Introduction

Diffusion magnetic resonance imaging (dMRI) allows examining microstructural changes occurring in multiple sclerosis (MS) that are more specific to the disease than conventional T1- or T2-weighted imaging¹. The intracellular volume fraction (ICVF), being one of the advanced dMRI measures, refers to the axon and dendrite density and is used to evaluate the distribution of white matter lesions in MS patients in comparison to healthy controls (HC)². Here, we aimed to classify MS patients and HC based on ICVF data using a convolutional neural network (CNN). Subsequently, we analyzed the relevant voxel information that contributed to the classification decisions using the DeepLIFT algorithm³. This relevance analysis⁴ helps to discover new MS markers or to confirm the current ones, thus supporting CNN-based clinical applications.

Materials and Methods

Diffusion MR imaging was performed with 64 MS patients and 64 HC using a 3T MRI (Siemens Prisma, 20 channel head coil). The MR diffusion acquisition used a multi-shell and multi-slice (SMS-slice-factor= 4) protocol with 1.5 mm isotropic resolution. The diffusion series was measured twice with reversed phase-encoding polarity to facilitate compensation of susceptibility induced geometric deformation. The microscopic diffusion anisotropy parameter ICVF was computed using the spherical mean technique (SMT⁵). For all subjects, we selected a two-dimensional image in transverse orientation at a predefined slice position from the three-dimensional ICVF volume. The resulting dataset of ICVF images was randomly split into a training group with 32 MS and 32 HC samples and a testing group with the remaining 64 samples.
To classify patients and controls, we used the following network architecture⁴. The 2D CNN had five convolutional layers (filter sizes = 16, 16, 32, 32, 64; kernel size = 3×3) and two fully connected layers (number of neurons = 8, 2) with rectified linear unit (ReLU) activation functions. For the training procedure, a data augmentation generator was used to perform image rotation, horizontal and vertical shifting, scaling, and horizontal flipping to a batch of samples. The trained CNN model showed sufficient performance on the test set (80% accuracy) and was used for further analysis.
After network model training, we applied the DeepLIFT algorithm³ to assign the relevance to the input image voxels in their ability to explain the contribution of these voxels to the output classification score. These relevance maps were qualitatively evaluated on a single subject and a population-based basis. To analyze the significance of lesion information for the classification, we extracted FLAIR-based lesion maps, which were initially defined using LST⁶ and manually adjusted by expert readers. We compared these lesion maps with the relevance maps for all correctly predicted MS patients and only for patients with the highest classification score (>0.9). We used the mean intersection over union (IoU) score as an evaluation metric in order to evaluate how much of the FLAIR-based lesion voxel contribute to the ICVF-based classification.

Results

Figure 1 shows the relevance maps thresholded for the first and the last percentile for two correctly predicted HC and two correctly predicted MS. We see that the relevance values for HC are smaller than for MS. For MS subjects, positive relevance is distinctly more pronounced and located in the central brain region. The higher degree of relevance for the MS population as well as the location of the most relevant voxels is well seen on the average maps in Figure 2.
FLAIR based lesion maps and corresponding ICVF-based positive relevance maps are shown for three MS patients in Figure 3 (in rows). We see that most of the relevant areas correspond to the lesions in all three subjects. Table 1 summarizes the mean IoU score across the subjects. The IoU increases when increasing the positive relevance threshold and is higher when computed only for MS patients with high classification accuracy.

Discussion and Conclusion

In our study, we identified MS patients using CNN on ICVF data. The CNN decision strategy was analyzed using DeepLIFT relevance maps. The relevance maps revealed that voxels in the central brain area contribute mostly to the classification. By comparing the lesion maps and the relevance maps of MS subjects, we find that the lesion voxel information in ICVF images is obviously important for the CNN classification. Specifically, we observed a relation between the high classification score and the high relevance of the voxels matching with the lesion voxels in MS patients. For future studies, the discrepancy between voxel locations identified as FLAIR-based lesion and relevant voxels for ICVF-based classification should be analyzed.
The study has some limitations. First, for the current study the performance of the model appears sufficient, but for future studies, we suggest to hyper-tune the parameters of the CNN and generate more specific relevance maps. A second limitation is the use of lesion maps derived from FLAIR images. Since the lesion information can differ between the two contrasts (FLAIR and ICVF), future combination of FLAIR-based and ICVF-based lesion maps might enable a more robust analysis.

Acknowledgements

This study was supported in parts by the Carl-Zeiss-Foundation (CZ-Project: Virtual Workshop), the German Research Foundation (RE1123/21-1), and the Austrian Science Fund (FWF3001-B27).