Automated Detection of Cerebral Microbleeds on Two-dimensional Gradient-recalled Echo T2* Weighted Images Using a Morphology Filter Bank and Convolutional Neural Network

Abstract

Purpose

We present a novel algorithm for the automated detection of cerebral microbleeds (CMBs) on 2D gradient-recalled echo T2* weighted images (T2*WIs). This approach combines a morphology filter bank with a convolutional neural network (CNN) to improve the efficiency of CMB detection. A technical evaluation was performed to ascertain the algorithm’s accuracy.

Methods

In this retrospective study, 60 patients with CMBs on T2*WIs were included. The gold standard was set by three neuroradiologists based on the Microbleed Anatomic Rating Scale guidelines. Images with CMBs were extracted from the training dataset comprising 30 cases using a morphology filter bank, and false positives (FPs) were removed based on the threshold of size and signal intensity. The extracted images were used to train the CNN (Vgg16). To determine the effectiveness of the morphology filter bank, the outcomes of the following two methods for detecting CMBs from the 30-case test dataset were compared: (a) employing the morphology filter bank and additional FP removal and (b) comprehensive detection without filters. The trained CNN processed both sets of initial CMB candidates, and the final CMB candidates were compared with the gold standard. The sensitivity and FPs per patient of both methods were compared.

Results

After CNN processing, the morphology-filter-bank-based method had a 95.0% sensitivity with 4.37 FPs per patient. In contrast, the comprehensive method had a 97.5% sensitivity with 25.87 FPs per patient.

Conclusion

Through effective CMB candidate refinement with a morphology filter bank and FP removal with a CNN, we achieved a high CMB detection rate and low FP count. Combining a CNN and morphology filter bank may facilitate the accurate automated detection of CMBs on T2*WIs.

Introduction

Cerebral microbleeds (CMBs) are small foci of chronic blood products in the brain parenchyma. Pathologically, they refer to lesions in which hemosiderin-containing macrophages accumulate around blood vessels.¹ CMBs can be detected in older adults without symptoms,² but they are often detected in patients with various diseases, including stroke,³ cerebral amyloid angiopathy (CAA),⁴ Moyamoya disease,⁵ and genetic diseases such as cerebral autosomal dominant arteriopathy with subcortical infarcts and leukoencephalopathy.⁶ They are also observed in traumatic or radiation-induced brain injuries.^7,8 Amyloid-related imaging abnormalities have been reported in patients with Alzheimer’s disease (AD). Amyloid-related imaging abnormalities with hemosiderin deposition represent newly emerging CMBs, particularly after amyloid-modifying therapies. Detecting them is crucial for making informed treatment decisions.⁹ CMBs are also associated with several risks, such as increased risks of ischemic stroke and intracerebral hemorrhage.^10,11 A high microbleed count is also associated with cognitive deterioration and dementia.¹¹ Therefore, detecting CMBs is important for predicting future risks of healthy older adults and patients and prognosis and selecting treatment options for amyloid-modifying therapies.

CMBs are detected using MRI. On gradient-recalled echo (GRE) T2* weighted imaging (T2*WI) or susceptibility-weighted imaging (SWI), they appear as round or ovoid hypointense lesions with clear margins, their sizes range from 2 to 10 mm, and a “blooming” effect may accompany them.¹ SWI typically employs 3D high spatial resolution. In contrast, conventional GRE T2*WI employs 2D spatial resolution with a thickness of approximately 5 mm. Therefore, 3D SWI is considered better for detecting CMBs than 2D GRE T2*WI.^12,13 In contrast, T2*WI is less time-consuming than SWI, does not require post-processing of phase images, and is more accessible than SWI in many facilities. Notwithstanding, T2*WI and SWI may show false-positive findings; veins, artifacts, and microcalcifications, among others,¹⁴ may also be observed as low-signal regions, just like true CMBs on T2*WI and SWI. Therefore, accurate detection and classification require knowledge and experience, as interpretations are affected by the level of experience of the observer.¹⁵ On the other hand, manual detection of true CMBs is time-consuming and particularly error-prone, especially when T2*WI images are reviewed by a single observer.¹⁶ Therefore, there is a need for systematic and automated detection of CMBs with high accuracy and reproducibility.

Shape filters are occasionally employed to extract lesions and structures from medical images. Morphology filter banks, a class of shape filters, have been used to extract microcalcifications from mammography images, delineate blood vessels in retinal fundus images, and remove normal blood vessels from chest CT images during reconstruction.¹⁷ Recent studies have reported automated CMB extraction using a combination of convolutional neural network (CNN) and shape filters.^18,19 However, these studies primarily employed SWI, and to the best of our knowledge, there is a lack of reports on automated CMB extraction from T2*WI.

We hypothesized that utilizing morphology filter banks in conjunction with a CNN would be effective for the automatic detection of CMBs in T2*WI, which is a more accessible sequence than SWI. In this study, we introduce this novel algorithm. Additionally, to validate the efficacy of the morphology filter bank, we compared the results of the combination of morphology filter banks and a CNN with those of a comprehensive, unfiltered extraction method combined with a CNN.

Materials and Methods

This retrospective, single-center study was approved by our institutional review board, and they waived the requirement for written informed consent.

Data sets

Patient data

All the data were collected at our institution. Study cohorts were included using the following criteria: (a) MRI examinations, including T2*WI, were conducted between August 2014 and October 2021, (b) one or more CMBs on T2*WI documented by radiologists, (c) patients aged 7 years or older, and (d) T2*WI scans free from significant artifacts. Sixty-three patients met the inclusion criteria. Among them, we excluded 3 who had a clinical history of traumatic diffuse axonal injury according to the “Recommended Criteria for CMB Identification” proposed by Greenberg et al.¹ CMBs caused by traumatic diffuse axonal injury often do not have the typical round shape, and they were considered confusing for the readers and CMBs’ extraction algorithms. Sixty patients were eligible for inclusion in the study. Their ages ranged from 15 to 87 years, comprising 30 males and 30 females. Details about the primary diseases of the participants are provided in Supplementary Table 1.

MRI technique

Sixty T2*WI data points were acquired using seven different machines from four companies with field strengths of 1.5T and 3T. Details about these systems and their respective parameters are presented in Supplementary Table 2.

Determination of gold standard

The CMBs were rated by visual inspection by two neuroradiologists with 10 and 6 years of experience, respectively, according to the Microbleed Anatomic Rating Scale (MARS) guidelines.¹⁵ In the MARS guideline, CMBs are classified as “definite” or “possible” based on how closely they resemble true CMBs. Definite CMBs were rounded or circular, well-defined hypointense lesions in the brain parenchyma with clear margins and sizes ranging from 2 to 10 mm on T2*WI. Possible CMBs were lesions that were not strictly rounded or circular, less well-defined, and less hypointense than definite CMBs.

Two neuroradiologists labeled the definite and possible CMBs for all slices of the T2*WI sequence for the 60 cases. If the decisions of the two neuroradiologists were contrasting, a third neuroradiologist with 25 years of experience would make the final decision. Labeling of the 60 cases by the neuroradiologists resulted in the detection of 341 definite and 3933 possible CMBs. Thereafter, the data of the 60 participants were sorted based on the number of definite CMBs in descending order and split into training and test sets in a 1:1 ratio to allow the number of definite CMBs to be approximately equal. The training dataset of 30 cases contained 181 definite and 1821 possible CMBs, while the test dataset of 30 cases contained 160 definite and 2112 possible CMBs.

Proposed method

Smoothing and brain mask

To reduce the effect of noise, we applied a 3 × 3 box filter and performed smoothing on the T2*WIs. To extract only the brain parenchyma where the intracerebral hemorrhage occurred, we used the following method and removed structures outside the brain parenchyma, such as the skull and subcutaneous fat. From the image histogram, we extracted only the signal-present area from the image using discriminant analysis. The brain parenchymal region was extracted using a 2D region-growing process with the center pixel of the image as the seed point (i.e., the calculation of the brain extraction mask) in each slice. The brain extraction mask was multiplied by the smoothed T2*WIs to obtain the brain-extracted image.

Morphology filter bank

The morphology filter bank applies a set of top-hat transformations with an opening operator that calculates a higher signal intensity area that cannot be penetrated by filling with certain structural elements from the lower signal intensity.¹⁷ This morphology filter bank was modified to extract low signal intensity areas with granular and linear structures in T2*WIs in this study. As shown in Fig. 1, there are eight types of linear structure elements with radial line structures spaced 22.5° apart and one type of circular structure element with a shape similar to a circle. Five kernel sizes (11 × 11, 9 × 9, 7 × 7, 5 × 5, and 3 × 3 pixels) were prepared.

Fig. 1.

Structural elements of the morphology filter bank.

These sizes were determined based on the following factors. The pixel size was 0.47 mm, and the line lengths and circular element structure diameters for the kernel sizes of 11 × 11, 9 × 9, and 7 × 7 pixels used in the granular morphology filter were about 5.2, 4.2, and 3.3 mm, respectively. Thus, the granular morphology filter could detect patterns with the largest lengths not exceeding 5.2 mm, including CMBs and vessels. The length of the line and the diameter of the circular element structure with 5 × 5 and 3 × 3 pixels used in the linear morphology filter were 2.4 and 1.4 mm, respectively. Thus, the linear morphology filter can detect linear patterns with lengths of ³ 2.4 mm and widths of < 2.4 (i.e., £1.9) mm or lengths of ³ 1.4 mm and widths of < 1.4 (i.e., £0.9) mm; these correspond to the partial length and diameter of brain surface vessels. Furthermore, it prevents the detection of granular patterns with widths of < 2 mm, which meets the MARS criteria. Accordingly, CMBs meeting the MARS criteria were maintained; however, vessels with these linear patterns were removed by subtracting the linear patterns from the granular patterns.

A top-hat transformation with a closing operator was applied to detect low-signal intensity areas with granular or linear structures in T2*WIs and obtain an area with high signal intensity (Fig. 2).

Fig. 2.

Granular and linear structures extracted from 2D-GRE T2*WI using morphology filter banks. (a) Smoothed, brain-extracted GRE T2*WI. (b) Granular structure extraction image. Granular structures extracted by the granular morphology filter are shown in high signal. (c) Linear structure extraction image. Linear structures extracted by the linear morphology filter are shown in high signal. GRE T2*WI, gradient-recalled echo T2* weighted imaging.

The image obtained after extracting granular structures using the granular morphology filter for the CMB candidate extraction is referred to as the “granular structure extraction image,” while the image obtained after extracting linear structures using the linear morphology filter to remove vascular structures is referred to as the “linear structure extraction image.”

Granular structure extraction

Granular structures with low signal intensity on T2*WIs were extracted from the granular structure extraction images via threshold processing. The granular structure extraction image is denoted as I_G, and the binary image M_a is calculated using Equation (1).

$${\mathrm{M}}_a\left(\mathrm{r}\right)=\left\{\begin{array}{cc}1& \mathrm{i}\mathrm{f}{\mathrm{I}}_{\mathrm{G}}\left(\mathrm{r}\right)>{\mathrm{S}}_{\mathrm{m}\mathrm{e}\mathrm{a}\mathrm{n}}{\mathrm{T}}_{\mathrm{a}}\\ 0& \mathrm{i}\mathrm{f}{\mathrm{I}}_{\mathrm{G}}\left(\mathrm{r}\right)\le {\mathrm{S}}_{\mathrm{m}\mathrm{e}\mathrm{a}\mathrm{n}}{\mathrm{T}}_{\mathrm{a}}\end{array}\right.$$ (1)

where “r” represents the image coordinates, S_mean represents the average signal value within the brain, and T_a represents the threshold value. A threshold value (T_a) of 0.1 was used.

Linear structure removal

We excluded the linear structures in the regions extracted as granular structures. After dividing the granular structure extraction image (I_G) by the linear structure extraction image (I_L), the pixels with the maximum values in the extracted region (cluster) below the threshold T_b were excluded. Let “v” be the range of the clusters of the binary image M_a created by granular structure extraction. Equation (2) was used to calculate the binary image M_b.

$${\mathrm{M}}_{\mathrm{b}}\left(\mathrm{v}\right)=\left\{\begin{array}{cc}1& \mathrm{i}\mathrm{f}\mathrm{m}\mathrm{a}\mathrm{x}\left[{\mathrm{M}}_a\left(\mathrm{v}\right){\mathrm{I}}_{\mathrm{G}}\left(\mathrm{v}\right)/{\mathrm{I}}_{\mathrm{L}}\left(\mathrm{v}\right)\right]>{\mathrm{T}}_{\mathrm{b}}\\ 0& \mathrm{i}\mathrm{f}\mathrm{m}\mathrm{a}\mathrm{x}\left[{\mathrm{M}}_a\left(\mathrm{v}\right){\mathrm{I}}_{\mathrm{G}}\left(\mathrm{v}\right)/{\mathrm{I}}_{\mathrm{L}}\left(\mathrm{v}\right)\right]\le {\mathrm{T}}_{\mathrm{b}}\end{array}\right.$$ (2)

where max[X] is an operator that calculates the maximum value of X. A threshold value (T_b) of 3 was used.

Figure 2 shows the extracted low-signal-intensity granular and linear structures in the T2*WI as bright regions.

Removal of false positives

Areas with extremely low signals not meeting the MARS criteria, as well as regions with a subtle low signal intensity not distinct enough to be classified as “definite,” were considered potential false positives (FPs). Two operations were performed to remove the FPs.

– Small granular structure removal

Clusters with pixel values lower than the threshold T_c were excluded from the analysis. Let M_b(v) be any cluster in binary image M_b. The binary image M_c was calculated using Equation (3).

$${\mathrm{M}}_{\mathrm{c}}\left(\mathrm{v}\right)=\left\{\begin{array}{cc}1& \mathrm{i}\mathrm{f}\sum _{\mathrm{v}}{\mathrm{M}}_{\mathrm{b}}\left(\mathrm{v}\right)>{\mathrm{T}}_{\mathrm{c}}\\ 0& \mathrm{i}\mathrm{f}\sum _{\mathrm{v}}{\mathrm{M}}_{\mathrm{b}}\left(\mathrm{v}\right)\le {\mathrm{T}}_{\mathrm{c}}\end{array}\right.$$ (3)

The threshold value (T_c) was set to 10.

– Subtle low signal intensity structure removal

If the minimum value in a cluster was greater than the user-defined coefficient multiplied by the average value of S_mean in the mask, the cluster was excluded. Let I_S be the T2*WIs after smoothing. The binary image M_d was calculated using Equation (4).

$${\mathrm{M}}_{\mathrm{d}}\left(\mathrm{v}\right)=\left\{\begin{array}{cc}1& \mathrm{i}\mathrm{f}\mathrm{m}\mathrm{i}\mathrm{n}\left[{\mathrm{M}}_{\mathrm{c}}\left(\mathrm{v}\right){\mathrm{I}}_{\mathrm{S}}\left(\mathrm{v}\right)\right]\le {\mathrm{S}}_{\mathrm{m}\mathrm{e}\mathrm{a}\mathrm{n}}{\mathrm{T}}_{\mathrm{d}}\\ 0& \mathrm{i}\mathrm{f}\mathrm{m}\mathrm{i}\mathrm{n}\left[{\mathrm{M}}_{\mathrm{c}}\left(\mathrm{v}\right){\mathrm{I}}_{\mathrm{S}}\left(\mathrm{v}\right)\right]>{\mathrm{S}}_{\mathrm{m}\mathrm{e}\mathrm{a}\mathrm{n}}{\mathrm{T}}_{\mathrm{d}}\end{array}\right.$$ (4)

where min[X] is an operator that calculates the minimum value within X, and T_d is the threshold value. The threshold value (T_d) was set to 0.8.

CNN

In this study, we utilized Vgg16,²⁰ a 16-layer CNN model that was trained on a large-scale image dataset called ImageNet. We employed transfer learning; we used a pre-trained network and trained the layers used for classification. The learning rate of the final layer was set to 20. The number of epochs was 20, the mini-batch size was 64, and the initial learning rate was set to 10^-4. All extracted CMB candidates were compiled into 16 × 16-pixel images and compared with the gold standard set by a neuroradiologist. The images with CMBs that matched the definite lesions identified by the neuroradiologists were considered correct images, while those with CMBs that did not match the definite or possible lesions were considered incorrect images. To mitigate the imbalance resulting from more incorrect than correct images, the number of correct images was doubled by augmenting them with left-right flips. The number of incorrect images was randomly reduced to match the number of correct images. We interpolated the 16 × 16 images into 224 × 224 images using linear interpolation and performed RGB conversion to train the CNN with these samples. A flowchart of the training data preparation is shown in Supplementary Fig. 1.

Evaluation

Detection of CMB candidates from test data

Proposed method

Smoothing and brain masking were used to analyze the data. Subsequently, we used a morphology filter bank to extract the granular structures and remove their linear features. FP removal was performed using two methods: small granular structure removal and subtle low-signal-intensity structure removal. The initially extracted CMB candidates were cropped into patches of 16 × 16 pixels. These patches were further expanded to 224 × 224 pixels using linear interpolation, followed by RGB conversion. Subsequently, a trained CNN was employed for processing, resulting in the detection of the final CMB candidates.

We investigated how the granular structures extracted by the granular morphology filter were refined by the removal of linear structures, two FPs, and a CNN.

Comprehensive method without morphology filter bank

Herein, we describe a comprehensive CMB candidate detection method that does not use a morphology filter bank. The smoothed T2*WIs multiplied by a brain extraction mask were divided into small regions (patches) with 16 × 16 pixels. Patches with pixel values of 0 were excluded. The coordinates of the minimum luminance for each patch were calculated, and patches of 16 × 16 pixels were cropped, centered on the minimum luminance, to form the initial CMB candidates. Figure 3 illustrates the comprehensive CMB candidate detection process without a morphology filter bank. These patches were enlarged to 224 × 224 pixels using linear interpolation, and RGB conversion was performed. All initially extracted CMB candidate patches obtained through this comprehensive approach were processed by a trained CNN, resulting in the final detection of CMB candidates.

Fig. 3.

Comprehensive method without morphology filter bank. (a) Images are divided into small regions (patches). (b) The coordinates of the minimum pixels within each patch are calculated. (c) A new patch centered on the coordinates of the minimum pixel is generated. CMB: cerebral microbleed.

Statistical analysis

We compared the numbers of initial CMB candidates and the final CMB candidates after CNN processing with the morphology-filter-bank-based approach and the comprehensive approach. The process for detecting CMB candidates from a test dataset using the two detection methods is shown in Supplementary Fig. 2.

Lesions identified as final CMB candidates by each method were cross-checked against the gold standard selected by the neuroradiologists. When the final CMB candidate matched a lesion that was considered definite by the neuroradiologists, the designation was a true positive (TP).

If the final CMB candidate matched the area that the neuroradiologist did not classify as definite or possible CMB, the designation was an FP. According to Gregoire et al,¹⁵ interrater reliability is significantly lower for possible than for definite CMBs. For this reason, we did not count the TPs or FPs for areas that the neuroradiologists considered as possible CMBs.

The performance of each detection method was assessed based on sensitivity to the detection of definite CMBs and the number of FPs per patient using a cutoff value of 0.5. We compared the sensitivities and the number of FPs per patient for the two detection methods. The sensitivity for definite CMBs was calculated by dividing the number of TPs for each method by the total number of definite CMBs. Free-response Operating Characteristic (FROC) curves were created for both detection methods. In detection tasks with multiple lesions per case, the FROC is an established method for evaluating detection performance.²¹ It compares the detection sensitivity of a model with the number of FP predictions over a continuous scale of operating points.

Results

Initial CMB detection by the morphology bank-based method

Figure 4 illustrates the CMB candidate detection process using the morphology-filter-bank-based approach. The candidates extracted at each stage are marked on the T2*WIs with green circles based on their image coordinates. Initially, several CMB candidates were extracted using a granular morphology filter (Fig. 4g). They decreased after removing the linear structures presumed to be vessels (Fig. 4h). Further reduction was achieved by removing small granular structures and subtle low-signal intensity structures, leading to a decrease in the number of CMB candidates (Fig. 4i, j).

Fig. 4.

(a) Smoothed, brain-extracted GRE T2*WI. (b) Extracted structures from granular structure extraction image using granular morphology filter. (c) Image after excluding linear structures using linear morphology filter. (d) Image after small granular structure removal. (e) Image after subtle low signal intensity structure removal. (f) Gold standard labeled by neuroradiologists. Red circles indicate “definite,” and yellow circles indicate “possible.” (g)–(j) Structures extracted from (b)–(e) on GRE T2*WIs. (k) Final candidates after processing by the trained CNN. Note that all three “definite” CMBs labeled by neuroradiologists are detected (red arrows). CMB, cerebral microbleeds; CNN, convolutional neural network; GRE T2*WI, gradient-recalled echo T2* weighted image.

A granular morphology filter was used to extract 22540 CMB candidates from the test data. After eliminating linear structures, the count was reduced to 7754. The subsequent removal of small granular structures further reduced the count to 5055, and the count dropped to 4632 after eliminating subtle low-signal intensity structures. The proposed method yielded an initial count of 4632 CMB candidates before CNN processing. The upper section of Table 1 shows the refinement of the CMB candidates in the morphology bank-based detection.

Table 1.

Comparison of the morphology-filter-bank-based and comprehensive methods

	Morphology-filter-bank-based detection			Comprehensive detection
	Granular morphology filter	22540
	Eliminating linear structures	7754
	Removal of small granular structures	5055
	Removal of subtle low signal intensity structures	4632
Initial CMB candidate		4632		137825
Final CMB candidate		812	Sensitivity	1585	Sensitivity
“Definite”		152	95.0%	156	97.5%
“Possible”		529	FP/patient	653	FP/patient
FP		131	4.37	776	25.87

The top section of the table represents the refinement of CMB candidates in the morphology-filter-bank-based detection. The bottom section represents the number of CMB candidates, sensitivity for “definite” and “possible,” and FP/patient ratio for the two detection methods. CMB, Cerebral Microbleed; FP, false positive.

Comparison of the morphology bank-based method and the comprehensive method

From the test set, our proposed morphology-filter-bank-based method extracted 4632 initial CMB candidates, whereas the comprehensive approach identified 138725 candidates. These initial candidates underwent CNN processing to eliminate the FPs. After processing, our method yielded 812 final CMB candidates, with 152 identified as definite (sensitivity: 95.0%) and 529 as possible CMBs. The remaining 131 candidates were neither definite nor possible CMBs (FP/patient ratio: 4.37, FP/slice ratio: 0.19). In contrast, the comprehensive method produced 1585 final CMB candidates, of which 156 were definite (sensitivity: 97.5%) and 653 were possible CMBs. However, 776 candidates were neither definite nor possible CMBs (FP/patient ratio: 25.87, FP/slice ratio: 1.14). The number of CMB candidates, sensitivity for definite cases, and FP/patient ratio for each of the two detection methods are presented in the bottom section of Table 1. The FROC curve analysis in Fig. 5 indicates the superior performance of our morphology-filter-bank-based method to that of the comprehensive approach.

Fig. 5.

FROC curves of the morphology-filter-bank-based and comprehensive detection methods. FROC, free-response operating characteristic.

Discussion

To our knowledge, this is the first study to apply a morphology filter bank, a type of shape filter, to the automatic detection of CMBs in the brain on 2D-T2*WI. During the approach based on the morphology filter bank, the initial CMB candidates were extracted using a granular morphology filter, yielding 22540 candidates. Subsequently, a linear morphology filter was applied to exclude vascular structures. Further refinement was achieved with two additional filters, reducing the number of initial candidates to 4632. In contrast, the comprehensive method did not involve such filtering, and the initial number of CMB candidates was large at 1368625. The sensitivity for definite CMBs among the final candidates after CNN processing was 95.0% for the morphology-filter-bank-based method and 97.5% for the comprehensive method. Although the sensitivity was slightly lower for the morphology-filter-bank-based approach, the number of FPs per patient was considerably reduced at 4.37 compared with 25.87 for the comprehensive method.

The superiority of the morphology-filter-bank-based approach over the comprehensive method was demonstrated through an analysis of FROC curves. The success of the morphology-filter-bank-based approach can be attributed to its effective initial CMB candidate refinement, which led to a substantial reduction in FPs after CNN processing. In addition, the incorporation of CMB candidates extracted through the morphology filter bank during CNN training played a significant role in the success of the training process. This highlights the effectiveness of the morphology filter bank.

Previous studies have proposed various automatic detection methods for CMBs,^{18,19,22–24} but the use of SWI alone, a combination of SWI and T2*WI, or the use of T2*WI alone has not been reported. Our study is the first to utilize T2*WI exclusively. Studies employing 3D-SWI have documented efforts in automatic CMB detection utilizing a combination of filters and CNN. Chen et al.¹⁸ and Liu et al.¹⁹ reported the use of fast radial symmetry transform (FRST) filters for CMB detection and CNNs for FP removal using 3D-SWI. FRST is a technique that enhances local objects with spherical or near-spherical geometry.²⁵ Chen et al. achieved a sensitivity of 94.7% and 11.6 FPs per patient for automatic detection of CMBs using 3D-SWI performed with the 7T MRI machine.¹⁸ Liu et al. achieved a sensitivity of 95.8% and 1.6 FPs per patient using the best model that combined phase imaging and SWI performed with 1.5T and 3T MRI machines, respectively.¹⁹ Our results are based on 2D-T2*WI acquired from both 1.5T and 3T MRI machines and cannot be directly compared with 3D-SWI. However, our results (a sensitivity of 95.0% and 4.37 FPs per patient, using a morphology-filter-bank-based approach) are comparable to those of previous reports in terms of sensitivity and the number of FPs.

The advantage of using 3D-SWI for automatic CMB detection is that there is no gap between slices. The slice is thin, making it easier to grasp the shape of lesions in the head-to-tail direction and identify vascular structures running in that direction for exclusion. However, 3D-SWI is time-consuming and is less accessible at facilities than 2D-T2*WI. In addition, 3D-SWI requires post-processing of phase images, and the processing methods vary among MRI manufacturers. On the other hand, 2D-T2*WI can be easily performed in most facilities, and the average image acquisition time is shorter than that of SWI. Considering the anticipated widespread clinical application of this algorithm in the future, the development of an algorithm that can automatically and accurately detect CMBs in 2D-T2*WI would also be beneficial. In addition, most previous studies on automatic CMB detection have used images from a single MRI scanner for machine learning or the training and validation of CNNs, which have not shown sufficient generalization across vendors or field strengths. In this study, we established an algorithm for automatic CMB detection using images from seven machines from four companies with field strengths of 1.5T and 3T. We believe that this highly versatile algorithm can be adapted to various machines.

Our study had several limitations. First, our approach used 2D-T2*WI only, making it difficult to compare it with studies using 3D-SWI. We did not apply our algorithm to 3D-SW images; they were not acquired for most of the 60 cases. In the future, it will be necessary to accumulate cases with available T2*WI and SWI acquired during the same examination to facilitate a comparison between the two techniques. Second, we did not directly compare the morphology filter bank used for CMB candidate detection with the widely used FRST filter. However, we believe that our results are comparable to those of previous studies using the FRST in 3D-SWI; they showed no significant difference in sensitivity or FP counts in detecting CMBs. Third, we compared the morphology-filter-bank-based approach with a comprehensive method utilizing the same CNN trained on candidates extracted using the morphology filter bank. If a different CNN trained on candidates from the comprehensive method were used, the effectiveness of the morphology filter bank could have been further demonstrated. Fourth, we used T2*WI obtained from multiple types of machines; our study was conducted in a single facility with a limited number of cases. Further investigation and refinement of the algorithm using data from other facilities or publicly available datasets with a larger number of cases are required to confirm our findings.

Conclusion

The automated detection of CMBs on 2D-GRE T2*WIs using a CNN and a morphology filter bank showed good accuracy. We achieved a reasonably high detection rate and low FP count through effective CMB candidate refinement with the morphology filter bank and FP removal with a CNN. Furthermore, the algorithm demonstrated high versatility, given that it used 2D-GRE T2*WIs acquired from seven machines manufactured by four companies with field strengths of 1.5T and 3T.

Supplementary Information

Supplementary Fig. 1.

Flowchart of the training data preparation. CMB, cerebral microbleed; CNN, Convolutional Neural Network.

Supplementary Fig. 2.

CMB candidate detection from a test dataset using the two detection methods. CMB, cerebral microbleed; CNN, Convolutional Neural Network.

Supplementary Table 1.

Breakdown of primary diseases in 60 patients.

Supplementary Table 2.

Models and parameters used for gradient-recalled echo T2* weighted imaging including the number of cases acquired with each model