A deep learning-based method for detecting cerebral microbleeds

Abstract

The present invention discloses a deep learning-based method for detecting cerebral microbleeds for automated CMB detection using both SWI and phase images. Valid CMB data is selected using a 3D fast radial symmetry transform, and then false positives are reduced using a deep residual neural network. Through data preprocessing and enhancement, the model provides high sensitivity and a low number of false positives, outperforming manual and single-channel models. This method is a two-stage model for CMB detection using deep learning and 3D multi-contrast MRI data. The use of phase images along with SWI images significantly improves performance. By controlling the maximum translation amount, the model is also focused on the central region of the 3D volume to reduce interference from nearby structures. Increased test time further stabilizes the prediction and reduces uncertainty caused by doubts about valid CMB data. Compared to the SWI model, the better performance of the phase image plus SWI model is not only due to the differentiation of calcifications, but also due to the removal of false positives associated with blood vessels.

Description

Technical Field

[0001] The present invention relates to a cerebral microbleed detection method based on deep learning, belonging to the technical field of deep detection. Background Art

[0002] Cerebral microbleeds (CMBs) are tiny, round, dark lesions most commonly detected on gradient-echo magnetic resonance imaging (MRI) and are frequently seen in patients with Alzheimer's disease, stroke, and traumatic brain injury (TBI). The location of CMBs reflects the etiology, while their number indicates the risk of future intracerebral hemorrhage (ICH) and cognitive impairment. Deep CMBs in the basal ganglia or thalamus are often associated with hypertension, while the presence of lobar CMBs suggests cerebral amyloid angiopathy. CMB detection is crucial for assessing the benefits and risks of antithrombotic therapy in stroke patients. Furthermore, CMB detection plays a crucial role in the diagnosis and prognosis of TBI. It is estimated that TBI affects 50 million people worldwide each year, with over 75% of these cases being mild. The high prevalence of CMBs in TBI makes them a core component of monitoring disease progression. Therefore, the ability to accurately and reliably detect CMBs is crucial.

[0003] CMBs are typically detected using magnetic resonance imaging (MRI). Susceptibility-weighted imaging (SWI) is a widely used post-processing technique for MRI data, primarily used to detect CMBs. Its advantage is its sensitivity to susceptibility effects in blood. In SWI, a susceptibility-weighted mask is generated from a high-pass filtered phase image and multiplied with an amplitude image to form a composite image (denoted as a "SWI image"). CMBs appear as small, spherical or elliptical areas of low intensity on SWI images, but there are various types of CMB mimics, such as veins, iron deposits in the basal ganglia, calcifications, and signal voids due to poor flow compensation or tip artifacts caused by coil combination failures. Therefore, manual detection of CMBs is time-consuming and prone to error. To address this issue, automated CMB detection models are often used.

[0004] Automated CMB detection is typically performed using a two-stage process: a valid data detection stage and a classification stage to reduce false positives. While high sensitivity can be easily achieved in the first stage, performance in the second stage is often poor, leading to false positives and low precision. A major drawback of traditional machine learning models is related to the requirement for feature engineering. Due to the variability in the shape and intensity of the CMB in SWI images, designing effective and robust features is very challenging. Summary of the Invention

[0005] In response to the problems existing in the above-mentioned prior art, the present invention provides a cerebral microbleed detection method based on deep learning, thereby solving the above-mentioned technical problems.

[0006] To achieve the above objectives, the present invention adopts a technical solution: a method for detecting cerebral microbleeds based on deep learning, characterized by comprising the following steps:

[0007] Step S1: Data preprocessing and normalization,

[0008] First, the N4 algorithm in the Advanced Normalization Tools (ANT) is used to correct the bias field effect in the amplitude image.

[0009] The phase image was processed using homodyne high-pass filtering with a k-space window size of 96 × 96. To generate SWI, the phase image was first scaled by (20 ms / TE)·(3T / B0), where TE and B0 are the echo time and main field strength, respectively. Normalized phase images were first used to generate SWI data according to conventional procedures.

[0010] The problem of having the same effective susceptibility weight for data collected at different echo times and field strengths is solved by reconstructing the susceptibility map;

[0011] Amplitude, phase, SWI, and QSM images were interpolated to 0.5-mm isotropic resolution; for amplitude and SWI images, the baseline intensity (Mb) was estimated from the internal brain region with a magnetic susceptibility-weighted mask equal to 1. The intensity of the amplitude and SWI images was then normalized as follows:

[0012] Equation 1:

[0013]

[0014] Phase image Normalized by:

[0015] Equation 2:

[0016]

[0017] Where P is the original phase image, B0 is the main field strength, and TE is the echo time;

[0018] QSM Image Normalized by:

[0019] Equation 3:

[0020]

[0021] Finally, the value range of pixel intensity on the normalized image is truncated to [-1, 1]; after normalization, the intensity of the background white matter area of ​​all types of images is close to 0, while the absolute value of the CMB intensity is close to 1;

[0022] Step S2: CMB valid data detection using 3D fast radial symmetry transform (3D-FRST);

[0023] The specific steps of CMB effective data detection are as follows: first, generate CMB effective data, generate initial data map by thresholding the SWI image after 3D-FRST conversion, and use th rst Threshold, using 2th within the globus pallidus rst Threshold, th rst It varies from 0.08 to 0.15 with a step size of 0.01, and the optimal th is selected based on the sensitivity and the average number of false positives in the training and validation data. rst Threshold;

[0024] By adopting the 3D version of the FRST algorithm, the scale parameter N ranges from 1 to 4 pixels, and the symmetry strictness parameter α is set to 2;

[0025] Assuming that no edge is detected on the straight line connecting them, two isolated pixels with a distance less than 4 pixels (approximately 2 mm) are connected; where an edge is defined as a region with a gradient greater than 80% of the gradients of all regions in the brain; then, for each connected region in the valid data map, one or two valid data regions are generated depending on the shape of the region; for a spherical region, only one valid data region is generated, located at the center of the region; for an elongated region, two valid data are generated, located at both ends of the long axis of the region; therefore, the CMB valid data corresponds to the center or either end of the potential CMB; for shape analysis, this method calculates the "fractional anisotropy" (FA) of each connected region using the following equation 4:

[0026] Equation 4:

[0027]

[0028] where λx ,λ y and λ z is the eigenvalue of the connected area in the x, y, and z directions, is the average eigenvalue;

[0029] Step S3: Reduce false positives through deep learning models, using both SWI and high-pass filtered phase images based on a deep residual neural network; this method uses phase and SWI images to train models, with an input scale of 16×16×16; for these models, the output size of the first convolutional layer is 16×16, and the first average pooling layer is discarded; the appropriate model is selected by comparing different input channels for verification.

[0030] Furthermore, in step S1, the non-rigid registration algorithm in the advanced normalization tool registers the corrected amplitude image to the MNI-152 template (1 mm isotropic resolution).

[0031] Furthermore, in step S1, the method for reconstructing the magnetic susceptibility map is specifically as follows:

[0032] First, use the Brain Extraction Tool (BET) to generate a brain mask;

[0033] Next, the Laplace expansion and the Sophisticated Harmonic Artifact Reduction for Phase (SHARP) algorithm of the phase data are used to reduce the residual background field and aliasing artifacts in the filtered phase image.

[0034] Finally, the morphology enabled dipole inversion (MEDI) algorithm with a regularization parameter λ of 300 was used to create the magnetic susceptibility map.

[0035] Furthermore, in step S3, the deep model architecture method is specifically as follows: the residual block is implemented in a complete pre-activation manner, and the rectified linear unit (ReLU) and batch normalization (BatchNorm) layers are located before the 3D convolution layer; the kernel size of all 3D convolution layers is 3×3×3; when the output size of the layer changes, 1×1×1 convolution is used to form a residual connection; when the spatial dimension is reduced, the convolution (including 1×1×1 convolution) uses a stride of 2, and other convolutions use a stride of 1; after the residual block, global average pooling is performed, followed by two fully-connected (FC) layers and a normalized exponential function (softmax) layer; except for the input layer, the sizes of all layers are the same for different types of inputs, and the number of trainable parameters is also the same.

[0036] Furthermore, the input channel combination can be single channel: amplitude (M), phase (P), SWI (S), QSM (Q); dual channel: MP, MQ, PS, SQ; triple channel: MPS, MPQ, MSQ, PSQ; quad channel: MPSQ.

[0037] The beneficial effects of the present invention are:

[0038] This method is a two-stage model for CMB detection using deep learning and 3D multi-contrast MRI data. Compared with the model using SWI images alone, using phase images together with SWI images significantly improved the performance. On the test data, the best model achieved outstanding performance.

[0039] By controlling the maximum translation, the model is also focused on the central region of the 3D volume to reduce interference from nearby structures; increasing the test time further stabilizes the prediction and reduces the uncertainty caused by doubts about the valid CMB data.

[0040] The better performance of the phase-plus-SWI model compared with the SWI model is not only due to the differentiation of calcifications but also due to the removal of false positives related to blood vessels, such as arteries or veins with problematic flow compensation. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 is the 3D-FRST threshold th of the embodiment of the present invention rst Schematic diagram of the impact of sensitivity and the average number of false positives (FPavg) on ​​CMB valid data detection; (a) Sensitivity and FPavg. (b) FPavg and th rst (c) Sensitivity and FP avg ;

[0042] Figure 2 Schematic diagram of the architecture of a deep learning model for reducing CMB false alarms according to an embodiment of the present invention;

[0043] Figure 3 Schematic diagram of the performance of the best single model on test data according to an embodiment of the present invention; (a) and (b) are models with three or four input channels; (c) and (d) are models with two input channels; (e) and (f) are models with a single input channel.

[0044] Figure 4 Schematic diagram showing the effects of input scale, model averaging, and CMB size on the performance of the best single model according to an embodiment of the present invention; (a) and (b) show the effects of input scale; (c) and (d) show the effects of model averaging; (e) and (f) show the performance of the best single model when excluding small lesions ≤ 2 pixels.

[0045] Figure 5 Schematic diagram of phase and SWI images used in an embodiment of the present invention to distinguish calcification from CMB; (a) amplitude image; (b) phase image; (c) SWI image; (d) quantitative magnetic susceptibility map; (e) SWI image after 3D-FRST transformation; (f) SWI image with valid CMB data detected; (g) The model observing SWI images alone failed to eliminate false positives caused by calcification (arrow). (h) The model combining phase and SWI images successfully eliminated one false positive.

[0046] Figure 6 Schematic diagram of the method of the present invention. DETAILED DESCRIPTION

[0047] In order to make the purpose, technical solutions and advantages of the present invention more clear, the present invention is further described in detail below with reference to the accompanying drawings and embodiments. However, it should be understood that the specific embodiments described herein are only used to illustrate the present invention and are not intended to limit the scope of the present invention.

[0048] Unless otherwise defined, all technical and scientific terms used herein have the same meanings as those commonly understood by those skilled in the art to which the present invention pertains. The terms used in the specification of the present invention herein are only for the purpose of describing specific embodiments and are not intended to limit the present invention.

[0049] Reference Figure 1This paper proposes a two-stage CMB detection method, using a 3D-FRST algorithm to detect CMB valid data from SWI images and a deep learning model to reduce false positives based on phase and SWI images. Data preprocessing and normalization reduce variations caused by imaging parameters, standardize the intensity scales of different image types, and maximize the contrast between CMBs and background tissue. These procedures are fundamental to fully exploiting the information in multi-contrast 3D MRI data.

[0050] 1. Data preprocessing and normalization

[0051] Data preprocessing and normalization were performed to reduce variations caused by imaging parameters. First, the N4 algorithm in the Advanced Normalization Tools (ANT) was used to correct for bias field effects in the magnitude image. The nonrigid registration algorithm in ANT was used to register the corrected magnitude image to the MNI-152 template (1 mm isotropic resolution). The template was used to localize the globus pallidus to avoid detecting iron deposits.

[0052] Phase images were processed using homodyne high-pass filtering with a k-space window size of 96 × 96. To generate SWI, the phase images were first scaled by (20 ms / TE)·(3T / B0), where TE and B0 are the echo time and main field strength, respectively. SWI data were generated using the normalized phase images according to conventional procedures.

[0053] This results in data collected at different echo times and field strengths having the same effective susceptibility weight. The susceptibility map is reconstructed using the following steps.

[0054] First, a brain mask was generated using the Brain Extraction Tool (BET). Next, residual background fields and aliasing artifacts in the filtered phase image were reduced using Laplace expansion and the Sophisticated Harmonic Artifact Reduction for Phase (SHARP) algorithm. Finally, magnetic susceptibility maps were created using the morphology-enabled dipole inversion (MEDI) algorithm with a regularization parameter λ of 300.

[0055] Amplitude, phase, SWI, and QSM images were interpolated to 0.5 mm isotropic resolution. For amplitude and SWI images, the baseline intensity (Mb) of the image was estimated based on the internal brain region with a magnetic susceptibility weighting mask equal to 1. The intensity of the amplitude and SWI images was then normalized by:

[0056] Equation 1:

[0057]

[0058] Phase image Normalized by:

[0059] Equation 2:

[0060]

[0061] Where P is the original phase image, B0 is the main field strength, and TE is the echo time.

[0062] QSM Image Normalized by:

[0063] Equation 3:

[0064]

[0065] Finally, the pixel intensity value range on the normalized image is truncated to [-1, 1]. After normalization, the background white matter intensity of all image types is close to 0, while the absolute value of the CMB intensity is close to 1.

[0066] 2. CMB effective data detection and 3D fast radial symmetry transform (3D-FRST)

[0067] This method uses the 3D version of the FRST algorithm, with the scale parameter N ranging from 1 to 4 pixels and the symmetry strictness parameter α set to 2.

[0068] CMB valid data were generated using the following steps. First, an initial data map was generated by thresholding the 3D-FRST converted SWI image. rst Threshold, using 2th within the globus pallidus rst The sensitivity of CMB effective data detection depends on the threshold value to avoid detecting increased iron deposition levels. rst To determine the optimal value, th rst It was varied between 0.08 and 0.15 with a step size of 0.01. The optimal thrst was chosen based on the sensitivity and the average number of false positives in the training and validation data.

[0069] Due to the noise and variations in CMB shape, there are a large number of isolated pixels, which causes unnecessary calculations in the false alarm reduction stage. Therefore, if no edge is detected on the straight line connecting them, two isolated pixels with a distance less than 4 pixels (approximately 2 mm) are connected. In this method, an edge is defined as a region with a gradient greater than 80% of the gradients of all regions within the brain. Next, for each connected region in the valid data map, one or two valid data regions are generated depending on the shape of the region. For spherical regions, such as conventional CMBs, only one valid data region is generated, located at the center of the region; for elongated regions, such as damaged veins, two valid data are generated, located at both ends of the long axis of the region. The reason is to provide enough data for CMBs of different shapes. Therefore, the valid CMB data in this method corresponds to the center or any end of the potential CMB. For shape analysis, this method calculates the "fractional anisotropy" (FA) of each connected region:

[0070] Equation 4:

[0071]

[0072] where λ x ,λ y and λ z are the eigenvalues ​​of the connected region in the x, y, and z directions, is the average eigenvalue. These eigenvalues ​​represent the shape of connected regions rather than the usual diffusion properties. For a sphere, FA is 0, and for a line, FA is 1. In this method, elongated regions are empirically defined as regions with a major axis length greater than 6 mm and an FA greater than 0.85 (in this method, a cuboid with an aspect ratio of 3:2:1 is considered spherical). Finally, 3D volumes with a matrix size of 32×32×32 are cropped from the amplitude, phase, SWI, and QSM images, with the CMB valid data located in the center. These 3D volumes will be classified by the deep learning model as true CMBs or false positives.

[0073] Figure 1 for th rst The impact of sensitivity and FPavg on the CMB valid data detection step was evaluated on the training and validation data. Based on sensitivity and FPavg, an optimal threshold of 0.09 was selected. This threshold achieved a sensitivity of 97.9% and an FPavg of 312.5. On the test data, the sensitivity of CMB valid data detection was 99.4% and the FPavg was 276.8.

[0074] 3. Reduce false positives through deep learning

[0075] The architecture of the deep learning model is as follows Figure 2As shown. The residual block is implemented in a full pre-activation manner, with rectified linear unit (ReLU) and batch normalization (BatchNorm) layers placed before the 3D convolutional layers. The kernel size of all 3D convolutional layers is 3×3×3. When the output size of the layer changes, 1×1×1 convolutions are used to form residual connections. When the spatial dimension is reduced, those convolutions (including 1×1×1 convolutions) use a stride of 2, and other convolutions use a stride of 1. After the residual block, global average pooling is performed, followed by two fully-connected (FC) layers and a normalized exponential function (softmax) layer. This method tests different combinations of input channels, including

[0076] (1) Single channel: amplitude (M), phase (P), SWI (S), QSM (Q);

[0077] (2) Dual channel: MP, MQ, PS, SQ;

[0078] (3) Three channels: MPS, MPQ, MSQ, PSQ;

[0079] (4) Four channels: MPSQ.

[0080] Except for the input layer, the sizes of all layers are the same for different types of inputs, as is the number of trainable parameters. Figure 2 The method uses phase and SWI images to train models with input size 3 of 16×16×16. For these models, the output size of the first convolutional layer is 16×16, the first average pooling layer is discarded, and the output size of all other layers is the same as Figure 2 Same as shown in .

[0081] Figure 3 Performance of the best single model with different types of input. The sensitivity of the CMB valid data detection stage has been taken into account, and the reported metrics reflect the performance of the entire two-stage CMB detection framework. Among the models with three or four input channels, the model using amplitude, phase and QSM (denoted as MPQ model) has the largest AUC-PR: 0.91 [0.85, 0.94], with lower and upper confidence limits provided in parentheses. In comparison, the model using all four types of images (denoted as MPSQ model) has an AUC-PR of 0.87 [0.82, 0.91]. As Figure 3 c and Figure 3As shown in d, the model using the combination of phase and SWI images (denoted as PS model, AUC-PR: 0.92 [0.88, 0.95]) outperformed all other types of models and achieved similar performance to the most experienced SWI data processor. Without test time augmentation (TTA), the performance of the PS model ( Figure 3 c and Figure 3 The dashed line in d, AUC-PR: 0.91 [0.87, 0.94]) was significantly lower (p = 0.013). Figure 3 e and Figure 3 Among the models with single-channel input (Figure 5), 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 magnitude (AUC-PR: 0.83 [0.78, 0.88]), reflecting strengths across image types. However, no significant differences were found 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 PS model showed a significant difference from the reference SWI model (p = 0.024). The PS model also showed a significant difference from the MPSQ model (p = 0.015). No significant differences were observed between the other model pairs.

[0082] Figure 4 a and Figure 4 Figure b shows that the performance of the PS model with 32×32×32 input scale is significantly better than that of the model with 16×16×16 input scale (p=0.0002). Figure 4 c and Figure 4 As shown in d, the performance of the model is not greatly improved by simple averaging. The sensitivity of scorer 1 is relatively low compared with other scorers, partly due to the lack of some small lesions, such as Figure 4 e and Figure 4 When small CMBs with a volume ≤ 2 pixels were excluded (a total of 15 CMBs, whose sizes were estimated from thresholded 3D-FRST images), the sensitivity of Rater 1 improved to 84.3% with an accuracy of 84.3%. Meanwhile, both the PS model and the MPQ model outperformed Raters 1 and 2. The other models still performed worse than the other Raters.

[0083] like Figure 5As shown by the arrows in the figure, the false positives caused by calcifications were not eliminated using the SWI model, but were successfully eliminated using the PS model. As predicted by the SWI model, the probability of this calcification being a true CMB was 60.0%, but the PS model reduced this probability to 19.3%. In the test data, a total of three calcifications were misclassified as CMBs by the SWI model, but all were eliminated by the PS model.

[0084] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions or improvements made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for detecting cerebral microbleeds based on deep learning, characterized in that: The following steps are involved: Step S1: Data preprocessing and normalization, First, the N4 algorithm in the advanced normalization tool ANT is used to correct the bias field effect in the amplitude image; The phase image was processed using homodyne high-pass filtering with a k-space window size of 96 × 96. To generate SWI, the phase image was first scaled by (20 ms / TE)·(3T / B0), where TE and B0 are the echo time and main field strength, respectively. Normalized phase images were first used to generate SWI data according to conventional procedures. The problem of having the same effective susceptibility weight for data collected at different echo times and field strengths is solved by reconstructing the susceptibility map; Amplitude, phase, SWI, and QSM images were interpolated to 0.5 mm isotropic resolution; for amplitude and SWI images, the baseline intensity M b It was estimated from the internal brain regions with a susceptibility-weighted mask equal to 1, and the intensity of the amplitude and SWI images was then normalized by: Equation 1: in, represents the normalized amplitude or SWI image intensity; M represents the original amplitude or SWI image intensity; M b Indicates a strong baseline; Phase image Normalized by: Equation 2: Where P is the original phase image, B0 is the main field strength, and TE is the echo time; QSM Image Normalized by: Equation 3: in, represents the normalized QSM image intensity; Q represents the original QSM image intensity; Finally, the value range of pixel intensity on the normalized image is truncated to [-1, 1]; after normalization, the intensity of the background white matter area of ​​all types of images is close to 0, while the absolute value of the CMB intensity of cerebral microbleeds is close to 1; Step S2: Detection of CMB valid data using 3D fast radial symmetry transform (3D-FRST); The specific steps of CMB effective data detection are as follows: first, generate CMB effective data, generate initial data map by thresholding the SWI image after 3D-FRST conversion, and use th rst Threshold, using 2th within the globus pallidus rst Threshold, th rst It varies from 0.08 to 0.15 with a step size of 0.01, and the optimal th is selected based on the sensitivity and the average number of false positives in the training and validation data. rst Threshold; By adopting the 3D version of the FRST algorithm, the scale parameter N ranges from 1 to 4 pixels, and the symmetry strictness parameter α is set to 2; Assuming that no edge is detected on the straight line connecting them, two isolated pixels with a distance less than 4 pixels are connected; an edge is defined as a region with a gradient greater than 80% of the gradients of all regions within the brain. Subsequently, for each connected region in the valid data map, one or two valid data regions are generated depending on the shape of the region. For spherical regions, only one valid data region is generated, located at the center of the region; for elongated regions, two valid data are generated, located at both ends of the long axis of the region. Therefore, the CMB valid data corresponds to the center or either end of the potential CMB. For shape analysis, the "fractional anisotropy" (FA) of each connected region is calculated using the following equation 4: Equation 4: where λ x ,λ y and λ z are the eigenvalues ​​of the connected region in the x, y, and z directions, is the average eigenvalue; Step S3: Reduce false positives through deep learning models, using both SWI and high-pass filtered phase images based on a deep residual neural network; this method uses phase and SWI images to train models with an input scale of 16×16×16; for these models, the output size of the first convolutional layer is 16×16, and the first average pooling layer is discarded; the best single model is obtained by comparing different input channels for verification.

2. The method for detecting cerebral microbleeds based on deep learning according to claim 1, characterized in that: In step S1, the non-rigid registration algorithm in the advanced normalization tool registers the corrected amplitude image to the MNI-152 template.

3. The method for detecting cerebral microbleeds based on deep learning according to claim 1, characterized in that: In step S1, the method for reconstructing the magnetic susceptibility map is specifically as follows: First, a brain mask was generated using the brain extraction tool BET; Next, the Laplace expansion and the complex harmonic artifact reduction SHARP algorithm of the phase data are used to reduce the residual background field and aliasing artifacts in the filtered phase image; Finally, the magnetic susceptibility maps were created using the morphology-enabled dipole inversion MEDI algorithm with a regularization parameter λ of 300.

4. The method for detecting cerebral microbleeds based on deep learning according to claim 1, characterized in that: In step S3, the deep model architecture method is specifically as follows: the residual block is implemented in a complete pre-activation manner, and the rectified linear unit ReLU and batch normalization BatchNorm layer are located before the 3D convolution layer; the kernel size of all 3D convolution layers is 3×3×3; when the output size of the layer changes, 1×1×1 convolution is used to form a residual connection; when the spatial dimension is reduced, the convolution uses a stride of 2, and other convolutions use a stride of 1; after the residual block, global average pooling is performed, followed by two fully connected FC layers and a normalized exponential function layer; except for the input layer, the sizes of all layers are the same for different types of inputs, and the number of trainable parameters is also the same.

5. The method for detecting cerebral microbleeds based on deep learning according to claim 4, characterized in that: The input channel combination can be single channel: amplitude M, phase P, SWI, QSM; dual channels: MP, MQ, PS, SQ; triple channels: MPS, MPQ, MSQ, PSQ; quad channels: MPSQ.

Images