Deep learning-based brain micro-hemorrhage automatic identification method and system

Abstract

The invention discloses a deep learning-based brain micro-hemorrhage automatic identification method and system. The method comprises the following steps of: obtaining brain micro-hemorrhage candidates in a standardized phase image and a standardized magnetic sensitive weighted imaging image based on a 3D fast radial symmetric transformation algorithm; constructing a double-input channel model; inputting the cerebral microhemorrhage candidate body into the double-input-channel model, extracting phase feature data of the cerebral microhemorrhage candidate body through the phase image input channel, extracting magnetic sensitive feature data of the cerebral microhemorrhage candidate body through the magnetic sensitive weighted imaging image input channel, and training the double-input-channel model to obtain an optimal double-input-channel model; and inputting the to-be-analyzed cerebral microhemorrhage candidate body into the optimal double-input channel model, and outputting a type corresponding to the to-be-analyzed cerebral microhemorrhage candidate body. By the adoption of the technical scheme, the problem that in the prior art, cerebral microhemorrhage detection accuracy is poor is solved.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a method and system for automatically identifying cerebral microbleeds based on deep learning. Background Art

[0002] Existing cerebral microbleeds (CMBs) detection technology is mainly based on magnetic resonance imaging (MRI), especially SWI susceptibility weighted imaging (SWI). SWI is an MRI post-processing technology that generates a susceptibility weighted mask and multiplies it with the amplitude image to obtain SWI. Since blood products have strong magnetic susceptibility, SWI can effectively display CMBs. On SWI, cerebral microbleeds appear as small spherical or elliptical areas with low signals.

[0003] Automatic cerebral microbleed detection is usually divided into two stages: candidate detection and false alarm reduction. In the candidate detection stage, various image processing techniques (such as two-dimensional radial symmetry transformation and shape feature extraction) are used to identify potential cerebral microbleeds. In the false alarm reduction stage, machine learning or deep learning algorithms (such as random forest, support vector machine, convolutional neural network, etc.) are used to further screen the candidate detection results to reduce the false alarm rate. For example, the SWI-based convolutional neural network model has demonstrated its effectiveness in cerebral microbleed detection.

[0004] However, existing technologies have problems such as high false alarm rate, limited differentiation ability and dependence on imaging parameters. The root cause of these problems is that existing technologies mainly rely on a single SWI mode and do not make full use of other potentially useful information, such as phase images or quantitative susceptibility mapping (QSM). In addition, traditional machine learning methods require manual feature design, and it is difficult to design universal and effective features for cerebral microbleeds of different shapes and intensities.

[0005] Therefore, there is an urgent need for a cerebral microbleed detection method that can integrate multi-source information and utilize deep learning technology to improve the detection accuracy of cerebral microbleeds. Summary of the invention

[0006] The purpose of the embodiments of the present invention is to provide a deep learning-based automatic identification system, method, electronic device and storage medium for cerebral microbleeds, so as to solve the problem of poor accuracy in cerebral microbleed detection in the prior art.

[0007] To achieve the above objectives, an embodiment of the present invention provides a method for automatically identifying cerebral microbleeds based on deep learning, and the method specifically includes: Acquire phase images and susceptibility-weighted imaging images; Performing image quality optimization processing on the phase image and the susceptibility weighted imaging image to obtain an optimized phase image and an optimized susceptibility weighted imaging image; Standardizing the intensity dimensions of the optimized phase image and the optimized susceptibility weighted imaging image to obtain a standardized phase image and a standardized susceptibility weighted imaging image; Acquire cerebral microbleed candidates in the standardized phase image and the standardized susceptibility-weighted imaging image based on a 3D fast radial symmetry transformation algorithm; Constructing a dual-input channel model, wherein the dual-input channel model includes a phase image input channel and a susceptibility weighted imaging image input channel, the dual-input channel model includes a residual block, and each convolution unit in each residual block includes a 3D convolution layer, an activation function layer, and a batch normalization layer; Inputting the cerebral microbleed candidate into a dual-input channel model, extracting phase feature data of the cerebral microbleed candidate through a phase image input channel, extracting magnetic susceptibility feature data of the cerebral microbleed candidate through a magnetic susceptibility weighted imaging image input channel, and training the dual-input channel model to obtain an optimal dual-input channel model; The cerebral microbleed candidate to be analyzed is input into the optimal dual-input channel model, and the type corresponding to the cerebral microbleed candidate to be analyzed is output, wherein the type includes a true cerebral microbleed or a false positive cerebral microbleed.

[0008] On the basis of the above technical solution, the present invention can also be improved as follows: Furthermore, the performing image quality optimization processing on the phase image and the susceptibility weighted imaging image to obtain an optimized phase image and an optimized susceptibility weighted imaging image includes: The elongated cerebral microbleeds were randomly deleted so that the number of elongated cerebral microbleeds matched that of spherical cerebral microbleeds.

[0009] Furthermore, the step of normalizing the intensity dimensions of the optimized phase image and the optimized susceptibility weighted imaging image to obtain a standardized phase image and a standardized susceptibility weighted imaging image comprises: The intensity size of the susceptibility weighted imaging image is normalized by formula (1): Formula (1); In the formula, is the normalized susceptibility-weighted imaging image intensity, is the intensity of the original susceptibility-weighted imaging image, is the baseline intensity; The intensity size of the phase image is normalized by formula (2): Formula (2); In the formula, is the normalized phase image intensity, is the original phase image, is the main magnetic field strength, is the echo time.

[0010] Further, obtaining the cerebral microbleed candidate in the standardized phase image and the standardized susceptibility weighted imaging image based on the 3D fast radial symmetry transformation algorithm includes: Threshold processing is performed on the susceptibility weighted imaging image transformed by the 3D fast radial symmetry transformation algorithm to generate an initial candidate image; The white matter spheres in the initial candidate image are located, and based on the characteristic differences of tissues inside and outside the white matter spheres, the cerebral microbleed candidates are accurately extracted, wherein a first threshold is used in the area outside the white matter spheres and a second threshold is used in the area inside the white matter spheres to improve detection sensitivity and accuracy.

[0011] Furthermore, after locating the white matter balls in the initial candidate image, the method further includes: Acquire isolated pixels in the initial candidate image whose distance is less than a preset number of pixels, and connect the isolated pixels to obtain a plurality of connected regions; For each of the connected regions in the initial candidate graph, one candidate object or two candidate objects are generated according to the shape of the connected region, wherein: For spherical regions, only one candidate is generated, located at the center of the connected region. For the elongated region, two candidates are generated, located at the two end points on the main axis of the connecting region.

[0012] Furthermore, the acquiring of isolated pixels in the initial candidate image whose distance is less than a preset number of pixels, and connecting the isolated pixels to obtain a plurality of connected regions, comprises: The fractional anisotropy value of each connected region is calculated by formula (3) to quantitatively describe the difference in eigenvalues ​​of the connected region in different directions: Formula (3); Where FA is the fractional anisotropy value of each connected region, For connecting area The eigenvalues ​​in the direction, is the eigenvalue of the connected area in the y direction, is the eigenvalue of the connected area in the z direction, is the mean eigenvalue.

[0013] Furthermore, after acquiring the cerebral microbleed candidate in the standardized image based on the 3D fast radial symmetry transformation algorithm, the method further includes: Data enhancement processing is performed on each cerebral microbleed candidate, wherein the data enhancement processing includes rotation and translation.

[0014] An automatic cerebral microbleed recognition system based on deep learning, comprising: An acquisition module, used for acquiring phase images and susceptibility-weighted imaging images; A data processing module, used for performing image quality optimization processing on the phase image and the susceptibility weighted imaging image to obtain an optimized phase image and an optimized susceptibility weighted imaging image; A standardization processing module, used for standardizing the intensity size of the optimized phase image and the optimized susceptibility weighted imaging image to obtain a standardized phase image and a standardized susceptibility weighted imaging image; A candidate acquisition module, used for acquiring cerebral microbleeding candidates in the standardized phase image and the standardized susceptibility-weighted imaging image based on a 3D fast radial symmetric transformation algorithm; A construction module, used to construct a dual-input channel model, wherein the dual-input channel model includes a phase image input channel and a susceptibility weighted imaging image input channel, the dual-input channel model includes a residual block, and each convolution unit in each residual block includes a 3D convolution layer, an activation function layer, and a batch normalization layer; a training module, configured to input the cerebral microbleed candidate into a dual-input channel model, extract phase feature data of the cerebral microbleed candidate through a phase image input channel, extract magnetic susceptibility feature data of the cerebral microbleed candidate through a magnetic susceptibility weighted imaging image input channel, and train the dual-input channel model to obtain an optimal dual-input channel model; The identification module is used to input the cerebral microbleed candidate to be classified into the optimal dual-input channel model, and output the type corresponding to the cerebral microbleed candidate to be classified, wherein the type includes true cerebral microbleed or false positive cerebral microbleed.

[0015] An electronic device comprises a memory, a processor and a computer program stored in the memory and executable on the processor, wherein the steps of the method are implemented when the processor executes the computer program.

[0016] A non-transitory computer-readable storage medium stores a computer program, which implements the steps of the method when executed by a processor.

[0017] The embodiments of the present invention have the following advantages: This deep learning-based automatic identification method for cerebral microbleeds has effectively improved the yield, quality, accuracy and efficiency of CMBs detection by developing a two-stage CMBs detection algorithm, and has significant technical advantages and application prospects in the field of cerebral microbleeds detection. The method of combining the 3D-FRST algorithm with the deep learning model has achieved high sensitivity and accuracy in CMBs detection.

[0018] This deep learning-based automatic identification method for cerebral microbleeds significantly reduces false positives by fusing SWI and phase images, especially in terms of vascular-related false positives. The dual-input channel model has strong generalization ability, shows high sensitivity and accuracy on different data sets, is suitable for a variety of clinical scenarios and disease types, and expands the application scope of CMBs detection. The dual-input channel model has a relatively simple deep learning architecture, which significantly reduces the number of trainable parameters, reduces the risk of overfitting, and improves the generalization ability of the model.

[0019] This deep learning-based automatic identification method for cerebral microbleeds uses data preprocessing and normalization techniques to effectively reduce the variation caused by changes in imaging parameters, standardize the intensity scale of different types of images, and maximize the contrast between CMBs and background tissue.

[0020] This deep learning-based automatic identification method for cerebral microbleeds further stabilizes the prediction results through data augmentation (such as 3D rotation and translation) and test-time enhancement.

[0021] This deep learning-based automatic identification method for cerebral microbleeds adopts an automated CMBs detection process, simplifies the operation steps, reduces dependence on manual operation, improves the automation level of the detection process, reduces operation complexity, and facilitates clinical application. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] In order to more clearly illustrate the implementation methods of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for the implementation methods or the description of the prior art. Obviously, the drawings in the following description are only exemplary, and for ordinary technicians in this field, other implementation drawings can be derived from the provided drawings without creative work.

[0023] The structures, proportions, sizes, etc. illustrated in this specification are only used to match the contents disclosed in the specification so as to facilitate understanding and reading by persons familiar with the technology. They are not used to limit the conditions under which the present invention can be implemented, and therefore have no substantial technical significance. Any structural modification, change in proportion or adjustment of size shall still fall within the scope of the technical contents disclosed in the present invention without affecting the effects and purposes that can be achieved by the present invention.

[0024] Figure 1 Schematic diagram of the process of the automatic identification method of cerebral microbleeds based on deep learning in an embodiment of the present invention; Figure 2 The architecture diagram of the dual input channel model of the present invention; Figure 3 (a) shows the 3D-FRST of the present invention Schematic diagram of the first effect on the sensitivity of CMB candidate detection and the average number of false detections per case; Figure 3 (b) shows the 3D-FRST of the present invention. Schematic diagram of the second impact on the sensitivity of CMB candidate detection and the average number of false positives per case; Figure 3 (c) shows the 3D-FRST of the present invention. Schematic diagram of the third effect on the sensitivity of CMB candidate detection and the average number of false detections per case; FIG4( a ) is a schematic diagram of the first performance of the best single model of the present invention on the test data; FIG4( b ) is a schematic diagram of the second performance of the best single model of the present invention on the test data; FIG4( c ) is a schematic diagram of the third performance of the best single model of the present invention on the test data; FIG4( d ) is a fourth performance diagram of the best single model of the present invention on the test data; FIG4( e ) is a schematic diagram of the fifth performance of the best single model of the present invention on the test data; FIG4( f ) is a sixth performance diagram of the best single model of the present invention on the test data; FIG5( a ) is a schematic diagram showing the first effect of input scale, model average and CMB size on the best single model performance of the present invention; Figure 5 (b) is a schematic diagram of the second effect of input scale, model average and CMB size on the best single model performance of the present invention; Figure 5 (c) is a schematic diagram of the third effect of input scale, model average and CMB size on the best single model performance of the present invention; Figure 5 (d) is a schematic diagram of the fourth effect of input scale, model average and CMB size on the best single model performance of the present invention; Figure 5 (e) is a schematic diagram of the fifth effect of input scale, model average and CMB size on the best single model performance of the present invention; Figure 5 (f) is a schematic diagram of the sixth effect of input scale, model average and CMB size on the best single model performance of the present invention; Figure 6 Schematic diagram of using phase images and susceptibility weighted imaging images to distinguish calcification and CMB in the present invention; Figure 7Schematic diagram of the sensitivity and accuracy of the best phase+SWI model measured in each subject of the present invention; Figure 8 Schematic diagram of the main components of the deep learning-based automatic identification system for cerebral microbleeds in an embodiment of the present invention; Fig. 9 This is a schematic diagram of the physical structure of the electronic device provided by the present invention. DETAILED DESCRIPTION

[0025] The following is a description of the implementation of the present invention by specific embodiments. People familiar with the art can easily understand other advantages and effects of the present invention from the contents disclosed in this specification. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0026] Figure 1 FIG. 1 is a flow chart of an embodiment of a method for automatically identifying cerebral microbleeds based on deep learning in an embodiment of the present invention. Figure 1 As shown, an automatic identification method for cerebral microbleeds based on deep learning provided by an embodiment of the present invention includes the following steps S101 to S107.

[0027] S101, acquiring a phase image and a susceptibility weighted imaging image.

[0028] The elongated cerebral microbleeds were randomly deleted so that the number of elongated cerebral microbleeds matched that of spherical cerebral microbleeds.

[0029] Fake samples (labeled as 0) are obtained from the following sources: CMBs mimics with weak signals on raw amplitude maps and susceptibility-weighted imaging images, or mimics with strong signals on QSM images; False positives identified by the 3D-FRST algorithm; Randomly selected background regions.

[0030] S102, performing image quality optimization processing on the phase image and the susceptibility weighted imaging image to obtain an optimized phase image and an optimized susceptibility weighted imaging image.

[0031] S103, performing standardization processing on the intensity size of the optimized phase image and the optimized susceptibility weighted imaging image to obtain a standardized phase image and a standardized susceptibility weighted imaging image.

[0032] Specifically, the intensity size of the susceptibility weighted imaging image is standardized by formula (1): Formula (1); In the formula, is the normalized susceptibility-weighted imaging image intensity, is the intensity of the original susceptibility-weighted imaging image, is the baseline intensity; The intensity size of the phase image is normalized by formula (2): Formula (2); In the formula, is the normalized phase image intensity, is the original phase image, is the main magnetic field strength, is the echo time.

[0033] S104, acquiring cerebral microbleed candidates in the standardized phase image and the standardized susceptibility weighted imaging image based on a 3D fast radial symmetry transformation algorithm.

[0034] Specifically, the susceptibility weighted imaging image transformed by the 3D fast radial symmetry transformation algorithm is threshold processed to generate an initial candidate image; The white matter spheres in the initial candidate image are located, and based on the characteristic differences of tissues inside and outside the white matter spheres, the cerebral microbleed candidates are accurately extracted, wherein a first threshold is used in the area outside the white matter spheres and a second threshold is used in the area inside the white matter spheres to improve detection sensitivity and accuracy.

[0035] Acquire isolated pixels in the initial candidate image whose distance is less than a preset number of pixels, and connect the isolated pixels to obtain a plurality of connected regions; For each of the connected regions in the initial candidate graph, one candidate object or two candidate objects are generated according to the shape of the connected region, wherein: For spherical regions, only one candidate is generated, located at the center of the connected region. For the elongated region, two candidates are generated, located at the two end points on the main axis of the connecting region.

[0036] The fractional anisotropy value of each connected region is calculated by formula (3) to quantitatively describe the difference in eigenvalues ​​of the connected region in different directions: Formula (3); Where FA is the fractional anisotropy value of each connected region, For connecting area The eigenvalues ​​in the direction, is the eigenvalue of the connected area in the y direction, is the eigenvalue of the connected area in the z direction, is the mean eigenvalue.

[0037] After acquiring the cerebral microbleed candidate in the standardized image based on the 3D fast radial symmetry transformation algorithm, the method further includes: Data enhancement processing is performed on each cerebral microbleed candidate, wherein the data enhancement processing includes rotation and translation.

[0038] S105, constructing a dual-input channel model.

[0039] Specifically, the dual input channel model includes a phase image input channel and a susceptibility weighted imaging image input channel, the dual input channel model includes a residual block, and each convolution unit in each residual block includes a 3D convolution layer, an activation function layer and a batch normalization layer.

[0040] Figure 2 For the dual input channel model architecture, each convolutional unit in the residual block consists of a batch normalization layer, a ReLU layer, and a 3×3×3 convolutional layer. The number below each layer corresponds to the output size of the layer. For models with different numbers of input channels, the output size of all layers remains the same. The total number of parameters of the model shown is 72082 (of which 71794 are trainable).

[0041] The residual block structure is adopted and implemented in a fully pre-activated manner. Each residual block contains a 3D convolution layer, an activation function layer (Rectified Linear Unit, ReLU) and a batch normalization layer (Batch Normalization, BatchNorm), which are arranged in order. The kernel size of all 3D convolution layers is 3×3×3. When the size of the output feature map changes, 1×1×1 convolutions are used to form residual connections. When the spatial dimension is reduced, the stride of these convolutions (including 1×1×1 convolutions) is set to 2, and the stride of other convolutions is set to 1. After all residual blocks, a global average pooling operation is performed to compress the feature map into a one-dimensional vector. Next, two fully connected layers and a softmax layer are used for classification prediction.

[0042] S106, inputting the cerebral microbleed candidate into the dual-input channel model, extracting the phase feature data of the cerebral microbleed candidate through the phase image input channel, extracting the magnetic susceptibility feature data of the cerebral microbleed candidate through the magnetic susceptibility weighted imaging image input channel, and training the dual-input channel model to obtain the optimal dual-input channel model.

[0043] S107, inputting the cerebral microbleed candidate to be analyzed into the optimal dual-input channel model, and outputting the type corresponding to the cerebral microbleed candidate to be analyzed, wherein the type includes a true cerebral microbleed or a false positive cerebral microbleed.

[0044] The determination process of the dual-input channel model used to automatically identify cerebral microbleeds is as follows: First, the brain mask was generated using the Brain Extraction Tool (BET) in the FMRIB Software Library (FSL). Then, the Laplacian Algorithm (Laplacian) phase development and the Sophisticated Harmonic Artifact Reduction for Phase data (SHARP) were used to reduce the residual background field and artifacts in the filtered phase image. Finally, the Morphology Enabled Dipole Inversion (MEDI) algorithm was used to generate the magnetic susceptibility map, and the regularization parameter Set to 300.

[0045] Amplitude, phase, SWI and QSM images were interpolated to 0.5 mm isotropic resolution, which is the highest planar resolution of the data in this invention. For amplitude and susceptibility weighted imaging images, the baseline intensity of the image It is estimated from the regions in the brain where the sensitivity-weighted mask is equal to 1.

[0046] The intensity of the amplitude and susceptibility-weighted imaging images was normalized by the following formula: Finally, the range of pixel intensities on the normalized image is truncated to After normalization, the intensity of the background white matter area in all types of images is close to 0, while the absolute value of the CMBs intensity is close to 1.

[0047] Construct and test various combinations of input channels, including single channel (M: amplitude, P: phase, S: CMBs, Q: QSM), dual channels (MP, MQ, PS, SQ), triple channels (MPS, MPQ, MSQ, PSQ) and quad channels (MPSQ).

[0048] The 3D version of the Fast Radial Symmetry Transform (FRST) algorithm is used, and the scale parameter N is set to the range of 1 to 4 pixels, and the strict symmetry parameter Set to 2. First, the 3D-FRST transformed susceptibility weighted imaging image was thresholded to generate an initial candidate map. To avoid detecting increased levels of iron deposition, a threshold was used in areas outside the white matter bulb. , using 2 in the white matter sphere The localization of white matter spheres is based on the Harvard-Oxford atlas provided by FSL. The sensitivity of CMB candidate detection depends on the threshold To find the optimal threshold , the present invention adjusted the threshold value from 0.08 to 0.15 with an interval of 0.01. Finally, the optimal threshold value was selected based on the sensitivity and the average number of false alarms in the training and validation datasets.

[0049] Due to the noise and shape changes in CMBs images, a large number of isolated pixels may appear, which may cause unnecessary calculations in the false positive reduction stage. To solve this problem, the present invention connects isolated pixels with a distance of less than 4 pixels (about 2 mm) and connects them if no edge is detected. The edge here refers to the area with more than 80% of the gradients of all regions in the brain. Next, for each connected region in the candidate map, one or two candidates are generated according to the shape of the region. For spherical regions (such as normal CMBs), only one candidate is generated, located at the center of the region; for elongated regions (such as damaged veins), two candidates are generated, located at the two end points on the main axis of the region. This is done to fully represent CMBs of different shapes. Therefore, in this study, CMBs candidates may correspond to the center or one end point of potential CMBs. For shape analysis, the present invention uses the same formula in diffusion tensor imaging to calculate the fractional anisotropy (FA) of each connected region.

[0050] The present invention defines the elongated area as an area with a main axis length greater than 6 mm and a FA greater than 0.85 (for example, a cuboid with an aspect ratio of 3:2:1 is considered sufficiently spherical). Finally, a 3D volume with a matrix size of 32×32×32 is cropped from the amplitude, phase, SWI, and QSM images, and the cerebral microbleed candidate object is placed in the center. These 3D volumes will be classified as true cerebral microbleeds or false positive cerebral microbleeds by a deep learning model.

[0051] Except for the input layer, all layers have the same size and number of trainable parameters for different input types. In addition to the model with an input scale of Models of , trained using phase images and susceptibility weighted imaging images. For these models, the output size of the first convolutional layer is ,The first average pooling layer is removed to study the contribution of individual channels, the effectiveness of channel combinations, and the impact of input scale.

[0052] The training process was repeated twice for each type of input, and the four best performing models were selected based on the Area Under the Curve (AUC) value on the validation set. In each training process, the model weights were first randomly initialized and then updated using the Adaptive Moment Estimation (Adam) and Adaptive Moment Estimation (Adam with Warm Restarts, AdamWR) optimizer. The learning rate was set to 0.002, the weight decay was 0.003, and the batch size was 256. The training used a cosine annealing strategy with an initial warm-up cycle of 4 cycles (epochs), followed by two full training cycles, each containing 16 cycles. In the last cycle, the learning rate and weight decay were reduced to 0.8 of the initial value, respectively.

[0053] To avoid overfitting, online data augmentation techniques such as 3D rotation, translation, and random left-right flipping are used during training. The rotation angles around the three axes are sampled independently and follow a zero-mean Gaussian distribution. The rotation ranges are and The translation distance is generated in a similar way, with a maximum translation distance of 1 pixel in each direction. The initial matrix size of the training sample is , to avoid introducing too many zero values ​​during 3D rotation. After the transformation, the samples are clipped to , and input into the deep learning model.

[0054] The present invention adopts the test-time augmentation (TTA) technology in the step of reducing false positives, and compares the model performance with and without TTA.

[0055] Specifically, each cerebral microbleed candidate is rotated along the Z axis to 0 degrees (original sample), 90 degrees, 180 degrees and 270 degrees. For each rotation, the corresponding 3D volume is also flipped in the Z direction. Using each model, the present invention can generate 8 prediction values ​​for each cerebral microbleed candidate, and then take the average of these prediction values ​​as the final prediction result of the model. In addition, when there are multiple candidates from the same elongated area, the two candidates are regarded as a potential CMBs, and their probability values ​​are averaged.

[0056] The present invention uses a series of evaluation indicators to evaluate the performance of cerebral microbleed candidate detection and false positive reduction, including sensitivity, precision, average false positives per subject (Average False Positives per Subject, FPavg), average false positives per CMBs (False Positives per CMBs, FP / CMBs) and area under the precision-recall curve (Area Under the Precision-Recall Curve, AUC-PR). For AUC-PR, the present invention estimates the 95% confidence interval (Confidence Interval, CI) using the bootstrap method. The AUC-PR values ​​of the models are compared using the permutation test, and the maximum statistical technique is used to control the overall error rate. A p value less than 0.05 is considered to be significantly different.

[0057] In order to check the model performance, the present invention selects two probability thresholds: The first threshold is the High Sensitivity Threshold: its performance is closest to that of the most experienced SWI data processors; the second threshold is the High Specificity Threshold: there are no false positives in the healthy control group.

[0058] The present invention uses the SWI model as a reference and only selects a high sensitivity threshold. For each input type, the model with the highest AUC-PR value is marked as the best single model. In addition, the prediction results of models with the same input type are averaged and marked as the model average.

[0059] Figure 3 shows 3D-FRST The impact on the sensitivity and average number of false positives per case (FPavg) of CMB candidate detection. Figure 3(b) FPavg vs. Figure 3(c) Relationship between sensitivity and FPavg. The circle indicates the best 3D-FRST is 0.09. The impact on sensitivity and FPavg in the CMB candidate detection step, which are evaluated on the training and validation data. Based on the sensitivity and FPavg, the best threshold is chosen to be 0.09. At this threshold, the sensitivity is 97.9% and the FPavg is 312.5. On the test data, the sensitivity of CMB candidate detection is 99.4% and the FPavg is 276.8.

[0060] Figures 4(a) and 4(b) represent models with three or four input channels; Figures 4(c) and 4(d) represent models with two input channels. Without test-time augmentation, the performance of the two-input channel model degrades. Figures 4(e) and 4(f) represent models with a single input channel. The first column (Figures 4(a), 4(c), and 4(e)) is the free operating characteristic curve, and the second column (Figures 4(b), 4(d), and 4(f)) shows the relationship between sensitivity and precision, which is actually the transpose of the traditional precision-recall curve. The SWI model is shown as a reference. For better visualization, only the 95% confidence interval of the SWI model is plotted in Figures 4(b), 4(d), and 4(f). The letters M, P, S, and Q represent the input channel type. M: amplitude; P: phase; S: SWI; Q: QSM. The sensitivity of the CMB candidate detection stage has been included, and the present invention reflects the performance of the entire CMB detection framework.

[0061] Among the models using three or four input channels, the model using amplitude, phase, and QSM (denoted as MPQ model) had the largest AUC-PR: 0.91 [0.85, 0.94], with confidence intervals in parentheses. In contrast, the model using all four images (denoted as MPSQ model) had an AUC-PR of 0.87 [0.82, 0.91]. As shown in Figures 4(c) and 4(d), the model using a combination of phase images and susceptibility-weighted imaging images (denoted as dual-input channel model, AUC-PR: 0.92 [0.88, 0.95]) outperformed all other types of models and performed similarly to the most experienced SWI data processor. Without test time augmentation (TTA), the performance of the dual-input channel model (AUC-PR: 0.91 [0.87, 0.94]) was significantly reduced (p = 0.013), but still similar to the performance of radiologists and SWI data processors.

[0062] Among the single-channel input models as shown in Figures 4(e) and 4(f), the QSM model performed best (AUC-PR: 0.88 [0.82, 0.92]), followed by SWI (AUC-PR: 0.87 [0.83, 0.91]), phase (AUC-PR: 0.85 [0.79, 0.90]), and amplitude (AUC-PR: 0.83 [0.78, 0.88]), reflecting the practicality of different types of images. However, there was no significant difference between the AUC-PRs of the single-channel models. By comparing the AUC-PRs of the reference SWI model and the best single models with different numbers of input channels (i.e., MPSQ, MPQ, PS, and Q), only the dual-input channel model was significantly different from the reference SWI model (p = 0.024). The dual-input channel model was also significantly different from the MPSQ model (p = 0.015). No significant differences were observed between the other models.

[0063] The SWI model is shown as a reference. For better visualization, only the 95% confidence intervals of the SWI model are plotted in (b), (d), and (f). The letters M, P, S, and Q indicate the input channel types. M: amplitude; P: phase; S: SWI; Q: QSM.

[0064] Figure 5 shows the effects of input scale, model averaging, and CMB size on the performance of the best single model. Figures 5(a) and 5(b): The effect of input scale. The SWI model at an input scale of 32³ (black lines in Figures 5(a) and 5(b)) is shown as a reference. Figures 5(c) and 5(d): The effect of model averaging. The solid line represents the best single model and the dashed line represents the model averaging. Figures 5(e) and 5(f): The performance of the best single model when small lesions of size ≤ 2 pixels are excluded. The first column (Figures 5(a), 5(c), and 5(e)) is the free operating characteristic curve, and the second column (Figures 5(b), 5(d), and 5(f)) shows the relationship between sensitivity and accuracy. For better visualization, only the 95% confidence intervals of the SWI model are plotted in Figures 5(b), 5(d), and 5(f). As shown in Figures 5(a) and 5(b), the dual-input channel model with a 32×32×32 input scale significantly outperforms the model with a 16×16×16 input scale (p = 0.0002). As shown in Figures 5(c) and 5(d), the performance of the models did not improve significantly by simple averaging. The poor sensitivity of Evaluator 1 was partly due to the omission of some small lesions, as shown in Figures 5(e) and 5(f). When small CMBs with a volume of 2 pixels (a total of 15 CMBs, the size of CMBs was estimated by thresholded 3D-FRST images) were excluded, the sensitivity of Evaluator 1 increased to 84.3% with an accuracy of 84.3%. Meanwhile, the dual-input channel model and the MPQ model both outperformed Evaluators 1 and 2. The other models were still inferior to the Evaluators. The performance of the Evaluators, the dual-input channel model, and the reference SWI model are provided in Table 1.

[0065] Table 1 like Figure 6 As shown, (a) is the amplitude image, and (b) is the phase image. Calcification is negative in the phase image, while CMB is positive. (c) is the susceptibility weighted imaging image, and (d) is the quantitative susceptibility map. The scale bar is only applicable to (d). (e) is the susceptibility weighted imaging image after 3D-FRST conversion. (f) is the susceptibility weighted imaging image of the detected CMB candidate area, where both CMB and calcification are detected. (g) The model using susceptibility weighted imaging images failed to eliminate the false detection caused by calcification. (h) The model using phase and susceptibility weighted imaging images successfully eliminated a false detection. (b) to (h) correspond to the dotted box area in (a). When using the SWI model, the false detection caused by calcification was not eliminated, but it was successfully eliminated under the dual input channel model. The SWI model predicts that the probability of this calcification being a true CMB is 0.0%, while the dual input channel model reduces it to 19.3%.

[0066] like Figure 7As shown, two probability thresholds were used to calculate sensitivity and precision, one for the high precision scenario ((a) and (c)) and the other for the high sensitivity scenario ((b) and (d)). (a) and (b): Number of CMBs per subject versus sensitivity; (c) and (d): Number of false positives (FPs) per subject versus number of CMBs. Number of CMBs per subject versus sensitivity and number of false positives in the test data using the dual-input channel model. For stroke and hemodialysis cases, no significant difference between model performance was found on 1.5T and 3T data using t-tests. Under the high precision threshold ((a) and (c)), the average sensitivity (mean ± SE) was 99.1 ± 0.9% for 1.5T data and 91.8 ± 5.4% for 3T data, respectively; while the average number of false positives was 2.2 ± 0.4 for 1.5T data and 1.8 ± 0.9 for 3T data, respectively. With the exception of one hemodialysis case and one other TBI case, sensitivity was 100% when there were fewer than five CMBs in any given case but decreased as the number of CMBs increased.

[0067] A small lesion in the corpus callosum was missed in the CMB candidate detection step in a hemodialysis case. A misclassified injured vein in a TBI case, the case with the lowest sensitivity for the best phase + SWI model. A false positive near an asymmetrically prominent cortical vein in a stroke case, with the most false positives detected by the best phase + SWI model. The poor sensitivity in the hemodialysis case was mainly due to the small size of the bleed, and the poor sensitivity in the TBI case was due to the omission of two injured veins with prediction probabilities below the threshold. At a high precision threshold, some false positives were still detected in patients without CMBs when there were no false positives in the healthy control group.

[0068] In one stroke case, the dual-input channel model detected the most false positives. It was found that the false positives in this stroke case were mainly related to abnormally high magnetic susceptibility of the veins, caused by decreased venous oxygen saturation. In fact, these asymmetric prominent cortical veins are a typical hallmark of stroke. In the stroke case, the performance of the dual-input channel model using a high-precision threshold was: sensitivity: 98.3%, precision: 69.0%, FPavg: 2.89, FP / cmb: 0.44; when using a high-sensitivity threshold, the performance was: sensitivity: 100%, precision: 63.4%, FPavg: 3.78, FP / emb: 0.58.

[0069] FPavg: average number of false positives per case; FPhc: average number of false positives per case in healthy controls; FP / cmb: average number of false positives per CMB.

[0070] On the test data, the computation time per case was 48 seconds. The CMB candidate detection step was the most time-consuming, taking 39 seconds per case. When only the best single model was used for prediction, the false positive reduction step took only 9 seconds.

[0071] In summary, by fusing susceptibility-weighted imaging images and phase images, false positives are significantly reduced, especially in terms of vascular-related false positives. The dual-input channel model has strong generalization ability, showing high sensitivity and accuracy on different data sets, and is suitable for a variety of clinical scenarios and disease types, expanding the application scope of CMBs detection.

[0072] The method for automatic identification of cerebral microbleeds based on deep learning collects a certain number of phase images and susceptibility weighted imaging images; performs data preprocessing on the phase images and susceptibility weighted imaging images to obtain preprocessed images; performs standardization on the intensity size of the preprocessed images to obtain standardized images; obtains cerebral microbleed candidates in the standardized images based on a 3D fast radial symmetric transformation algorithm; constructs a dual-input channel model, wherein the dual-input channel model includes a phase image input channel and a susceptibility weighted imaging image input channel, the dual-input channel model includes a residual block, and each convolution unit in each residual block includes a 3D convolution layer, an activation function layer, and a batch normalization layer; uses the cerebral microbleed candidate as an input of the dual-input channel model, trains the dual-input channel model to obtain an optimal dual-input channel model; inputs the cerebral microbleed candidate to be analyzed into the optimal dual-input channel model, and outputs the type corresponding to the cerebral microbleed candidate to be analyzed, wherein the type includes real cerebral microbleeds or false-positive cerebral microbleeds. The problem of poor accuracy in cerebral microbleed detection in the prior art is solved.

[0073] Figure 8 Schematic diagram of the main components of the deep learning-based automatic identification system for cerebral microbleeds in an embodiment of the present invention; Figure 8 As shown, an automatic cerebral microbleed recognition system 1 based on deep learning provided by an embodiment of the present invention includes an acquisition module 10, a data processing module 20, a standardization processing module 30, a candidate acquisition module 40, a construction module 50, a training module 60 and a recognition module 70: The acquisition module 10 is used to acquire a certain number of phase images and susceptibility weighted imaging images.

[0074] The acquisition module 10 is also used to perform image quality optimization processing on the phase image and the susceptibility weighted imaging image to obtain an optimized phase image and an optimized susceptibility weighted imaging image.

[0075] The data processing module 20 is used to perform standardization processing on the intensity size of the optimized phase image and the optimized susceptibility weighted imaging image to obtain a standardized phase image and a standardized susceptibility weighted imaging image.

[0076] The standardization processing module 30 is used to obtain the cerebral microbleed candidate in the above-mentioned standardized phase image and the standardized susceptibility weighted imaging image based on the 3D fast radial symmetry transformation algorithm.

[0077] The candidate acquisition module 40 is used to acquire the cerebral microbleed candidate in the above-mentioned standardized image based on a 3D fast radial symmetric transformation algorithm.

[0078] The construction module 50 is used to construct a dual-input channel model, wherein the dual-input channel model includes a phase image input channel and a susceptibility weighted imaging image input channel, and the dual-input channel model includes a residual block, and each convolution unit in each residual block includes a 3D convolution layer, an activation function layer and a batch normalization layer.

[0079] The training module 60 is used to input the above-mentioned cerebral microbleeding candidates into the dual-input channel model, extract the phase feature data of the cerebral microbleeding candidates through the phase image input channel, extract the magnetic susceptibility feature data of the cerebral microbleeding candidates through the magnetic susceptibility weighted imaging image input channel, and train the dual-input channel model to obtain the optimal dual-input channel model.

[0080] The identification module 70 is used to input the cerebral microbleed candidate to be classified into the optimal dual-input channel model, and output the type corresponding to the cerebral microbleed candidate to be classified, wherein the type includes true cerebral microbleed or false positive cerebral microbleed.

[0081] Fig. 9 A schematic diagram of the physical structure of an electronic device provided in an embodiment of the present invention, such as Fig. 9 As shown, the electronic device 80 includes: a processor 801 (processor), a memory 802 (memory) and a bus 803; The processor 801 and the memory 802 communicate with each other via the bus 803; The processor 801 is used to call the program instructions in the memory 802 to execute the methods provided by the above-mentioned method embodiments.

[0082] This embodiment provides a non-transitory computer-readable storage medium, which stores computer instructions. The computer instructions enable a computer to execute the methods provided by the above method embodiments.

[0083] A person skilled in the art can understand that all or part of the steps of implementing the above method embodiment can be completed by hardware related to program instructions, and the aforementioned program can be stored in a computer-readable storage medium. When the program is executed, it executes the steps of the above method embodiment; and the aforementioned storage medium includes: ROM, RAM, magnetic disk or optical disk, etc., various storage media that can store program codes.

[0084] The device embodiments described above are merely illustrative, wherein the units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, i.e., they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of this embodiment. Those of ordinary skill in the art may understand and implement it without creative effort.

[0085] Through the description of the above implementation methods, those skilled in the art can clearly understand that each implementation method can be implemented by means of software plus a necessary general hardware platform, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solution is essentially or the part that contributes to the prior art can be embodied in the form of a software product, and the computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a disk, an optical disk, etc., including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods of each embodiment or some parts of the embodiment.

[0086] Although the present invention has been described in detail above by general description and specific embodiments, it is obvious to those skilled in the art that some modifications or improvements can be made to the present invention. Therefore, these modifications or improvements made without departing from the spirit of the present invention all belong to the scope of protection claimed by the present invention.

Claims

1. A method for automatic identification of cerebral microbleeds based on deep learning, characterized in that: include: Acquire phase images and susceptibility-weighted imaging images; Performing image quality optimization processing on the phase image and the susceptibility weighted imaging image to obtain an optimized phase image and an optimized susceptibility weighted imaging image; Standardizing the intensity dimensions of the optimized phase image and the optimized susceptibility weighted imaging image to obtain a standardized phase image and a standardized susceptibility weighted imaging image; Acquire cerebral microbleed candidates in the standardized phase image and the standardized susceptibility-weighted imaging image based on a 3D fast radial symmetry transformation algorithm; Constructing a dual-input channel model, wherein the dual-input channel model includes a phase image input channel and a susceptibility weighted imaging image input channel, the dual-input channel model includes a residual block, and each convolution unit in each residual block includes a 3D convolution layer, an activation function layer, and a batch normalization layer; Inputting the cerebral microbleed candidate into a dual-input channel model, extracting phase feature data of the cerebral microbleed candidate through a phase image input channel, extracting magnetic susceptibility feature data of the cerebral microbleed candidate through a magnetic susceptibility weighted imaging image input channel, and training the dual-input channel model to obtain an optimal dual-input channel model; The cerebral microbleed candidate to be analyzed is input into the optimal dual-input channel model, and the type corresponding to the cerebral microbleed candidate to be analyzed is output, wherein the type includes a true cerebral microbleed or a false positive cerebral microbleed.

2. The method for automatic identification of cerebral microbleeds based on deep learning according to claim 1, characterized in that: The performing image quality optimization processing on the phase image and the susceptibility weighted imaging image to obtain an optimized phase image and an optimized susceptibility weighted imaging image comprises: The elongated cerebral microbleeds were randomly deleted so that the number of elongated cerebral microbleeds matched that of spherical cerebral microbleeds.

3. The method for automatic identification of cerebral microbleeds based on deep learning according to claim 1, characterized in that: The step of normalizing the intensity size of the optimized phase image and the optimized susceptibility weighted imaging image to obtain a standardized phase image and a standardized susceptibility weighted imaging image comprises: The intensity size of the susceptibility weighted imaging image is normalized by formula (1): Formula (1): In the formula, is the normalized susceptibility-weighted imaging image intensity, is the intensity of the original susceptibility-weighted imaging image, is the baseline intensity; The intensity size of the phase image is normalized by formula (2): Formula (2): In the formula, is the normalized phase image intensity, is the original phase image, is the main magnetic field strength, is the echo time.

4. The method for automatic identification of cerebral microbleeds based on deep learning according to claim 1, characterized in that: The method of acquiring the cerebral microbleed candidate in the standardized phase image and the standardized susceptibility weighted imaging image based on the 3D fast radial symmetry transformation algorithm comprises: Threshold processing is performed on the susceptibility weighted imaging image transformed by the 3D fast radial symmetry transformation algorithm to generate an initial candidate image; The white matter spheres in the initial candidate image are located, and based on the characteristic differences of tissues inside and outside the white matter spheres, the cerebral microbleed candidates are accurately extracted, wherein a first threshold is used in the area outside the white matter spheres and a second threshold is used in the area inside the white matter spheres to improve detection sensitivity and accuracy.

5. The method for automatic identification of cerebral microbleeds based on deep learning according to claim 4, characterized in that: After locating the white matter ball in the initial candidate image, the method further includes: Acquire isolated pixels in the initial candidate image whose distance is less than a preset number of pixels, and connect the isolated pixels to obtain a plurality of connected regions; For each of the connected regions in the initial candidate graph, one candidate object or two candidate objects are generated according to the shape of the connected region, wherein: For spherical regions, only one candidate is generated, located at the center of the connected region. For the elongated region, two candidates are generated, located at the two end points on the main axis of the connecting region.

6. The method for automatic identification of cerebral microbleeds based on deep learning according to claim 5, characterized in that: The step of obtaining isolated pixels in the initial candidate image whose distance is less than a preset number of pixels, and connecting the isolated pixels to obtain a plurality of connected regions comprises: The fractional anisotropy value of each connected region is calculated by formula (3) to quantitatively describe the difference in eigenvalues ​​of the connected region in different directions: Formula (3): Where FA is the fractional anisotropy value of each connected region, For connecting area The eigenvalues ​​in the direction, is the eigenvalue of the connected area in the y direction, is the eigenvalue of the connected area in the z direction, is the mean eigenvalue.

7. The method for automatic identification of cerebral microbleeds based on deep learning according to claim 6, characterized in that: After acquiring the cerebral microbleed candidate in the standardized image based on the 3D fast radial symmetry transformation algorithm, the method further includes: Data enhancement processing is performed on each cerebral microbleed candidate, wherein the data enhancement processing includes rotation and translation.

8. A deep learning-based automatic identification system for cerebral microbleeds, characterized in that: include: An acquisition module, used for acquiring phase images and susceptibility-weighted imaging images; A data processing module, used for performing image quality optimization processing on the phase image and the susceptibility weighted imaging image to obtain an optimized phase image and an optimized susceptibility weighted imaging image; A standardization processing module, used for standardizing the intensity size of the optimized phase image and the optimized susceptibility weighted imaging image to obtain a standardized phase image and a standardized susceptibility weighted imaging image; A candidate acquisition module, used for acquiring cerebral microbleeding candidates in the standardized phase image and the standardized susceptibility-weighted imaging image based on a 3D fast radial symmetric transformation algorithm; A construction module, used to construct a dual-input channel model, wherein the dual-input channel model includes a phase image input channel and a susceptibility weighted imaging image input channel, the dual-input channel model includes a residual block, and each convolution unit in each residual block includes a 3D convolution layer, an activation function layer, and a batch normalization layer; a training module, configured to input the cerebral microbleed candidate into a dual-input channel model, extract phase feature data of the cerebral microbleed candidate through a phase image input channel, extract magnetic susceptibility feature data of the cerebral microbleed candidate through a magnetic susceptibility weighted imaging image input channel, and train the dual-input channel model to obtain an optimal dual-input channel model; The identification module is used to input the cerebral microbleed candidate to be classified into the optimal dual-input channel model, and output the type corresponding to the cerebral microbleed candidate to be classified, wherein the type includes true cerebral microbleed or false positive cerebral microbleed.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 7 are implemented.

10. A non-transitory computer readable medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 7 are implemented.

Images