Introduction

Multiple sclerosis is characterized by inflammatory and demyelinating lesions in the Central Nervous System¹.Focal inflammation with breakdown of the blood-brain barrier leads to gadolinium extravasation in the parenchyma after contrast injection, highlighting contrast-enhanced lesions (CEL)².
The detection of CELs is fundamental for the diagnosis and monitoring of MS^3,4.
The main challenges for CEL identification and segmentation are the low frequency, heterogeneous shape and often low volume of CELs⁵. Manual detection of CELs is highly time-consuming. For this reason, an automatic tool can be useful and reduce the variability of the results in clinical practice. Moreover, in literature, only a few methods performed this task. This work aims to develop a deep learning tool to detect and segment especially low-volume and sparse lesions in MRI images.

Methods

The dataset was obtained by 1.5 and 3 T MRI systems in 7 centres belonging to the SMSC⁶.
The MRI acquisitions and their respective TR [ms], TI [ms] and slice thickness [mm] ranges are: T1-weighted (T1w) MPRAGE ([1720-2700]; [900-110];[1-1.2]), T1w-MPRAGE with gadolinium contrast agent([30-2700]; [0-1100];[1-3])and FLuid Attenuated Inversion Recovery (FLAIR) ([5000-7500]; [1600-3000];[1-1.3]).
Furthermore, the automatically segmented white matter lesion (WML) and ground truth masks are manually corrected and segmented by expert neurologists and radiologists.
The dataset included 373 scans from 280 patients, divided into scans with CELs and without CELs (Figure 1). These scans are introduced to mimic the real clinical situation, which shows a high percentage of MRI scans without CELs⁷.
62 patients had multiple visits, considered independently and included in the same dataset for training, validation or testing.
The images of the same acquisition were co-registered and skull-stripped using FSL⁸ and HD-BET⁹, respectively.
A 3D patch-based UNet-based network for cortical lesions¹⁰, was expanded with two more convolutional layers of 256 and 512 filters, respectively.
Patches were sampled on the WMLs, indeed during the training process, the ratio of the CEL patches (positive patches) over WML patches (negative patches) was 1/3. In control patients, instead, all the patches were cropped in WML.
In the validation process, the sliding window inference was replaced by the sample of WML. Consequently, the validation strategy is similar to the training one, and all the possible regions in which CELs can be present are given as input to the model (Figure 2).
The further implementation takes into account the loss function.
The starting loss function¹¹ was:
$$$Loss_{function}=0.5*Loss_{Dice}+Loss_{focal}$$$
However, the dataset is unbalanced and the number of negative patches is higher than the positive ones $$$(rate_{unbalance}=2/9)$$$. Moreover, the Dice loss in negative patches is usually 1 if the number of False Positive (FP) voxels is nonzero. As a consequence, the learning process gives low importance to the $$$Loss_{Dice_{positive}}$$$ decrease but concentrates more on lowering $$$Loss_{Dice_{negative}}$$$. Therefore, we defined a new loss function to reduce the weight of $$$Loss_{Dice_{negative}}$$$ multiplying it with the rate of patches unbalance: $$$Loss_{function}=mean(rate_{unbalance}*Loss_{Dice_{negative}}+Loss_{Dice_{positive}})+mean(Loss_{focal_{negative}}+Loss_{focal_{positive}})$$$
The dataset was randomly subdivided into 11 groups. A group was randomly chosen as the testing dataset. The rest of the groups formed 10-fold cross-validation. An example of the training pipeline is in Figure 3.The Dice Score Coefficient (DSC) was calculated in patches and whole images during the validation process.
The output of the model was filtered removing the prediction outside the WML mask. Consequently, 3 FP were deleted.

Results

The weighted loss function increased the weights of $$$Loss_{Dice_{positive}}$$$,thereby leading to a DSC improvement in positive patches. Furthermore, it improved the segmentation and detection performance both in the validation and test datasets. (Figure 4)
The model reached over the whole images a DSC of 0.76, a True Positive (TP) rate of 0.93, and an FP rate of 0.02. The wrong detections were localized on low-volume lesion (Figure 5). Regarding the segmentation in these groups, the DSC reached values similar to the ones in other volume ranges.

Discussion

The sampling strategy was fundamental in decreasing the FP prediction and achieving a stable DSC and loss function during the training process. Furthermore, the weighted loss function permits high performance in validation and test datasets.

Conclusion

The strategies implemented in our UNet-based network permit the detection and segmentation of low-volume and sparse lesions in MRI images.
As a future step, we need to consider the application of the tool in a large dataset, to achieve generalizable results and aim at further reducing the number of FN lesions.
Another possible future improvement might be the achievement of a similar accuracy without the need of a WML mask as an input, which relies on its segmentation accuracy.

Acknowledgements

No acknowledgement found.