeddings. In another embodiment, modal embeddings are generated based on domain codes, and these modal embeddings are used to encode the first set of features.

[0009] In one embodiment, the medical imaging analysis task includes medical image synthesis.

[0010] According to one or more embodiments, a system and method are provided for training a general base model to perform medical imaging analysis tasks. The system receives 1) one or more training medical images, each in a domain, and 2) domain codes identifying the domain for each of the one or more training medical images. Patches of the one or more training medical images are masked. For each corresponding domain code, one or more weights are determined based on the corresponding domain code. One or more parameters of a dynamic convolutional layer are updated based on one or more additional weights. Using the dynamic convolutional layer with one or more updated parameters, a first set of features is extracted from unmasked patches of the one or more training medical images. The first set of features is encoded into a second set of features using a machine learning-based encoder. Labels representing the masked patches of the one or more training medical images are inserted into the second set of features. The one or more training medical images are reconstructed based on the second set of features with the inserted labels using a machine learning-based decoder. The dynamic convolutional layer and the machine learning-based encoder are trained based on the one or more training medical images and the one or more reconstructed training medical images. The trained dynamic convolutional layer and the trained machine learning-based encoder are output.

[0011] These and other advantages of the invention will be apparent to those skilled in the art from the following detailed description and accompanying drawings. Attached Figure Description

[0012] Figure 1 A method for performing a medical imaging analysis task using a general base model, according to one or more embodiments, is illustrated; Figure 2 A workflow for performing medical imaging analysis tasks using a common base model, according to one or more embodiments, is illustrated. Figure 3 A method for training a general base model for performing medical imaging analysis tasks, according to one or more embodiments, is illustrated; Figure 4 The workflow for training a general base model for performing medical imaging analysis tasks according to one or more embodiments is illustrated; Figure 5 A table comparing the performance of AMAE according to the embodiments described herein with conventional methods is shown; Figure 6 A table comparing AMAE according to the embodiments described herein with conventional methods is shown, wherein all subjects lacked AxT2; Figure 7 A table is shown that statistically compares AMAE with conventional methods according to the embodiments described herein with and without AMAE pre-trained weights. Figure 8 An exemplary artificial neural network that can be used to implement one or more embodiments is shown; Figure 9 This illustrates a convolutional neural network that can be used to implement one or more embodiments; Figure 10 A data flow diagram is shown that can be used to implement one or more embodiments of a generative adversarial network; Figure 11 A schematic structure is shown that can be used to implement a recurrent machine learning model of one or more embodiments; Figure 12 A high-level block diagram of a computer that can be used to implement one or more embodiments is shown. Detailed Implementation

[0013] This invention generally relates to methods and systems for a general foundational model framework for medical imaging. Embodiments of the invention are described herein to provide an intuitive understanding of these methods and systems. Digital images typically consist of digital representations of one or more objects (or shapes). The digital representations of objects are generally described herein in relation to identifying and manipulating them. Such manipulation is a virtual manipulation performed in the memory or other circuitry / hardware of a computer system. Therefore, it should be understood that embodiments of the invention can be executed within a computer system using data stored within the computer system. Furthermore, image pixels mentioned herein can also be equivalently referred to as image voxels, and vice versa.

[0014] The embodiments described herein provide a novel, general-purpose base model framework for efficiently processing varying numbers of input medical images. This framework eliminates the need to train multiple models or encoders to adapt to different numbers of input images. The base model according to the embodiments described herein can utilize all available input data with minimal preprocessing, and the pre-trained base model can be used to handle a variety of downstream tasks with diverse requirements and varying numbers of input medical images.

[0015] Figure 1 A method 100 for performing a medical imaging analysis task using a general base model, according to one or more embodiments, is illustrated. The steps and sub-steps of method 100 can be performed by one or more suitable computing devices, such as… Figure 12 Computer 1202. Figure 2 A workflow 200 for performing medical imaging analysis tasks using a common base model, according to one or more embodiments, is shown. Figure 1 and Figure 2 They will be described together.

[0016] exist Figure 1In step 102, 1) one or more input medical images are received, each image in a domain, and 2) a domain code for each of the one or more input medical images, identifying its domain. Figure 2 As shown in workflow 200, one or more input medical images can be input medical images 202-A, 202-B, ..., 202-N (collectively referred to as input medical images 202) in domains 1, 2, ..., N, and the domain codes can be domain codes 204-A, 204-B, ..., 204-N (collectively referred to as domain codes 204) that identify the domains of input medical images 202-A, 202-B, ..., 202-N, respectively. Specifically, domain code 204-A is [0,0,0,1] identifying domain 1, domain code 204-B is [0,0,1,0] identifying domain 2, and domain code 204-N is [1,1,1,1] identifying domain N.

[0017] One or more input medical images can depict any anatomical object of interest to the patient and can belong to any suitable domain. As used herein, the domain of a medical image refers to the modality of the medical image and the protocol used to acquire that modality. The modality of one or more input medical images can include, for example, MRI (Magnetic Resonance Imaging), CT (Computed Tomography), US (Ultrasound), X-ray, SPECT (Single Photon Emission Computed Tomography), PET (Positron Emission Tomography), or any other medical imaging modality or combination of medical imaging modalities. The protocol used to acquire the medical images can include, for example, acquisition sequences or techniques for acquiring medical images, such as, for example, T1-weighted, T2-weighted, proton density-weighted MRI images, contrast and non-contrast images, CT images captured at low kV and high kV, or low-resolution and high-resolution medical images. Thus, these domains can be completely different medical imaging modalities or different imaging protocols within the same overall imaging modality. One or more input medical images can be represented in image space (e.g., as pixel or voxel values ​​in spatial coordinates) or latent space (e.g., as a low-dimensional compressed representation of one or more medical images in the form of feature vectors). One or more input medical images in the image space can be 2D (two-dimensional) images and / or 3D (three-dimensional) volumes.

[0018] Domain codes identify one or more domains of an input medical image. Domain codes can be represented in any suitable form. In one embodiment, domain codes are vectors. Each combination of domain code values ​​is associated with a predefined domain. The association between domain codes and predefined domains can be predefined by the user or learned during training. Other methods for encoding domain identifiers into domain codes are also envisioned.

[0019] For example, one or more input medical images and / or domain codes can be received from an image acquisition device (e.g., Figure 12 The image acquisition device 1214 directly receives one or more input medical images as one or more input medical images for acquisition; from the storage device or memory of the computer system (e.g., Figure 12 Computer 1202 loads one or more input medical images and / or field codes from storage device 1212 or memory 1210; or from a remote computer system (e.g., Figure 12 The computer (1202) receives one or more input medical images and / or domain codes. Such a computer system or remote computer system may include one or more patient databases, such as, for example, EHR (Electronic Health Record), EMR (Electronic Medical Record), PHR (Personal Health Record), HIS (Health Information System), RIS (Radiology Information System), PACS (Picture Archiving and Communication System), LIMS (Laboratory Information Management System), or any other suitable database or system.

[0020] exist Figure 1 In step 104, for each corresponding domain code, one or more weights are determined based on the corresponding domain code. These one or more weights are used to update one or more parameters of the dynamic convolutional layer (in...). Figure 1 (Updated in step 106). In one embodiment, one or more weights include a weight vector. and bias vector These are used to update the weight parameters of the dynamic convolutional layer. and bias parameters .

[0021] One or more weights can be determined using any suitable method. In one embodiment, a learnable linear projector is used to determine one or more weights. In one example, such as... Figure 2 As shown in workflow 200, each linear projector 206-A, 206-B, ..., 206-N (collectively referred to as linear projectors 206) determines one or more weights based on domain codes 204-A, 204-B, ..., 204-N. A linear projector is a linear transformation learned during the training process. A linear projector projects or maps a domain code to one or more weights using a set of parameters optimized during training. Although linear projectors 206-A, 206-B, ..., 206-N are shown individually in workflow 200 to illustrate the linear projection of domain codes 204-A, 204-B, ..., 204-N to one or more weights, it should be understood that linear projectors 206-A, 206-B, ..., 206-N share the same weights and therefore represent the same linear projector 206.

[0022] exist Figure 1 In step 106, one or more parameters of the dynamic convolutional layer are updated based on one or more weights. In one example, such as Figure 2 As shown in workflow 200, one or more parameters of the dynamic convolutional layer / tokenizer 208 are updated based on one or more weights output by the linear projector 206.

[0023] Dynamic convolutional layers include convolutional filters (or kernels) that are not fixed but dynamically adapt by updating one or more parameters based on domain codes. Examples of dynamic convolutional layers include dynamic filter networks and attention mechanisms. In one embodiment, one or more parameters of a dynamic convolutional layer include weight parameters. and bias parameters In one embodiment, one or more parameters are updated by calculating a dot product. For example, weight parameters can be updated by calculating weight parameters. With weight vector The dot product (i.e., The bias parameter can be updated by calculating the bias parameter. With bias vector The dot product (i.e., The parameters are updated using [the appropriate method]. Therefore, the domain code controls the behavior of the dynamic convolutional layer for each domain to extract the corresponding features. One or more parameters can be updated according to any other suitable method.

[0024] exist Figure 1 In step 108, a first set of features is extracted from one or more input medical images using a dynamically generated convolutional layer with one or more updated parameters. In one example, such as... Figure 2 As shown in workflow 200, a first set of features 210 is extracted from medical image 202 using dynamic convolutional layer / labeler 208. The dynamic convolutional layer receives patches extracted from one or more input medical images as input, maps each patch to a label of a feature, and generates labels representing the first set of features as output. These features are low-dimensional compressed representations of one or more input medical images as feature vectors.

[0025] Dynamic convolutional layers allow for the extraction of domain-specific features from each of one or more input medical images, while also enabling the machine learning encoder (in...) to use the tile size as the kernel and stride of the convolution and the output dimension of the convolution as the embedding dimension of the machine learning encoder. Figure 1 In step 110, non-overlapping markers are created. Let... Represents one of one or more input medical images (or image tensors), where , and These are the height, width, and depth of the input medical image, respectively. The convolution operation with updated weight and bias parameters is as follows: in, It is an output label generated by a dynamic convolutional layer for an input medical image.

[0026] exist Figure 1 In step 110, a machine learning-based encoder is used to encode the first set of features into a second set of features. In one example, such as... Figure 2 As shown in workflow 200, the first set of features 210 is encoded into the second set of features 214 by transformer encoder 212.

[0027] The labels of the first set of features are combined (e.g., concatenated) into a single sequence and input into a machine learning-based encoder. The labels of the first set of features are encoded using positional embeddings (e.g., sinusoidal embeddings or learnable embeddings) based on the location of their tiles within one or more input medical images. Therefore, labels from different domains but with the same relative tile location will have the same positional embedding. Positional embeddings help preserve the spatial relationships of the labels. Furthermore, modal embeddings are generated from the domain codes using another linear projector, and the labels of the first set of features are encoded using these modal embeddings. The machine learning-based encoder receives the first set of features (represented as a sequence of labels encoded using positional and modal embeddings) as input and generates a second set of features as output. In one embodiment, the machine learning-based encoder is a transformer encoder. However, the machine learning-based encoder can be implemented in any other suitable form.

[0028] exist Figure 1 In step 112, a medical imaging analysis task is performed based on the second set of features. The medical imaging analysis task can be performed using a machine learning-based task network, such as, for example, a task-specific decoder head (e.g., a transformer decoder). In one example, such as... Figure 2 As shown in workflow 200, the transformer decoder 216 generates a synthetic medical image 218 based on a second set of features 214. A machine learning-based task network receives the second set of features as input and generates the result of a medical imaging analysis task as output. In one embodiment, the medical imaging analysis task is image synthesis for generating a synthetic image from one or more medical images. The synthetic image may reside in a domain outside the domain of the one or more input medical images. The medical imaging analysis task may alternatively or additionally include any other suitable task, such as, for example, detection, segmentation, classification, quantization, etc.

[0029] exist Figure 1 In step 114, the results of the medical imaging analysis task are output. For example, the results of the medical imaging analysis task can be output by displaying the results on a display device of the computer system (e.g., ...). Figure 12 The computer 1202's I / O 1208); stores the results in the computer system's memory or storage device (e.g., Figure 12 The computer 1202's memory 1210 or storage device 1212; or the results are transmitted to a remote computer system (e.g., Figure 12 Computer 1202).

[0030] The dynamic convolutional layers and machine learning-based encoders are trained during previous offline or training phases, for example, according to Figure 3 Method 300 and / or Figure 4 The workflow is 400. Once trained, dynamic convolutional layers and machine learning-based encoders are applied in the online or inference phase, for example, to perform... Figure 1 Steps 108 and 110.

[0031] Figure 3 A method 300 for training a general base model for performing medical imaging analysis tasks, according to one or more embodiments, is illustrated. The steps and sub-steps of method 300 can be performed by one or more suitable computing devices, such as… Figure 12 Computer 1202. Figure 4 A workflow 400 for training a general base model for performing medical imaging analysis tasks according to one or more embodiments is shown. Figure 3 and Figure 4 They will be described together. Figure 3 Method 300 and Figure 4 The workflow consists of 400 steps performed offline or during the training phase to train dynamic convolutional layers and machine learning-based encoders.

[0032] exist Figure 3 In step 302, 1) one or more training medical images, each image in a domain, and 2) a domain code for each of the one or more training medical images, identifying its domain. Figure 4As shown in workflow 400, one or more training medical images can be training medical images 402-A, 402-B, ..., 402-N (collectively referred to as training medical images 402) in domains 1, 2, ..., N, and the domain codes can be domain codes 404-A, 404-B, ..., 404-N (collectively referred to as domain codes 404) that identify the domains of training medical images 402-A, 402-B, ..., 402-N, respectively. Specifically, domain code 404-A is [0,0,0,1] identifying domain 1, domain code 404-B is [0,0,1,0] identifying domain 2, and domain code 404-N is [1,1,1,1] identifying domain N.

[0033] One or more training medical images can belong to any suitable domain. One or more training medical images can be represented in an image space or a latent space. One or more training medical images in the image space can be 2D images and / or 3D volumes.

[0034] Domain codes identify one or more domains of training medical images. Domain codes can be represented in any suitable form (e.g., vectors). Each combination of domain code values ​​is associated with a predefined domain (e.g., predefined by the user or learned during training). Other methods for encoding domain identifiers into domain codes are also envisioned.

[0035] For example, one or more training medical images and / or domain codes can be received from an image acquisition device (e.g., Figure 12 The image acquisition device 1214 directly receives one or more training medical images as acquired training medical images; from the storage device or memory of the computer system (e.g., Figure 12 Computer 1202 loads one or more training medical images and / or domain codes from storage device 1212 or memory 1210; or from a remote computer system (e.g., Figure 12 The computer (1202) receives one or more training medical images and / or domain codes.

[0036] exist Figure 3 In step 304, one or more training medical images are masked. In one example, such as... Figure 4 As shown in workflow 400, patches of training medical images 402-A, 402-B, ..., 402-N are masked to provide masked training medical images 406-A, 406-B, ..., 406-N. Patches of one or more training medical images may be masked according to any suitable method (e.g., random).

[0037] exist Figure 3In step 306, for each corresponding domain code, one or more weights are determined based on the corresponding domain code. These one or more weights are used to update one or more parameters of the dynamically convolutional layer (in...). Figure 3 (Updated in step 308). In one embodiment, one or more weights include a weight vector. and bias vector Used to update the weight parameters of dynamic convolutional layers and bias parameters .

[0038] One or more weights can be determined using any suitable method. In one embodiment, a learnable linear projector is used to determine one or more weights. In one example, such as... Figure 4 As shown in workflow 400, each linear projector 406-A, 406-B, ..., 406-N (collectively referred to as linear projector 406) determines one or more weights based on domain codes 404-A, 404-B, ..., 404-N, respectively. Although linear projectors 406-A, 406-B, ..., 406-N are shown individually in workflow 400 to illustrate the linear projection of domain codes 404-A, 404-B, ..., 404-N to one or more weights, it should be understood that linear projectors 406-A, 406-B, ..., 406-N share the same weights and therefore represent the same linear projector 406.

[0039] exist Figure 3 In step 308, one or more parameters of the dynamic convolutional layer are updated based on one or more weights. In one example, such as Figure 4 As shown in workflow 400, one or more parameters of the dynamic convolutional layer / labeler 410 are updated based on one or more weights output by the linear projector 406. This dynamic convolutional layer can be... Figure 1 The dynamic convolutional layer used in step 108 can be... Figure 2 208 dynamic convolutional layers / labelers.

[0040] In one embodiment, one or more parameters of the dynamic convolutional layer include weight parameters. and bias parameters In one embodiment, one or more parameters are updated by calculating a dot product. For example, weight parameters can be updated by calculating weight parameters. With weight vector The dot product (i.e., The bias parameter can be updated by calculating the bias parameter. With bias vector The dot product (i.e., This can be used to update. One or more parameters can also be updated using any other suitable method.

[0041] exist Figure 3 In step 310, a first set of features is extracted from unmasked patches of one or more training medical images using a dynamically generated convolutional layer with one or more updated parameters. In one example, such as... Figure 4 As shown in workflow 400, a first set of features 412 is extracted from unmasked patches of training medical images 402 using a dynamic convolutional layer / labeler 410. The dynamic convolutional layer receives unmasked patches extracted from one or more training medical images as input, maps each patch to a label of a feature, and generates labels representing the first set of features as output. These features are low-dimensional compressed representations of one or more input medical images as feature vectors.

[0042] exist Figure 3 In step 312, a machine learning-based encoder is used to encode the first set of features into a second set of features. In one example, such as... Figure 4 As shown in workflow 400, the first set of features 412 is encoded into the second set of features 416 by the transformer encoder 414. The machine learning-based encoder can be... Figure 1 The machine learning-based encoder used in step 108 may be... Figure 2 The converter encoder 212.

[0043] The labels of the first set of features are combined (e.g., concatenated) into a single sequence and input into a machine learning-based encoder. The labels of the first set of features are encoded using positional embeddings based on the location of their tiles within one or more input medical images. Furthermore, modal embeddings are generated from the domain code using another linear projector, and these modal embeddings are used to encode the labels of the first set of features. The machine learning-based encoder receives the first set of features (represented as a sequence of labels encoded using both positional and modal embeddings) as input and generates a second set of features as output. In one embodiment, the machine learning-based encoder is a transformer encoder. However, the machine learning-based encoder can be implemented in any other suitable form.

[0044] exist Figure 3 In step 314, labels representing one or more masked patches of training medical images are inserted into the second set of features. In one example, such as... Figure 4 As shown in workflow 400, markers representing masked patches are inserted into a second set of features 416 to provide feature 418. The markers representing masked patches are placeholder markers that are inserted at the position of each masked patch in the second set of features.

[0045] exist Figure 3 In step 316, a machine learning-based decoder is used to reconstruct one or more training medical images based on a second set of features with insertion tags. In one example, such as... Figure 4 As shown in the workflow 400, the transformer decoder 420 reconstructs the training medical images 402-A, 402-B, ..., 402-N based on features 418 to generate reconstructed images 422-A, 422-B, ..., 422-N.

[0046] In one embodiment, the machine learning-based decoder is a transformer decoder. However, the machine learning-based decoder can be implemented according to any other suitable architecture. The machine learning-based decoder receives a second set of features with insertion tags as input and generates a reconstructed image as output. In one embodiment, decoding can be performed by self-attention of the transformer decoder. In this embodiment, self-attention is used to decode the second set of features with insertion tags, and then a linear projector is used to map the decoded tags to the original tile size. In another embodiment, decoding can be performed by cross-attention. In this embodiment, a query is generated using the second set of features with insertion tags from a specific domain, cross-attention is performed using keys and values ​​generated from all encoded tags from all domains, and a linear projector is used to map the decoded tags to the original tile size. Both self-attention and cross-attention utilize positional embeddings and modal embeddings.

[0047] exist Figure 3 In step 318, a dynamic convolutional layer and a machine learning-based encoder are trained based on one or more training medical images and one or more reconstructed training medical images. In one embodiment, the dynamic convolutional layer and the machine learning-based encoder are trained by comparing one or more training medical images and one or more reconstructed training medical images, according to a loss function. In one embodiment, the loss function is an L2 loss function. However, the loss function may include any other suitable loss function.

[0048] exist Figure 3 In step 320, the trained dynamic convolutional layer and the trained machine learning-based encoder are output. For example, the trained dynamic convolutional layer and the trained machine learning-based encoder can be output by storing them in the memory or storage device of a computer system (e.g., ...). Figure 12 The memory 1210 or storage device 1212 of the computer 1202; or the trained dynamic convolutional layer and the trained machine learning-based encoder are transferred to a remote computer system (e.g., Figure 12(Computer 1202). Trained dynamic convolutional layers and trained machine learning-based encoders can be applied during online or inference phases, for example, to perform... Figure 1 Method 100 or Figure 2 Workflow 200.

[0049] In one embodiment, the converter encoder and converter decoder used in this paper can be replaced by a state-space model to reduce computational costs.

[0050] The embodiments described herein have been experimentally validated. For pre-training, 45,374 MRI studies were used. For each case, a different number of acquired MRI contrasts were used, including a set of seven contrasts: apparent diffusion coefficient (ADC), trace weighting, T2 weighting, gradient echo (GRE), sensitivity weighting, T1 weighting, and fluid attenuation inversion recovery (FLAIR), or a subset thereof when available. For fine-tuning, a subset of 1,648 training subjects, 193 validation subjects, and 215 test subjects were used, with acute / subacute ischemic regions manually segmented.

[0051] Acute and subacute infarct lesions were manually segmented on axial trace-weighted (AxTrace) contrast image sequences by radiologists (TJRs, 10 years of experience). AxTrace image sequences and corresponding ADC images were reviewed by radiologists. Regions within the brain parenchyma exhibiting high signal intensity in TraceW image sequences but low or isointense signal intensity in ADC images were identified by radiologists as recent (acute to subacute) infarct-positive and delineated in image masks. Inclusion criteria for the study were: at least AxTrace and AxADC contrast were presented, and at least one infarct lesion was identified by the radiologist. Additionally, axial T2-weighted contrast was also used if available.

[0052] Two approaches were used for fine-tuning the infarct segmentation task. The first approach, called Adaptive UNETR (UNet Transformer), preserves the DCT and transformer encoder while removing the transformer decoder. A UNETR-based segmentation decoder is then used. Since the number of inputs varies for each subject, the number of labels also varies. At each layer of the UNETR decoder head, adaptive max-pooling layers are used to extract the necessary information across all domains, thus transforming the unequal number of labels into tensors of fixed size to generate the segmentation mask.

[0053] make Indicates the output of the converter encoder: in It refers to the batch size. It is the number of 3D input images. It is the number of tiles / markers, and It is the transformer dimension.

[0054] Information was extracted from all tags using an adaptive max-pooling layer, and the results are as follows: .

[0055] In another approach, the original UNETR model is used as is. All images from different modalities are concatenated along the channel dimension. If any input modality is missing, a zero-padding tensor is used. The weights of the pre-trained ViT encoder are passed from the pre-trained model to the UNETR encoder.

[0056] The initial learning rate for self-pre-training was 1e-5 with weight decay of 0.05. The L2 loss function and weighted Adam optimizer were used. For pre-training, the variable-mask autoencoder model was run for 500 epochs with a masking rate of 70% and a batch size of 4.

[0057] For fine-tuning the infarction segmentation task, the fine-tuning architecture was run for 200 epochs, and the best-performing validation model was saved. The performance of different models was then compared using a test dataset. The performance of the model according to the embodiment described in this paper was compared with UNet and UNETR. The fine-tuning was performed with a learning rate of 0.0001, a learning rate scheduler, and a weight decay of 0.05. Diceloss and a weighted ADAM optimizer were used for training. Additionally, a test was performed where model performance on the test data was evaluated by omitting the optional contrastive AxT2 to observe how the model performs when contrastives / modalities are missing.

[0058] The Dice similarity coefficient was used to quantitatively evaluate the test data. A two-sided pairwise Wilcoxon signed-rank test was used to compare the Dice scores of two models, one pre-trained and one unpre-trained.

[0059] Figure 5 Table 500 presents the effectiveness of the proposed AMAE (Adaptive Mask Autoencoder) pre-trained using adaptive UNETR and the original UNETR, according to the embodiments described herein. The adaptive UNETR with AMAE pre-trained weights outperforms the adaptive UNETR without AMAE pre-trained weights by 3.7%. Furthermore, in the case of the original UNETR, using weights pre-trained from AMAE improves performance by 2.8%.

[0060] Figure 6The results in Table 600 show the performance of different models using the average Dice score on test subjects with optional contrast AxT2 missing. The AMAE pre-trained weights using the original UNETR decoder achieved an average Dice score of 0.575, only 2.9% lower than the performance using all contrasts.

[0061] Figure 7 Table 700 shows the statistical analysis of Dice scores for two fine-tuning methods (Adaptive UNETR and UNETR) with and without pre-training. In both cases, the p-values ​​are less than 0.05, indicating a significant difference in segmentation results when applying pre-trained weights compared to the corresponding model without pre-trained weights.

[0062] The embodiments described herein relate to the claimed system and the claimed method. Features, advantages, or alternative embodiments described herein can be assigned to other claimed objects, and vice versa. In other words, the system claims and embodiments can be modified using features described or claimed in the context of the corresponding method. In this case, the functional features of the method are implemented by the physical units of the system.

[0063] Furthermore, certain embodiments described herein relate to methods and systems for utilizing trained machine learning models, as well as methods and systems for providing trained machine learning models. Features, advantages, or alternative embodiments described herein can be assigned to other claimed objects, and vice versa. In other words, the claims and embodiments for providing trained machine learning models can be modified by features described or claimed in the context of utilizing trained machine learning models, and vice versa. Specifically, the dataset in the methods and systems for utilizing trained machine learning models can have the same properties and characteristics as the corresponding dataset in the methods and systems for providing trained machine learning models, and the trained machine learning model provided by the corresponding method and system can be used in the methods and systems for utilizing trained machine learning models.

[0064] Generally, trained machine learning models simulate human cognitive functions related to other aspects of human thought. Specifically, through training on training data, machine learning models can adapt to new environments and detect and infer various patterns. Another term for a "trained machine learning model" is "trained function."

[0065] Generally, the parameters of a machine learning model can be adapted through training. Specifically, supervised training, semi-supervised training, unsupervised training, reinforcement learning, and / or active learning can be used. Additionally, representation learning (an alternative term is "feature learning") can be employed. Specifically, the parameters of a machine learning model can be iteratively adapted through several training steps. Specifically, during training, a cost function can be minimized. Specifically, in the training of neural networks, the backpropagation algorithm can be used.

[0066] In particular, machine learning models, such as Figure 1 The linear projector used in step 104, the dynamic convolutional layer used in step 108, the machine learning-based encoder used in step 110, and the machine learning-based decoder used in step 112, Figure 2 The linear projector 206, dynamic convolutional layer 208, transformer encoder 212, and transformer decoder 216, Figure 3 The linear projector used in step 306, the dynamic convolutional layer used in step 310, the machine learning-based encoder used in step 312, and the machine learning-based decoder used in step 316, and Figure 4 The linear projector 406, dynamic convolutional layer 410, transformer encoder 414, and transformer decoder 420 may include, for example, neural networks, support vector machines, decision trees, and / or Bayesian networks, and / or machine learning models may be based on, for example, k-means clustering, Q-learning, genetic algorithms, and / or association rules. Specifically, the neural network may be, for example, a deep neural network, a convolutional neural network, or a convolutional deep neural network. Furthermore, the neural network may be, for example, an adversarial network, a deep adversarial network, and / or a generative adversarial network, etc.

[0067] Figure 8 An embodiment of an artificial neural network 800 that can be used to implement one or more machine learning models described herein is shown. Alternative terms for “artificial neural network” are “neural network,” “artificial neural network,” or “neural network.”

[0068] The artificial neural network 800 includes nodes 820,…,832 and edges 840,…,842, where each edge 840,…,842 is a directed connection from a first node 820,…,832 to a second node 820,…,832. Generally, the first node 820,…,832 and the second node 820,…,832 are different nodes 820,…,832, but they can also be the same node. For example, in… Figure 8In the diagram, edge 840 is a directed connection from node 820 to node 823, and edge 842 is a directed connection from node 830 to node 832. Edges 840, ..., 842 from the first node 820, ..., 832 to the second node 820, ..., 832 are also labeled as the "incoming edge" of the second node 820, ..., 832 and the "outgoing edge" of the first node 820, ..., 832.

[0069] In this embodiment, nodes 820,...,832 of the artificial neural network 800 can be arranged in layers 810,...,813, wherein each layer may include an inherent order introduced by edges 840,...,842 between nodes 820,...,832. Specifically, edges 840,...,842 may exist only between adjacent node layers. In the illustrated embodiment, there is an input layer 810 comprising only nodes 820,...,822 without incoming edges, an output layer 813 comprising only nodes 831, 832 without outgoing edges, and hidden layers 811, 812 between the input layer 810 and the output layer 813. Generally, the number of hidden layers 811, 812 can be arbitrarily chosen. The number of nodes 820,...,822 in the input layer 810 is typically related to the number of input values ​​of the neural network, and the number of nodes 831, 832 in the output layer 813 is typically related to the number of output values ​​of the neural network.

[0070] Specifically, (real) numbers can be assigned as values ​​to each node 820,…,832 of the neural network 800. Here, This represents the value of the i-th node 820, ..., 832 in the n-th layer 810, ..., 813. The values ​​of nodes 820, ..., 822 in the input layer 810 are equivalent to the input values ​​of the neural network 800, and the values ​​of nodes 831 and 832 in the output layer 813 are equivalent to the output values ​​of the neural network 800. Furthermore, each side 840, ..., 842 may include weights, which are real numbers, specifically real numbers within the interval [-1, 1] or [0, 1]. This indicates the weight of the edge between the i-th node 820,...,832 of layer m (810,...,813) and the j-th node 820,...,832 of layer n (810,...,813). Additionally, abbreviations... Defined as weight .

[0071] Specifically, to calculate the output value of neural network 800, the input values ​​are propagated through the neural network. Specifically, the values ​​of nodes 820, ..., 832 in the (n+1)th layer 810, ..., 813 can be calculated based on the values ​​of nodes 820, ..., 832 in the nth layer 810, ..., 813 using the following formula: .

[0072] In this paper, the function f is the transfer function (another term is "activation function"). Known transfer functions include step functions, sigmoid functions (e.g., logic functions, generalized logic functions, hyperbolic tangent functions, arctangent functions, error functions, smooth step functions), or rectifier functions. Transfer functions are primarily used for standardization purposes.

[0073] Specifically, these values ​​are propagated layer by layer through the neural network, where the value of the input layer 810 is given by the input of the neural network 800, the value of the first hidden layer 811 can be calculated based on the value of the input layer 810 of the neural network, the value of the second hidden layer 812 can be calculated based on the value of the first hidden layer 811, and so on.

[0074] To set the value of the edge Training data is required to train the neural network 800. Specifically, the training data includes training input data and training output data (labeled as...). For the training step, a neural network 800 is applied to the training input data to generate computed output data. Specifically, the training data and the computed output data include multiple values, the number of which is equal to the number of nodes in the output layer.

[0075] Specifically, the comparison between the calculated output data and the training data is used to recursively adapt the weights within the neural network (backpropagation algorithm). Specifically, the weights are changed as follows: in, The learning rate, and the number It can be recursively calculated as follows: based on If the (n+1)th layer is not an output layer, and: If the (n+1)th layer is the output layer 813, where f' is the first derivative of the activation function, and It is the comparison training value of the j-th node of the output layer 813.

[0076] A convolutional neural network (CNN) is a neural network that uses convolution operations instead of general matrix multiplication in at least one of its layers (so-called "convolutional layers"). Specifically, a convolutional layer performs a dot product of one or more convolutional kernels with the input data / image of the convolutional layer, where the entries of the one or more convolutional kernels are parameters or weights adapted through training. In particular, Frobenius inner products and ReLU activation functions can be used. CNNs can include additional layers such as pooling layers, fully connected layers, and normalization layers.

[0077] Convolutional neural networks (CNNs) can process input images very efficiently because convolution operations based on different kernels can extract various image features, allowing relevant image features to be found during training by adjusting the weights of the convolution kernels. Furthermore, due to weight sharing within the convolution kernels, fewer parameters need to be trained, preventing overfitting during training and allowing for faster training or more layers in the network, thus improving network performance.

[0078] Figure 9 An embodiment of a convolutional neural network 900 that can be used to implement one or more machine learning models described herein is shown. In the shown embodiment, the convolutional neural network 900 includes an input node layer 910, a convolutional layer 911, a pooling layer 913, a fully connected layer 914, and an output node layer 916, as well as hidden node layers 912 and 914. Alternatively, the convolutional neural network 900 may include a plurality of convolutional layers 911, a plurality of pooling layers 913, and a plurality of fully connected layers 915, as well as other types of layers. The order of the layers can be chosen arbitrarily; typically, the fully connected layer 915 is used as the last layer before the output layer 916.

[0079] Specifically, within the convolutional neural network 900, nodes 920, 922, and 924 of node layers 910, 912, and 914 can be considered as arranged as a d-dimensional matrix or a d-dimensional image. In particular, in the two-dimensional case, the values ​​of nodes 920, 922, and 924 indexed by i and j in the nth node layer 910, 912, and 914 can be represented as... However, the arrangement of nodes 920, 922, and 924 in a node layer 910, 912, and 914 does not affect the computation performed within the convolutional neural network 900, as these are given solely by the structure and weights of the edges.

[0080] Convolutional layer 911 is a connection layer between the preceding node layer 910 (node ​​value x(n-1)) and the following node layer 912 (node ​​value x(n)). Specifically, convolutional layer 911 is characterized by the structure and weights of the input edges that form the convolution operation based on a certain number of kernels. In particular, the structure and weights of the edges of convolutional layer 911 are chosen such that the value x(n) of node 922 of the following node layer 912 is computed as a convolution x(n) = K * x(n-1) based on the value x(n-1) of node 920 of the preceding node layer 910, where convolution * is defined in the two-dimensional case as: .

[0081] Here, the kernel K is a d-dimensional matrix (a two-dimensional matrix in this embodiment), which is typically small compared to the number of nodes 920 and 922 (e.g., a 3×3 or 5×5 matrix). Specifically, this implies that the weights of the edges in convolutional layer 911 are not independent, but are chosen such that they produce the convolution equation. In particular, for a 3×3 kernel, regardless of the number of nodes 920 and 922 in the preceding and following layers 910, there are only 9 independent weights (each entry in the kernel matrix corresponds to one independent weight).

[0082] Generally, convolutional neural networks 900 use node layers 910, 912, and 914 with multiple channels, especially since multiple kernels are used in convolutional layer 911. In these cases, the node layer can be viewed as a (d+1)-dimensional matrix (the first dimension indexes the channels). The action of convolutional layer 911 is then illustrated in a two-dimensional example, defined as: in Corresponding to the a-th channel of the previous node layer 910, Corresponding to the b-th channel of the subsequent node layer 912, and This corresponds to one of the kernels. If convolutional layer 911 acts on a preceding node layer 910 with A channels and outputs a following node layer 912 with B channels, then there exist A ∙ B independent d-dimensional kernels. .

[0083] Generally, convolutional neural networks 900 use activation functions. In this embodiment, ReLU (an abbreviation for "Rectified Linear Unit") is used, where R(z) = max(0, z), such that the action of convolutional layer 911 in the two-dimensional example is... .

[0084] Other activation functions can also be used, such as ELU (an abbreviation for "Exponential Linear Unit"), LeakyReLU, Sigmoid, Tanh, or Softmax.

[0085] In the shown embodiment, the input layer 910 includes 36 nodes 920 arranged in a two-dimensional 6×6 matrix. The first hidden node layer 912 includes 72 nodes 922 arranged in two two-dimensional 6×6 matrices, each of which is the result of convolving the values ​​of the input layer with a 3×3 kernel within the convolutional layer 911. Equivalently, the nodes 922 of the first hidden node layer 912 can be interpreted as being arranged in a three-dimensional 2×6×6 matrix, where the first dimension corresponds to the channel dimension.

[0086] The advantage of using convolutional layers 911 is that it can take advantage of the spatial local correlation of the input data by enforcing a local connectivity pattern between nodes in adjacent layers, in particular by having each node connect only to a small region of the node in the previous layer.

[0087] Pooling layer 913 is the connecting layer between the preceding node layer 912 (node ​​value x(n-1)) and the following node layer 914 (node ​​value x(n)). Specifically, pooling layer 913 can be characterized by the structure and weights of the edges and activation functions that form the pooling operation based on the nonlinear pooling function f. For example, in the two-dimensional case, the value x(n) of node 924 in the following node layer 914 can be calculated based on the value x(n-1) of node 922 in the preceding node layer 912 as follows: .

[0088] In other words, by using pooling layer 913, the number of nodes 922 and 924 can be reduced by replacing d1·d2 neighboring nodes 922 in the preceding node 912 with a single node 922 in the following node 914, the value of which is calculated as a function of the number of neighboring nodes. Specifically, the pooling function f can be a maximum function, an average function, or an L2 norm. Specifically, for pooling layer 913, the weights of the incoming edges are fixed and not modified during training.

[0089] The advantage of using pooling layer 913 is that it reduces the number of nodes 922 and 924 and the number of parameters. This leads to a reduction in the computational cost of the network and controls overfitting.

[0090] In the illustrated embodiment, pooling layer 913 is a max-pooling layer that replaces four adjacent nodes with only one node whose value is the maximum of the values ​​of those four adjacent nodes. Max pooling is applied to each d-dimensional matrix of the previous layer; in this embodiment, max pooling is applied to each of the two 2D matrices, reducing the number of nodes from 72 to 18.

[0091] Generally, the last few layers of a convolutional neural network (CNN) are fully connected layers (915). Fully connected layer 915 connects the preceding node layer (914) and the following node layer (916). A characteristic of fully connected layer 913 is that most (especially all) edges exist between node 914 of the preceding node layer (914) and the following node layer (916), and the weights of each of these edges can be adjusted individually.

[0092] In this embodiment, the nodes 924 of the preceding node layer 914 of the fully connected layer 915 are all displayed as a two-dimensional matrix, and also additionally as unrelated nodes (indicated as a row of nodes, where the number of nodes is reduced for better presentation). This operation is also called "flattening". In this embodiment, the number of nodes 926 in the following node layer 916 of the fully connected layer 915 is less than the number of nodes 924 in the preceding node layer 914. Alternatively, the number of nodes 926 can be equal to or greater than the number of nodes 924.

[0093] Furthermore, in this embodiment, the Softmax activation function is used within the fully connected layer 915. By applying the Softmax function, the sum of the values ​​of all nodes 926 in the output layer 916 is 1, and all values ​​of all nodes 926 in the output layer 916 are real numbers between 0 and 1. Specifically, if the input data is classified using a convolutional neural network 900, the values ​​of the output layer 916 can be interpreted as the probability that the input data falls into one of the different categories.

[0094] Specifically, the convolutional neural network 900 can be trained based on the backpropagation algorithm. To prevent overfitting, regularization methods can be used, such as dropping nodes 920,…,924, random pooling, using artificial data, weight decay based on L1 or L2 norm, or maximum norm constraint.

[0095] According to one aspect, a machine learning model may include one or more Residual Networks (ResNets). Specifically, a ResNet is an artificial neural network that includes at least one jump or skip connection used to skip at least one layer of the artificial neural network. In particular, a ResNet may be a convolutional neural network, including one or more skip connections that skip one or more convolutional layers respectively. According to some examples, a ResNet can be represented as an m-layer ResNet, where m is the number of layers in the corresponding architecture, and according to some examples, it can take values ​​of 34, 50, 101, or 152. According to some examples, such an m-layer ResNet may include (m-2) / 2 skip connections respectively.

[0096] Skip connections can be viewed as bypasses that feed the output of a preceding layer directly to one or more bypassed layers. The bypassed layers then have to fit a residual map that "balances" the directly fed output, rather than directly fitting the desired map.

[0097] Fitting residual mappings is computationally easier to optimize than directional mappings. More importantly, this alleviates the vanishing / exploding gradient problem during optimization when training machine learning models: if a bypassed layer encounters such a problem, its contribution can be skipped by regularizing the output directly fed to it. Therefore, one advantage of using ResNet is that it allows for the training of deeper networks.

[0098] Generative adversarial models (GA models) consist of a generator function and a discriminator function. The generator function creates synthetic data, and the discriminator function distinguishes between synthetic and real data. By training the generator and / or discriminator functions, the generator function is configured to create synthetic data that the discriminator function misclassifies as real data, while the discriminator function is configured to distinguish between real data and synthetic data generated by the generator function. From a game theory perspective, GA models can be interpreted as zero-sum games. The training of the generator and / or discriminator functions is specifically based on minimizing a cost function.

[0099] By using a Generative Analytics (GA) model, synthetic data with the same characteristics as the training dataset can be generated based on a set of training data. The GA model can be trained on unlabeled data (unsupervised learning), making the training workload minimal.

[0100] Figure 10 The diagram shows a data flow graph according to one or more embodiments of a generative adversarial network used to create synthetic output data G(x) 1008 based on input data x 1002 that is indistinguishable from real output data y 1004. The synthetic output data G(x) 1008 has the same structure as the real output data y 1004, but its content is not derived from real-world data.

[0101] The Generative Adversarial Network (GAN) consists of a jointly trained generator function G1006 and a classifier function C1010. The generator function G1006 is tasked with providing realistic synthetic output data G(x)1008 based on the input data x1002, and the classifier function C1010 is tasked with distinguishing between the real output data y1004 and the synthetic output data G(x)1008. Specifically, the output of the classifier function C1010 is a real number between 0 and 1, corresponding to the probability that the input value is real data, such that an ideal classifier function would compute an output value of C(y)1014≈1 for the real data y1004 and an output value of C(G(x))1012≈0 for the synthetic data G(x)1008.

[0102] During training, the parameters of the generator function G 1006 are adjusted so that the synthetic output data G(x) 1008 has the same features as the real output data y 1004, making it impossible for the classifier function C 1010 to distinguish between real and synthetic data. Simultaneously, the parameters of the classifier function C 1010 are adjusted to optimally distinguish between real and synthetic data. Here, training relies on pairs including input data x 1002 and the corresponding real output data y 1004. Within a single training step, the generator function G 1006 is applied to the input data x 1002 to generate the synthetic output data G(x) 1008. Further, the classifier function C 1010 is applied to the real output data y 1004 to generate the first classification result C(y) 1014. Additionally, the classifier function C 1010 is applied to the synthetic output data G(x) 1008 to generate the second classification result C(G(x)) 1012.

[0103] The parameters of the generating function G 1006 and the classifier function C 1010 are adjusted based on minimizing the cost function using the backpropagation algorithm. In this embodiment, the cost function K of the classifier function C 1010 is... C For K C ∝ - BCE(C(y),1) - BCE(C(G(x)), 0), where the BCE indices are defined as the binary cross-entropy of BCE(z, z′) = z′·log(z) + (1 − z′)·log(1 − z). By using this cost function, misclassifying true output data as synthetic data (indicated by C(y)≈0) and misclassifying synthetic output data as true data (indicated by C(G(x))1012≈1) both result in the cost function K to be minimized being less than the cost function K. C Increase. Furthermore, the cost function K of the generator function G 1006... G For K G ∝ -BCE(C(G(x)), 1) = -log(C(G(x)). Using this cost function, correctly classified synthetic output data (indicated as C(G(x))1012≈0) results in a cost function K to be minimized. G The increase.

[0104] In particular, a recurrent machine learning model is a machine learning model whose output depends not only on the input values ​​and parameters of the machine learning model adjusted during the training process, but also on the hidden state vector, which is based on the previous inputs used in the recurrent machine learning model. Specifically, a recurrent machine learning model may include additional stored states or additional structures that include time delays or feedback loops.

[0105] In particular, the underlying structure of a recurrent machine learning model can be a neural network, which can be called a recurrent neural network. Such a recurrent neural network can be described as an artificial neural network where the connections between nodes form a directed graph along a time series. Specifically, a recurrent neural network can be interpreted as a directed acyclic graph. Furthermore, a recurrent neural network can be a finite-spiking recurrent neural network or an infinite-spiking recurrent neural network (where a finite-spiking network can be unfolded and replaced with a strictly feedforward neural network, and an infinite-spiking network cannot be unfolded and replaced with a strictly feedforward neural network).

[0106] In particular, training recurrent neural networks can be based on the BPTT algorithm (an acronym for "backward propagation in time"), the RTRL algorithm (an acronym for "real-time recurrent learning"), and / or genetic algorithms.

[0107] By using recurrent machine learning models, input data that includes sequences of variable length can be used. Specifically, this implies that the method can be used not only with a fixed number of input datasets (and requiring different training for every other dataset used as input), but also with any number of input datasets. This implies that regardless of the number of input datasets contained in different sequences, the entire set of training data can be used for training, and the training data will not be reduced to the training data corresponding to a fixed number of consecutive input datasets.

[0108] Figure 11 A schematic structure of a recurrent machine learning model F, represented by both loop representation 1102 and expanded representation 1104, is shown. This model can be used to implement one or more machine learning models described in this paper. The recurrent machine learning model takes several input datasets x, x1, ..., x2. n 1106 is used as input and corresponding output datasets y, y1, …, y are created. n 1108. Furthermore, the output depends on the so-called hidden vectors h, h1, ..., h2. n 1110, the hidden vectors implicitly include information about the input dataset previously used as input to the recurrent machine learning model F 1112. By using these hidden vectors h, h1, ..., h n 1110, the order of the input dataset can be utilized.

[0109] In a single step of the process, the recurrent machine learning model F1112 takes the hidden vector h created in the previous step. n-1 and input dataset x n As input. Within this step, the recurrent machine learning model F generates the updated hidden vector h. n and output dataset y n As output. In other words, one step of the processing calculates (y). n, h n ) = F(x n , h n-1 Alternatively, by splitting the recurrent machine learning model F1112 into a step of computing the output data F(y) and computing the hidden vector F(h), the processing step computes y. n = F （y） (x) n , h n-1 ) and h n = F （h） (x) n , h n-1 For the first processing step, h0 can be randomly selected, or all entries can be filled with zeros. The parameters of the recurrent machine learning model F 1112, previously trained on the training dataset, do not change between different processing steps.

[0110] Specifically, the output data and hidden vectors of the processing step depend on all the previous input datasets used in the preceding steps. n = F （y） (x) n , F （h） (x) n-1 , h n-2 )) and h n = F(h)(x) n , F （h） (x) n-1 , h n-2 )).

[0111] The systems, apparatus, and methods described herein can be implemented using digital circuitry or using one or more computers employing known computer processors, memory units, storage devices, computer software, and other components. Typically, a computer includes a processor for executing instructions and one or more memories for storing instructions and data. A computer may also include or be coupled to one or more mass storage devices, such as one or more disks, internal hard disks and removable disks, magneto-optical disks, optical disks, etc.

[0112] The systems, apparatus, and methods described herein can be implemented using computers operating in a client-server relationship. Typically, in such a system, the client computer is located remotely from the server computer and interacts via a network. The client-server relationship can be defined and controlled by computer programs running on the respective client and server computers.

[0113] The systems, apparatus, and methods described herein can be implemented in a network-based cloud computing system. In such a system, a server or other processor connected to the network communicates with one or more client computers via the network. For example, a client computer may communicate with the server via a web browser application residing on and running on the client computer. The client computer may store and access data on the server via the network. The client computer may transmit data requests or online service requests to the server via the network. The server may perform the requested service and provide data to one or more client computers(s). The server may also transmit data adapted to enable the client computer to perform specific functions, such as performing calculations, displaying specific data on a screen, etc. For example, the server may transmit requests adapted to enable the client computer to perform one or more steps or functions of the methods and workflows described herein, including... Figures 1-4 One or more steps or functions. Some steps or functions of the methods and workflows described herein include... Figures 1-4 One or more steps or functions may be performed by a server or another processor in a network-based cloud computing system. Some steps or functions of the methods and workflows described herein include... Figures 1-4 One or more steps can be performed by a client computer in a web-based cloud computing system. The steps or functions of the methods and workflows described herein include... Figures 1-4 One or more steps can be performed by servers and / or client computers in a web-based cloud computing system in any combination.

[0114] The systems, apparatus, and methods described herein can be implemented using a computer program product tangibly embodied in an information carrier, such as in a non-transitory machine-readable storage device, for execution by a programmable processor; and the methods and workflow steps described herein include Figures 1-4 One or more steps or functions can be implemented using one or more computer programs that can be executed by such a processor. A computer program is a set of computer program instructions that can be used directly or indirectly in a computer to perform a specific activity or produce a specific result. Computer programs can be written in any form of programming language, including compiled or interpreted languages, and can be deployed in any form, including as a standalone program or as a module, component, subroutine, or other unit suitable for use in a computing environment.

[0115] Figure 12A high-level block diagram of an example computer 1202 that can be used to implement the systems, apparatus, and methods described herein is depicted. Computer 1202 includes a processor 1204 operatively coupled to a data storage device 1212 and a memory 1210. Processor 1204 controls the operation of computer 1202 by executing computer program instructions that define the overall operation. The computer program instructions may be stored in data storage device 1212 or other computer-readable medium and loaded into memory 1210 when execution of the computer program instructions is desired. Therefore, Figures 1-4 The methods and workflow steps or functions can be defined by computer program instructions stored in memory 1210 and / or data storage device 1212, and controlled by processor 1204 that executes the computer program instructions. For example, the computer program instructions can be implemented as computer executable code programmed by those skilled in the art to perform... Figures 1-4 The methods and workflow steps or functions. Therefore, by executing computer program instructions, processor 1204 performs... Figures 1-4 The computer 1202 may include one or more network interfaces 1206 for communicating with other devices via a network. The computer 1202 may also include one or more input / output devices 1208 that enable a user to interact with the computer 1202 (e.g., a monitor, keyboard, mouse, speakers, buttons, etc.).

[0116] Processor 1204 may include both general-purpose and special-purpose microprocessors, and may be the sole processor of computer 1202 or one of multiple processors. For example, processor 1204 may include one or more central processing units (CPUs). Processor 1204, data storage device 1212 and / or memory 1210 may include one or more application-specific integrated circuits (ASICs) and / or one or more field-programmable gate arrays (FPGAs), or be supplemented or incorporated therein.

[0117] Each of the data storage device 1212 and the memory 1210 includes a tangible, non-transitory computer-readable storage medium. The data storage device 1212 and the memory 1210 may each include high-speed random access memory, such as dynamic random access memory (DRAM), static random access memory (SRAM), double data rate synchronous dynamic random access memory (DDR RAM), or other random access solid-state memory devices, and may include non-volatile memory, such as one or more disk storage devices, such as internal hard disks and removable disks, magneto-optical disk storage devices, optical disk storage devices, flash memory devices, semiconductor memory devices, such as erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), optical disc read-only memory (CD-ROM), digital versatile optical disc read-only memory (DVD-ROM), or other non-volatile solid-state storage devices.

[0118] Input / output device 1208 may include peripheral devices such as printers, scanners, displays, etc. For example, input / output device 1208 may include display devices such as cathode ray tube (CRT) or liquid crystal display (LCD) monitors for displaying information to a user, keyboards, and pointing devices such as mice or trackballs, through which the user can provide input to computer 1202.

[0119] Image acquisition device 1214 can be connected to computer 1202 to input image data (e.g., medical images) into computer 1202. Image acquisition device 1214 and computer 1202 can be implemented as a single device. Image acquisition device 1214 and computer 1202 can also communicate wirelessly via a network. In a possible embodiment, computer 1202 can be remotely located relative to image acquisition device 1214.

[0120] Any or all of the systems, apparatuses, and methods discussed herein can be implemented using one or more computers, such as computer 1202.

[0121] Those skilled in the art will recognize that actual computer or computer system implementations may have other structures and may include other components, and Figure 12 It is a high-level representation of some components of this computer for illustrative purposes.

[0122] Regardless of the use of grammatical terms, individuals with both male and female gender identities are included in the terminology.

[0123] The foregoing detailed embodiments should be understood as illustrative and exemplary in every respect, not restrictive, and the scope of the invention disclosed herein is not determined by the detailed embodiments, but by the claims interpreted to the maximum extent permitted by patent law. It should be understood that the embodiments shown and described herein are merely illustrative of the principles of the invention, and various modifications can be made by those skilled in the art without departing from the scope and spirit of the invention. Various other combinations of features can be implemented by those skilled in the art without departing from the scope and spirit of the invention.

[0124] The following is a list of non-limiting illustrative embodiments disclosed herein: Illustrative Example 1. A computer-implemented method comprising: receiving 1) one or more input medical images, each image in a domain, and 2) a domain code for each of the one or more input medical images, identifying its domain; for each corresponding domain code, determining one or more weights based on the corresponding domain code; updating one or more parameters of a dynamically convolutional layer based on the one or more weights; extracting a first set of features from the one or more input medical images using the dynamically convolutional layer having one or more updated parameters; encoding the first set of features into a second set of features using a machine learning-based encoder; performing a medical imaging analysis task based on the second set of features; and outputting the result of the medical imaging analysis task.

[0125] Illustrative Example 2. A computer implementation of the method of Illustrative Example 1, wherein the dynamic convolutional layer and the machine learning-based encoder are trained by: receiving 1) one or more training medical images, each image in a domain, and 2) training domain codes for each of the one or more training medical images, identifying its domain; masking patches of the one or more training medical images; determining one or more additional weights based on the corresponding training domain code for each corresponding one of the training domain codes; updating one or more additional parameters of the dynamic convolutional layer based on the one or more additional weights; extracting a third set of features from unmasked patches of the one or more training medical images using the dynamic convolutional layer with the one or more updated additional parameters; encoding the third set of features into a fourth set of features using the machine learning-based encoder; inserting labels representing the masked patches of the one or more training medical images into the fourth set of features; reconstructing the one or more training medical images using a machine learning-based decoder based on the fourth set of features with the inserted labels; training the dynamic convolutional layer and the machine learning-based encoder based on the one or more training medical images and the one or more reconstructed training medical images; and outputting the trained dynamic convolutional layer and the trained machine learning-based encoder.

[0126] Illustrative Example 3. A computer implementation of any of Illustrative Examples 1-2, wherein determining one or more weights based on corresponding domain codes includes: projecting the corresponding domain codes onto one or more weights using a linear projector.

[0127] Illustrative Example 4. A computer implementation of any of Illustrative Examples 1-3, wherein updating one or more parameters of a dynamic convolutional layer based on one or more weights includes: updating the weight parameters and bias parameters of the dynamic convolutional layer.

[0128] Illustrative Example 5. A computer implementation of any of Illustrative Examples 1-4, wherein updating one or more parameters of a dynamic convolutional layer based on one or more weights includes: determining the dot product of one or more parameters of the dynamic convolutional layer with a corresponding one of the one or more weights.

[0129] Illustrative Example 6. A computer implementation method as described in any of Illustrative Examples 1-5, wherein encoding a first set of features into a second set of features using a machine learning-based encoder includes: encoding the first set of features using positional embedding.

[0130] Illustrative Example 7. A computer implementation method as described in any of Illustrative Examples 1-6, wherein encoding a first set of features into a second set of features using a machine learning-based encoder includes: generating a modal embedding based on domain codes; and encoding the first set of features using the modal embedding.

[0131] Illustrative Example 8. A computer implementation method as described in any of Illustrative Examples 1-7, wherein the medical imaging analysis task includes medical image synthesis.

[0132] Illustrative Example 9. An apparatus comprising: means for receiving 1) one or more input medical images, each image in a domain, and 2) a domain code for each of the one or more input medical images identifying its domain; means for determining one or more weights based on a corresponding domain code for each corresponding domain code; means for updating one or more parameters of a dynamically convolutional layer based on the one or more weights; means for extracting a first set of features from the one or more input medical images using the dynamically convolutional layer having one or more updated parameters; means for encoding the first set of features into a second set of features using a machine learning-based encoder; means for performing a medical imaging analysis task based on the second set of features; and means for outputting the result of the medical imaging analysis task.

[0133] Illustrative Example 10. The apparatus of Illustrative Example 9, wherein the dynamic convolutional layer and the machine learning-based encoder are trained by: receiving 1) one or more training medical images, each image in a domain, and 2) training domain codes for each of the one or more training medical images, identifying its domain; masking patches of the one or more training medical images; determining one or more additional weights based on the corresponding training domain code for each corresponding one of the training domain codes; updating one or more additional parameters of the dynamic convolutional layer based on the one or more additional weights; extracting a third set of features from unmasked patches of the one or more training medical images using the dynamic convolutional layer with the one or more updated additional parameters; encoding the third set of features into a fourth set of features using the machine learning-based encoder; inserting labels representing the masked patches of the one or more training medical images into the fourth set of features; reconstructing the one or more training medical images using a machine learning-based decoder based on the fourth set of features with the inserted labels; training the dynamic convolutional layer and the machine learning-based encoder based on the one or more training medical images and the one or more reconstructed training medical images; and outputting the trained dynamic convolutional layer and the trained machine learning-based encoder.

[0134] Illustrative embodiment. 11. An apparatus as described in any of illustrative embodiments 9-10, wherein the component for determining one or more weights based on corresponding domain codes includes: a component for projecting the corresponding domain codes onto one or more weights using a linear projector.

[0135] Illustrative Example 12. An apparatus as described in any of Illustrative Examples 9-11, wherein the component for updating one or more parameters of a dynamic convolutional layer based on one or more weights includes: components for updating weight parameters and bias parameters of the dynamic convolutional layer.

[0136] Illustrative Example 13. An apparatus as described in any of Illustrative Examples 9-12, wherein the component for updating one or more parameters of a dynamic convolutional layer based on one or more weights includes: a component for determining the dot product of one or more parameters of the dynamic convolutional layer with a corresponding one of the one or more weights.

[0137] Illustrative Example 14. A non-transitory computer-readable storage medium includes instructions that, when executed by a computer, cause the computer to perform operations including: receiving 1) one or more input medical images, each image in a domain, and 2) a domain code for each of the one or more input medical images, identifying its domain; for each corresponding domain code, determining one or more weights based on the corresponding domain code; updating one or more parameters of a dynamically convolutional layer based on the one or more weights; extracting a first set of features from the one or more input medical images using the dynamically convolutional layer having one or more updated parameters; encoding the first set of features into a second set of features using a machine learning-based encoder; performing a medical imaging analysis task based on the second set of features; and outputting the results of the medical imaging analysis task.

[0138] Illustrative Example 15. As described in the non-transitory computer-readable storage medium of illustrative embodiment 14, the dynamic convolutional layer and the machine learning-based encoder are trained by: receiving 1) one or more training medical images, each in a domain, and 2) training domain codes for each of the one or more training medical images, identifying its domain; masking the patches of the one or more training medical images; determining one or more additional weights based on the corresponding training domain code for each corresponding one of the training domain codes; updating one or more additional parameters of the dynamic convolutional layer based on the one or more additional weights; extracting a third set of features from the unmasked patches of the one or more training medical images using the dynamic convolutional layer with the one or more updated additional parameters; encoding the third set of features into a fourth set of features using the machine learning-based encoder; inserting labels representing the masked patches of the one or more training medical images into the fourth set of features; reconstructing the one or more training medical images using the machine learning-based decoder based on the fourth set of features with the inserted labels; training the dynamic convolutional layer and the machine learning-based encoder based on the one or more training medical images and the one or more reconstructed training medical images; and outputting the trained dynamic convolutional layer and the trained machine learning-based encoder.

[0139] Illustrative Example 16. A non-transitory computer-readable storage medium as described in any of Illustrative Examples 14-15, wherein encoding a first set of features into a second set of features using a machine learning-based encoder includes: encoding the first set of features using positional embedding.

[0140] Illustrative Example 17. A non-transitory computer-readable storage medium as described in any of Illustrative Examples 14-16, wherein encoding a first set of features into a second set of features using a machine learning-based encoder includes: generating a modal embedding based on domain codes; and encoding the first set of features using the modal embedding.

[0141] Illustrative Example 18. A non-transitory computer-readable storage medium as described in any of Illustrative Examples 14-17, wherein the medical imaging analysis task includes medical image synthesis.

[0142] Illustrative Example 19. A computer-implemented method comprising: receiving 1) one or more training medical images, each image in a domain, and 2) domain codes for each of the one or more training medical images, identifying its domain; masking patches of the one or more training medical images; determining one or more weights based on the corresponding domain code for each corresponding domain code; updating one or more parameters of a dynamic convolutional layer based on one or more additional weights; extracting a first set of features from unmasked patches of the one or more training medical images using the dynamic convolutional layer having one or more updated parameters; encoding the first set of features into a second set of features using a machine learning-based encoder; inserting labels representing the masked patches of the one or more training medical images into the second set of features; reconstructing the one or more training medical images based on the second set of features having the inserted labels using a machine learning-based decoder; training the dynamic convolutional layer and the machine learning-based encoder based on the one or more training medical images and the one or more reconstructed training medical images; and outputting the trained dynamic convolutional layer and the trained machine learning-based encoder.

[0143] Illustrative Example 20. A computer implementation of the method as described in Illustrative Example 19, wherein determining one or more weights based on corresponding domain codes includes: projecting the corresponding domain codes onto one or more weights using a linear projector.

Claims

1. A computer implementation method, comprising: Receive 1) one or more input medical images, each image in a domain, and 2) a domain code for each of the one or more input medical images that identifies its domain; For each corresponding domain code, one or more weights are determined based on the corresponding domain code; Based on the one or more weights, update one or more parameters of the dynamic convolutional layer; Using the dynamic convolutional layer with one or more updated parameters, a first set of features is extracted from the one or more input medical images; The first set of features is encoded into the second set of features using a machine learning-based encoder; Perform medical imaging analysis tasks based on the second set of features; and Output the results of the medical imaging analysis task.

2. The computer implementation method as described in claim 1, wherein, The dynamic convolutional layer and the machine learning-based encoder are trained in the following manner: Receive 1) one or more training medical images, each image in a domain, and 2) a training domain code for each of the one or more training medical images, identifying its domain; Mask the patches of the one or more training medical images; For each corresponding training domain code, one or more additional weights are determined based on the corresponding training domain code; Based on the one or more additional weights, update one or more additional parameters of the dynamic convolutional layer; Using the dynamic convolutional layer with one or more updated additional parameters, a third set of features is extracted from the unmasked patches of the one or more training medical images; The machine learning-based encoder is used to encode the third set of features into the fourth set of features; The markers representing the masked patches of the one or more training medical images are inserted into the fourth set of features; The one or more training medical images are reconstructed using a machine learning-based decoder based on a fourth set of features with inserted labels. The dynamic convolutional layer and the machine learning-based encoder are trained based on the one or more training medical images and the one or more reconstructed training medical images. as well as Outputs trained dynamic convolutional layers and trained machine learning-based encoders.

3. The computer implementation method as described in claim 1, wherein, Determining one or more weights based on the corresponding domain code includes: The corresponding domain code is projected onto the one or more weights using a linear projector.

4. The computer implementation method as described in claim 1, wherein, Updating one or more parameters of the dynamic convolutional layer based on the one or more weights includes: Update the weight and bias parameters of the dynamic convolutional layer.

5. The computer implementation method as described in claim 1, wherein, Updating one or more parameters of the dynamic convolutional layer based on the one or more weights includes: Determine the dot product of one or more parameters of the dynamic convolutional layer with a corresponding one of the one or more weights.

6. The computer implementation method as described in claim 1, wherein, Encoding the first set of features into the second set of features using a machine learning-based encoder includes: The first set of features is encoded using positional embedding.

7. The computer implementation method as described in claim 1, wherein, Encoding the first set of features into the second set of features using a machine learning-based encoder includes: Modal embeddings are generated based on the domain codes; and The first set of features is encoded using the modality embedding.

8. The computer implementation method as described in claim 1, wherein, The medical imaging analysis task includes medical image synthesis.

9. An apparatus comprising: Components for receiving 1) one or more input medical images, each image in a domain, and 2) a domain code for each of the one or more input medical images that identifies its domain; For each corresponding domain code, a component is used to determine one or more weights based on the corresponding domain code; A component for updating one or more parameters of a dynamic convolutional layer based on the one or more weights; A component for extracting a first set of features from the one or more input medical images using the dynamic convolutional layer with one or more updated parameters; A component used to encode a first set of features into a second set of features using a machine learning-based encoder; Components used to perform medical imaging analysis tasks based on the second set of features; as well as A component used to output the results of the medical imaging analysis task.

10. The apparatus of claim 9, wherein, The dynamic convolutional layer and the machine learning-based encoder are trained in the following manner: Receive 1) one or more training medical images, each image in a domain, and 2) a training domain code for each of the one or more training medical images, identifying its domain; Mask the patches of the one or more training medical images; For each corresponding training domain code, one or more additional weights are determined based on the corresponding training domain code; Based on the one or more additional weights, update one or more additional parameters of the dynamic convolutional layer; Using the dynamic convolutional layer with one or more updated additional parameters, a third set of features is extracted from the unmasked patches of the one or more training medical images; The machine learning-based encoder is used to encode the third set of features into the fourth set of features; The markers representing the masked patches of the one or more training medical images are inserted into the fourth set of features; The one or more training medical images are reconstructed using a machine learning-based decoder based on a fourth set of features with inserted labels. The dynamic convolutional layer and the machine learning-based encoder are trained based on the one or more training medical images and the one or more reconstructed training medical images. as well as Outputs trained dynamic convolutional layers and trained machine learning-based encoders.

11. The apparatus of claim 9, wherein, The component for determining one or more weights based on the corresponding domain code includes: A component for projecting the corresponding domain code onto the one or more weights using a linear projector.

12. The apparatus of claim 9, wherein, The component for updating one or more parameters of a dynamic convolutional layer based on the one or more weights includes: A component used to update the weight and bias parameters of the dynamic convolutional layer.

13. The apparatus of claim 9, wherein, The component for updating one or more parameters of a dynamic convolutional layer based on the one or more weights includes: A component for determining the dot product of one or more parameters of the dynamic convolutional layer with a corresponding one of the one or more weights.

14. A non-transitory computer-readable storage medium comprising instructions that, when executed by a computer, cause the computer to perform operations including: Receive 1) one or more input medical images, each image in a domain, and 2) a domain code for each of the one or more input medical images that identifies its domain; For each corresponding domain code, one or more weights are determined based on the corresponding domain code; Based on the one or more weights, update one or more parameters of the dynamic convolutional layer; Using the dynamic convolutional layer with one or more updated parameters, a first set of features is extracted from the one or more input medical images; The first set of features is encoded into the second set of features using a machine learning-based encoder; Perform medical imaging analysis tasks based on the second set of features; and Output the results of the medical imaging analysis task.

15. The non-transitory computer-readable storage medium of claim 14, wherein, The dynamic convolutional layer and the machine learning-based encoder are trained in the following manner: Receive 1) one or more training medical images, each image in a domain, and 2) a training domain code for each of the one or more training medical images, identifying its domain; Mask the patches of the one or more training medical images; For each corresponding training domain code, one or more additional weights are determined based on the corresponding training domain code; Based on the one or more additional weights, update one or more additional parameters of the dynamic convolutional layer; Using the dynamic convolutional layer with one or more updated additional parameters, a third set of features is extracted from the unmasked patches of the one or more training medical images; The machine learning-based encoder is used to encode the third set of features into the fourth set of features; The markers representing the masked patches of the one or more training medical images are inserted into the fourth set of features; The one or more training medical images are reconstructed using a machine learning-based decoder based on a fourth set of features with inserted labels. The dynamic convolutional layer and the machine learning-based encoder are trained based on the one or more training medical images and the one or more reconstructed training medical images. as well as Outputs trained dynamic convolutional layers and trained machine learning-based encoders.

16. The non-transitory computer-readable storage medium of claim 14, wherein, Encoding the first set of features into the second set of features using a machine learning-based encoder includes: The first set of features is encoded using positional embedding.

17. The non-transitory computer-readable storage medium of claim 14, wherein, Encoding the first set of features into the second set of features using a machine learning-based encoder includes: Modal embeddings are generated based on the domain codes; and The first set of features is encoded using the modality embedding.

18. The non-transitory computer-readable storage medium of claim 14, wherein, The medical imaging analysis task includes medical image synthesis.

19. A computer-implemented method, comprising: Receive 1) one or more training medical images, each image in a domain, and 2) a domain code for each of the one or more training medical images that identifies its domain; Mask the patches of the one or more training medical images; For each corresponding domain code, one or more weights are determined based on the corresponding domain code; Based on the one or more additional weights, update one or more parameters of the dynamic convolutional layer; Using the dynamic convolutional layer with one or more updated parameters, a first set of features is extracted from unmasked patches of the one or more training medical images; The first set of features is encoded into the second set of features using a machine learning-based encoder; The markers representing the masked patches of the one or more training medical images are inserted into the second set of features; The one or more training medical images are reconstructed using a machine learning-based decoder based on a second set of features with inserted labels. The dynamic convolutional layer and the machine learning-based encoder are trained based on the one or more training medical images and the one or more reconstructed training medical images. as well as Outputs trained dynamic convolutional layers and trained machine learning-based encoders.

20. The computer implementation method as described in claim 19, wherein, Determining one or more weights based on the corresponding domain code includes: The corresponding domain code is projected onto the one or more weights using a linear projector.

Images