Keywords

1 Introduction

Cerebral microbleeds (CMBs) are hemosiderin deposit caused by structural abnormalities of the blood vessels [1]. They are prevalent in people suffering from cognitive and cerebrovascular disease such as stroke and Alzheimer’s disease. They are also asymptomatic and present in cognitively normal individuals. The detection of CMB from MRI is thus clinically important to assess cerebrovascular burden.

Magnetic susceptibility is affected by the presence of CMB in brain parenchyma [2], and susceptibility-weighted imaging (SWI) [2] has been shown to image CMB with excellent sensitivity, limited mostly by the imaging resolution (typically 1 × 1 × 2 mm3). In SWI scans, CMBs appear as small spherical hypo-intense areas in tissue with a similar appearance than the numerous blood vessels, also seen in SWI, when observed in cross-section. For this reason, the clinical observation of CMB, often based on the Microbleed Anatomical Rating Scale (MARS) [3], is tedious and time-consuming because of the large number of similar features from blood vessels. Indeed, CMB mimics include vessel cross section, calcification, and cavernous malformation on SWI.

Automated methods are also challenged by the large number of mimics resulting in numerous false positives while there is a relatively small number of true positives since the prevalence is about a few CMB in 10% of asymptomatic elderly subjects [4]. This creates a large unbalance classes hampering supervised training, further compounded by the limited size of datasets where ground truth (human expert annotation) is available. To address this problem, data augmentation methods are often used [5]. Synthetic Minority Oversampling Technique SMOTE [5], is one strategy to balance the classes by generating new pseudo TP (true positives) instances from the existing minority TP cases. In this paper, we investigate data augmentation by generating synthetic CMBs. We compare this new approach to SMOTE and the other widely used data augmentation method using rotation and translation. We describe the method to create realistic synthetic CMB and demonstrate that performance of a neural network classifier can be improved by transferring the learning from the synthetic data to real clinical data. We do not aim at presenting a new classifier, but at demonstrating that using synthetic lesions can improve the performance of supervised learning. We use in this paper a simple feed forward Artificial Neural Network as a case example.

2 Method

2.1 Dataset

Data are from the Australian Imaging Biomarkers & Lifestyle (AIBL). We used 64 scans from 39 participants (including some repeat scans at 18 months interval). There were 24 patients clinically diagnosed with Alzheimer’s disease (AD), 34 subjects with mild cognitive impairment (MCI) and 6 healthy controls (HC), with an average age of 79.21, 77.74 and 79.98, respectively. The scans comprised 27 females and 37 males. All subjects underwent an anatomical T1-weighted (T1 W) and a SWI acquisitions on a 3T Siemens TRIO scanner. SWI was reconstructed online using the scanner system (software VB17). The 3D SWI parameters were: 0.93 × 0.93 mm2 in-plane resolution and 1.75 mm slice thickness, repetition time/echo time of 27/20 ms, and flip angle 20°. T1 W images were acquired using a standard 3D magnetization-prepared rapid gradient echo (MPRAGE) sequence with in-plane resolution 1.0 × 1.0 mm2, slice thickness 1.2 mm, repetition-time/echo-time/TI = 2.300/2.98/900, flip angle 9°, field of view 240 × 256, and 160 slices. Scans were reviewed by two clinical experts and marked using the MARS criteria as definite and possible.

2.2 Preprocessing

For all the scans, N4 bias field correction technique [6], skull-stripping masking to exclude non-brain tissues [4], and histogram matching were applied before any processing. All SWI images were normalized between [0 1]. Brain tissues were segmented from the T1W into gray matter (GM), white matter (WM), and cerebrospinal fluid (CSF) using posterior probability classification after fitting a mixture of Gaussian distribution to the histogram using the expectation-minimization method. A final region of interest (ROI) mask was built by merging the WM and GM voxels.

2.3 Features of Actual CMBs

From the analysis of the SWI datasets, CMBs were characterized using four parameters: size, shape, intensity, and location. In our dataset, the CMBs were uniformly located in the WM, GM tissues, and therefore location of synthetic lesions was uniformly distributed in the corresponding mask.

CMB Intensity:

It was observed that large CMB intensity had a minimum value of 0, while the intensity of smaller size lesions became closer to tissue intensity. Figure 1 top left panel shows the minimum value of expert identified CMB. We assumed that this effect was entirely due to partial volume effect and that the minimum intensity value of a CMB is 0. This is consistent with SWI processing where the square of the high frequency filtered phase multiplies the signal magnitude, resulting in 0 intensity.

Fig. 1.
figure 1

The size distribution (top panels) and minimum intensity (bottom panels) for the real CMB in blue (left panels) and the generated synthetic CMB in red (right panels). The minimum intensities are comparable and entirely due to partial volume effect in the synthetic case. (Color figure online)

Full size image

CMB Size:

11 × 11 × 11 patches with definite CMB at their center were extracted from the 64 SWI images, resulting in 144 patches making up the TP class. For each patch, a mixture of distributions was fitted using the expectation minimization algorithm: a Gaussian to model brain tissues and a uniform distribution to model outliers due to partial volume effect, and low intensity pixels from vessels and CMBs. A maximum a posterior classification was performed to create a mask of the CMB in the patch (cleaned using connected components). The mean of the Gaussian was assumed to be the tissue mean intensity while an intensity of zero was assumed to be the intensity of the CMB. Fractional CMB content was estimated for each pixel of the CMB mask by using a standard partial volume model. Eventually, fractions of CMB for each pixel were summed up to estimate the volume of each CMB as shown in Fig. 1.

2.4 Generation of Synthetic CMBs

It is our goal to create realistic synthetic CMB (sCMB), which should thus have the same characteristics as the real ones defined in the previous section.

A 3D Gaussian function was generated on a 110 × 110 × 110 patch (ratio matching the SWI resolution ratio). Its location was the center of the middle pixel of the patch with added uniform random noise in all three axis within [±15, ±15, ±15] voxels. In order to simulate variation around a spherical shape, given a spherical volume, two of the 3D Gaussian standard deviations were multiplied by a random number between 0.5 and 0.9, while the third one was defined so that the overall volume was conserved. The shape was adjusted using the 3D Gaussian standard deviations to match the actual SWI resolution ratio. Random variation of the three standard deviation values of the 3D Gaussian allowed random variation of shape around a mean ellipsoid while keeping the same overall volume. The volume of the sCMB was randomly sampled from a smooth approximation of the size distribution of real CMBs (shown in Fig. 1 top left panel). Uniform random rotation was also added on the three axis within ±30°.

The Gaussian intensity profile was thresholded at half maximum to create a mask simulating a high resolution sCMB (0 within the mask and 1 outside). The patch was then down sampled to 11 × 11 × 11 to simulate partial volume effect (values ranging between 0 and 1 to simulate PVE). The resulting patches were then multiplied with patches extracted from the SWI images without any CMB, whose selection is explained in the next section. As a results, a synthetic TP patch (sTP) could be created for which the lesion size and appearance could be computed in the same way as actual CMB patches as explained in the previous section and displayed in Fig. 1 right panels.

2.5 Patch Extraction

From the SWI dataset, patches were extracted and assigned to two categories: (1) Patches with real CMB defined as definite by the two experts (TP), (2) Patches with no actual CMB (neither definite nor possible). The second type was further divided into those patches containing either blood vessels or some CSF (seen as dark and possibly confounding CMB), and patches containing mostly GM or WM tissues. They were labeled as true negative (TN) with and without blood vessels. Vessels were detected using the Radial Symmetry Transform RST [7]. In order to create a patch with a synthetic CMB, the 11 × 11 × 11 mask described in the previous section was multiplied with a patch from the TN class, creating a synthetic TP patch (sTP). We selected the TN patches so that the proportion of sTP with and without vessels was 50%/50%.

2.6 Experimental Method

We used a 2-fold with repeated 25 random sub-sampling validation: for each of the 25 sampling, 19 random subjects (out of 39 subjects) were set aside as the testing data, while the remaining were used for training. This resulted in a total number of TP (actual CMB patches) to be around 70 for testing and 70 for training (actual number depends on the subjects selected). The description of the training strategy for the 8 models can be found in Table 1. The concatenation of the results for the 25 sampling for each model was used to smooth the ROC curves.

Table 1. Models considered in this study with approximate number of patches in each class.
Full size table

2.7 Classifier

To test our hypothesis that synthetic lesion would improve the performance of a classifier, we used a single-hidden layer feed-forward neural-network [8], trained with scaled conjugate gradient [9]. This ANN was implemented with Matlab. The activation function in the hidden layer was Leaky ReLU [10] with 55 neurons. The activation function of the output unit was the logistic sigmoid (LOSI) [11]. The number of epochs was 2000, the mean square error was used as the cost function.

3 Results and Discussion

3.1 Synthetic CMB Generation

Synthetic CMB could be created with no discernable differences from actual lesions. The size and intensity were similar between actual and synthetic lesions as shown in Fig. 1. The location of synthetic samples was uniform in the GM and WM tissues as observed for real ones. Figure 2 shows examples of axial sections from the 11 × 11 × 11 patches.

Fig. 2.
figure 2

Typical examples of axial section for the different classes. Note that the sCMB is not centered (second left panel), similar to the real one (left panel), because we added random variation in location and shape.

Full size image

3.2 Results of CMB Classification Using ANN

Results for the ANN classification on the testing datasets is reported in Table 2. We show the area under ROC (AUR), mean sensitivity, specificity and accuracy along with the standard deviation over the 25 sampling 2-fold cross validation design. Model 2 would not converge to any meaningful results due to the unbalanced classes and is omitted here. Figure 3 shows the ROC curve for all the models. The results using SMOTE (model 8) were the lowest of the three data augmentation methods. Using synthetic data provided good results, and the best model was the one training on synthetic data alone. Merging synthetic with the TP class had also good results close to training on synthetic data alone. Using standard geometric data augmentation improved performance but not as much strategy involving using synthetic data.

Table 2. Performance of the models on the test datasets. Values are averages ± standard deviation over the 25 draws from averaging the outputs. Bold shows the best performance models.
Full size table
Fig. 3.
figure 3

ROC curve for the different models using a 2-fold with repeated 25 random sub-sampling validation design.

Full size image

3.3 Discussion

Our results suggest that augmenting data with synthetic lesions could help increasing supervised classifier performance to detect microbleeds in MRI SWI. The best results were obtained by training on synthetic data alone even though the performance were measured on real data.

Our proposed approach addresses one critical issue for lesion detection task: when the prevalence of a lesion type is low, a very small number of TP are available for training supervised classifier limiting the complexity of the possible models to consider. Another consequence is the class unbalanced between the positive and negative. Indeed, the naĂŻve approach of training the ANN on very unbalanced data (1:10) failed to produce any meaningful results. Adding sCMB allowed to balance the classes and increased the number of examples resulting in much improved performance.

Our proposed approach is very dependent on how realistic the synthetic lesions are created. We described a simple method based on a Gaussian profile that is added to the magnitude SWI images (magnitude multiplied by filtered phase). A better characterization of actual microbleeds using high resolution MRI, such as 7T, could help improve the generation of simulated lesions. In addition, adding the synthetic lesions on the complex image, before SWI processing, might improve the realism of the synthetic data.

Despite those limitations, augmenting medical data using synthetic lesions could allow to use complex classifiers, such as deep learning network, with large number of parameters since arbitrary large training datasets could be created along with the perfect ground truth necessary for supervised learning. Of course that approach is limited by how close the synthetic ground truth is from the actual and elusive ground truth. This is tantamount to converting human expertise (a priori knowledge about what lesions look like) into data that is then transfer through learning into the weights of a classifier. More research on different applications would inform whether our proposed approach could allow automated methods to push performance beyond that of experts.