A multi-sequence MRI integration framework using SwinUNETR-v2 for multiple sclerosis lesion segmentation

Abstract

Multiple Sclerosis (MS) is a chronic brain disease that affects the brain and spinal cord, where Magnetic Resonance Imaging (MRI) plays a key role in diagnosis. While manual analysis of brain MRIs is important, it is time-consuming and prone to human error. Artificial Intelligence (AI)-driven Computer Aided Diagnostic (CAD) systems have therefore gained traction due to their ability to provide more consistent and reliable assessments. This study presents a multi-sequence framework that integrates four MRI modalities with the SwinUNETR-v2 backbone for MS lesion segmentation. The main contribution is a task-oriented integration of multi-sequence input design (implemented as a four-channel volume representation), refined preprocessing (including balanced foreground/background patch extraction), and a weighted loss formulation under a controlled five-fold evaluation protocol. This approach achieved a peak DSC of 90.7% and a mean DSC of 88.3%. Moreover, when compared against other state-of-the-art segmentation methods—including AttentionUNet, DenseResidualUNet, SegResNet, FCNN, and nnUNet-v2—the multi-sequence SwinUNETR-v2 setup consistently outperformed these models across all key metrics, demonstrating strong effectiveness in identifying MS lesions.

Introduction

Multiple sclerosis (MS) is a chronic autoimmune disorder that affects the brain and spinal cord, leading to progressive neurological damage  [1]. Early and accurate lesion identification is essential for diagnosis, treatment planning, and longitudinal monitoring. However, manual delineation of MS lesions remains labor-intensive, observer-dependent, and difficult to scale in routine clinical workflows.

Magnetic resonance imaging (MRI) provides a non-invasive basis for detecting MS-related abnormalities. Different MRI sequences emphasize complementary tissue characteristics and are influenced by acquisition protocols. Commonly used modalities include T1-weighted (T1-w)  [2, 3], T2-weighted  [4, 5], Proton Density-weighted (PD-w)  [6], and Fluid-Attenuated Inversion Recovery (FLAIR)  [7, 8]. Representative examples of these modalities are shown in Fig. 1.

Fig. 1.

Illustration of different MRI modalities as presented by  [3, 5, 6, 8]

Recent advances in Artificial Intelligence (AI), particularly Deep Learning (DL), have made automated medical image segmentation increasingly effective  [9, 10]. By learning hierarchical representations directly from data, DL models can capture discriminative lesion patterns with limited handcrafted feature design  [11]. U-Net and its variants remain among the most influential segmentation architectures  [12], while transformer-based models have shown strong performance in both computer vision and natural language processing  [13–15]. In particular, Vision Transformers (ViTs) have achieved state-of-the-art results on several benchmarks  [15]. Prior studies also suggest that DL methods can better handle heterogeneous acquisitions and imaging noise, which is essential for multi-center datasets collected with different scanners and quality-control protocols  [16–18].

Motivated by these advances, this work presents an integrated framework for MS lesion segmentation that combines multi-sequence MRI with transformer-based 3D backbones. Throughout this manuscript, we use multi-sequence as the primary term for combining the four MRI sequences (T1-w, T1-wKS, FLAIR, and T2-w), while multi-channel is used only to describe their tensor representation along the channel axis.

To strengthen training under limited data, we employ a refined preprocessing pipeline that includes balanced 50/50 foreground-background patch extraction to improve sample diversity and data consistency. We also adopt an enhanced hybrid loss that combines Dice and cross-entropy terms with weights and .

Building on this pipeline, we implement and evaluate the residual-convolutional SwinUNETR-v2 architecture, which couples transformer-based global context modeling with local residual convolutional refinement. Experiments on the public 3D-MR-MS dataset using five-fold cross-validation show consistent improvements in Dice similarity coefficient (DSC), sensitivity, and precision over state-of-the-art methods and conventional 2D/3D single-sequence approaches, while preserving a practical balance between segmentation accuracy and computational cost.

Related works

Many literature attempted to analyze and solve the problem of MS detection and segmentation. Using a hybrid CNN-RNN architecture,  [19] developed a framework that took advantage of an adaptive CNN architecture featuring an attention mechanism  [20] and a convolutional Long Short-Term Memory (LSTM) layer  [21]. The framework allowed for the segmentation and quantification of lesions from MRI of MS. Throughout the work they employed a set of preprocessing methods, including image registration  [22, 23], bias correction  [24], and skull stripping  [25]. They applied their model to FLAIR modality volumes in the ISBI 2015 challenge dataset  [26] and scored 89% using the Dice Similarity Coefficient (DSC).

One more CNN model that extracts lesion patches, modified from the U-Net architecture, was used by  [27] to localize and segment MS lesions on MICCAI 2008  [28], MICCAI 2016 (MSSEG)  [29], ISBI 2015  [26], OASIS-3  [30], BraTS 2019  [31], and clinical (private) datasets. The proposed data was subsequently preprocessed using FSL  [32–34], and a DSC score of 67.63% was obtained.

In order to separate lesions, Gray Matter (GM), Cerebrospinal Fluid (CSF), White Matter (WM), and individually among 1000 MS patients,  [35] employed a fully convolutional neural network with a 2D U-Net architecture. The input of the network consisted of fifteen permutations of FLAIR, T1-w, T2-w, and PD-w images. A range of preprocessing techniques were used, such as data augmentation  [36, 37], image denoising  [38–40], bias correction  [24], intensity normalization  [41], and skull stripping  [25]. With a DSC of 90%, their performance was encouraging.

Dense U-Net, a version of the original U-Net, was presented by  [42]. The four individual input image modalities (T1-w, T1-wc, FLAIR, and T2-w) were first fed into a dense block by their model, where certain convolutions and skip connection procedures were conducted. The output was then forwarded to the network’s max pooling layer. Until they hit the encoding section bottleneck, they kept doing the same procedure over and over. The same technique was performed in reverse during the decoding phase. Their approach substituted dense blocks for the first three convolution layers at the start and finish of the original U-Net. Using MICCAI 2016 (MSSEG)  [29] as their dataset, by employing strategies like data augmentation, they expanded the data  [36, 37], scaling intensity levels in the direction of zero  [43], and removing null (no lesions) data samples. They received an 86.6% DSC score.

An autoencoder and regularization network architecture was presented by  [44]. Three distinct components were necessary for this architecture: the regularization network, the decoder, and the encoder. The autoencoder network’s encoder and decoder are its main components, and the regularization network was the main center for enhancing the MS lesion segmentation task’s effectiveness. A softmax layer and three perceptron layers made up the regularization network. The primary objective was to supply the regularization network with the encoder’s output feature vector, which was subsequently fed to the decoder. This greatly enhanced the outcomes since the extra network assisted in avoiding overfitting. A clinical dataset was utilized, consisting of 56 participants with four modalities (T1-w, FLAIR, T2-w, and PD-w). They simply utilized two preprocessing techniques—skull stripping  [25] and image registration  [22, 25]—on a clinical dataset. A 50% DSC was obtained.

By combining attention gate (AG)  [45], efficient channel attention (ECA)  [46], and Atrous Spatial Pyramid Pooling (ASPP),  [47] introduced a novel dense residual U-Net model that improved the effectiveness of automated MS lesion segmentation using 3D single-sequence MRI data. Their methodology was validated using the publicly accessible ISBI 2015  [26] and MSSEG 2016  [29] challenge datasets. The model was trained on ISBI 2015  [26] and tested with MSSEG 2016  [29] and got an average DSC of 67.27%, an average Positive Predictive Value of 65.19%, and an average sensitivity of 74.4%.

Using 1488 datasets of 934 MS patients from 81 scanners,  [48] introduced a novel technique for externally training a three-dimensional convolutional neural network for Contrast Enhancing (CE) lesion segmentation. T1-wc and FLAIR imaging were used. In clinical settings, 504 T1-w post-contrast and FLAIR pictures from MS patients were used to evaluate the externally trained model. They applied registration  [22, 23] and patches extraction as their preprocessing methods. The approach was a U-Net like structure where two encoders and one decoder were used. The first encoder encoded and extracted features from FLAIR input, while the second one extracted features from T1-w. The extracted features from both encoders were concatenated along the encoding process. Finally, the decoder tried to get lesions mask. In the clinical test dataset, CE lesions were recognized as the gold standard definition by two neuroradiologists (R1 and R2). The algorithmic outcome was assessed at both the patient and lesion level. They achieved a sensitivity of 75%, specificity of 99%, precision of 91%, and accuracy of 96%.

For T2-w lesion segmentation,  [49] developed a 3D single-sequence, multi-arm U-Net that was trained on a sizable, multicenter clinical trial dataset of relapse MS. Its generalizability was evaluated using an external dataset from the MICCAI 2016  [29] MS lesion segmentation challenge, as well as other relapsing and main progressive MS clinical trial datasets. In comparison to manual lesion annotations, they assessed the model’s ability to reliably distinguish T2-w lesion volumes from treatment and control groups and how baseline T2-w lesion volumes and clinical impairment ratings relate to one another. Preprocessing methods such as image registration  [22, 23], bias field correction  [24], and skull stripping  [25] were used. On internal test datasets, the trained model showed a mean DSC of % and a lesion detection sensitivity of %; on the external test dataset, the model showed good performance with a mean DSC of 62% and a lesion detection sensitivity of 68%.

For  [50], they retrospectively included 746 patients from their MS clinic’s routine assessments. Using one of seven different scanners and adhering to an MRI procedure that incorporates 2D or 3D FLAIR, T1-w, and T2-w volumes were prerequisites for inclusion. Reference lesion masks were created using an initial segmentation model and manually adjusted for the training (n = 571) and validation (n = 70) datasets. Their segmentation approach was based on the popular nnU-Net  [12], which had won a set of biomedical segmentation challenges. The model’s robustness was tested on data from unseen scanners using a leave-one-out experiment. They applied Re-orientation  [41], image registration  [22, 23], and skull stripping  [25] as preprocessing method prior to the training procedure. They achieved a DSC of 68%, a precision of 90%, a sensitivity of 70%, and an F1-score of 78%.

[51] proposed a three-dimensional deep autoencoder network-based technique for automatically detecting and segmenting brain abnormalities without prior training on abnormal samples. The model was trained and validated using 578 normal T2-w MR volumes with no evident anomalies. They used 3 different datasets, where two of them were public including IXI  [52] and BraTs 2019  [31], along with one clinical dataset for the purposes of training and testing of the suggested 3D autoencoder. Three preprocessing methods were applied involving skull removal  [25], image registration  [22, 23], and image normalization (i.e., converting to range [0–1]). They got an Area Under the Curve (AUC) of 85.8%, Area Under the Precision-Recall Curve (AUPRC) of 73.1%, and sensitivity of 83.7%.

Recent multi-modal studies further highlight robustness-oriented fusion strategies. MGML  [53] introduces a meta-guided framework for incomplete-modality MRI segmentation, where adaptive modality fusion and consistency regularization are jointly used to preserve performance when one or more input sequences are missing. In terms of data preparation, the method explicitly constructs incomplete-modality training scenarios (missing-modality combinations) and enforces consistency constraints across available modality subsets. On the BraTS2020 dataset  [54], MGML reported 87.55% Dice (WT) across 15 missing-modality combinations.

Similarly, DGSSA  [55] emphasizes robustness to domain shift through a preprocessing/augmentation pipeline that combines structural augmentation (space colonization to generate vascular-like patterns) with enhanced Pix2Pix and PixMix synthesis. This strategy increases structural diversity before segmentation training and improves cross-domain generalization. DGSSA reported a state-of-the-art average DSC of 77.98% on retinal vessel segmentation.

Methodology

This section outlines the comprehensive framework for automated segmentation of MS lesions in brain MRI. The methodology integrates multi-sequence 3D MRI data, a robust preprocessing pipeline, and advanced deep learning architectures to achieve precise lesion delineation. The workflow begins with the publicly available 3D-MR-MS dataset  [56], comprising T1-w, T1-wKS, FLAIR, and T2-w volumes from 30 MS patients. Following preprocessing steps to standardize and augment the data, the study evaluates two transformer-based models—multi-sequence UNETR and multi-sequence SwinUNETR-v2—designed to leverage spatial and modality-specific features through four-channel input representations. The section concludes with implementation details, loss functions, and performance metrics to validate the framework’s clinical applicability.

Image dataset

3D-MR-MS  [56], a publicly accessible image dataset provided by Laboratory of Imaging Technologies  [57], is the dataset used in this study. The collection includes segmentations of white matter lesions in 3D MR volumes and volumes of individuals with multiple sclerosis. The entire dataset consists of 30 patient MRI images obtained using a 3.0T MR scanner, with multi-rater consensus used to define lesion segmentations for each participant. The dataset is explained as follows: each patient’s data represents a distinct raw of MRI images/volumes using four different sequences (T1-w, T1-wKS (i.e., contrast enhanced T1-w), FLAIR, and T2-w); the brain mask image; and the consensus sclerosis lesions segmentation mask, which is derived from multi-rater and preprocessed versions of the four modality volumes. The sizes of all volumes fall between 176512512 and 192512512. Each example includes the original MR volume and the volume following bias field correction and co-registration (geometric alignment of the images) to the FLAIR sequence space. Axial, Coronal, and Sagittal features of a sample patient volume in FLAIR, T1-w, T1-wKS, and T2-w, together with the associated lesion segmentation mask, are represented in Fig. 2.

Fig. 2.

A sample from 3D-MR-MS dataset. Each row represents one of the three different volume aspects, from top to bottom: axial, coronal, and sagittal. Each view consists of four different image modalities defined as follows: (a) FLAIR, (b) T1-w, (c) T1-wKS, (d) T2-w modalities, in addition to (e) the ground truth segmentation mask for the MS lesions

Preprocessing

The preprocessing pipeline involves several steps to prepare the dataset for training and validation. For all data volumes, the preprocessing begins with skull stripping to remove non-brain tissues from the MRI volumes and their associated modalities. This is followed by re-orienting each skull-stripped volume to the standard RAS+ coordinate space  [58], ensuring consistency across the dataset. Following the skull stripping and re-orientation, the next step is cropping the foreground region, which is typically a plane, from the volumes. To achieve this, the FLAIR modality is used as a reference to extract the foreground region because it is considered the most reliable modality for foreground tissue extraction, as demonstrated by our prior work  [59]. According to the prior study, the FLAIR modality provides the best representation for tissue segmentation, especially when the objective is to crop the foreground area. After cropping the foreground region in the FLAIR modality, the same cropping coordinates are applied to all other modalities, ensuring that the extracted foreground region is consistent across modalities.

Afterwards, for the training volumes only, random patches of size 96 96 96 are extracted. Each modality (sequence) has its own set of four random patches, taking into consideration that the coordinates of these patches are the same across all modalities. For example, the first set of four patches extracted for the first modality will share the same coordinates as the corresponding patches for the second, third, and fourth modalities. The same applies to the ground truth segmentation mask. Using the segmentation mask as a reference, each patch is extracted with a probabilistic criterion: each patch has a 50% chance of having the center from a lesion (foreground) and a 50% chance of being from a non-lesion region (background). The objective of this random extraction is to focus on extracting regions that either contain foreground or background. The selected patch thus represents a plane of lesions or non-lesions at the center of the patch.

Once the random patch extraction is completed, data augmentation methods are applied to allow further diversify of the data. This firstly includes random rotation, where each volume is rotated independently along its three axes: axial, coronal, and sagittal. Following the rotation, random flipping is applied to all three axes. Then, intensity shifting is applied, where each volume has its intensity value adjusted randomly with a probability of 50% by a factor of 0.1.

After applying data augmentation, the first patches from each modality are concatenated together to form a single large volume. For instance, the first patch from the first modality, the first patch from the second modality, the first patch from the third modality, and the first patch from the fourth modality are concatenated along the channel axis, resulting in one large volume. This operation is repeated for the second, third, and fourth patches, resulting in four concatenated volumes. Each concatenated volume has a shape of where the first three dimensions H, W, and D represent the height, width, and depth of the volume, respectively, while the final dimension C represents the channel dimension, which equals four (corresponding to the four modalities).

It is important to note that, although the 3D-MR-MS dataset comprises only 30 patients, the adopted patch-based strategy substantially increases the number and diversity of training samples. From each training volume, four 3D patches of size are sampled per iteration using a balanced 50% foreground / 50% background criterion, and each patch is further transformed by random rotations, flips, and intensity shifts. As a result, the DL model is exposed to a large variety of lesion and non-lesion configurations across epochs, which helps to mitigate overfitting and improves generalization despite the limited number of subjects.

For training volumes, this concatenation process is applied only to the extracted patches, while the corresponding lesion masks are not concatenated, as they represent the output of each concatenated volume. For validation volumes, the same concatenation procedure is followed, and no random patch extraction is applied. The re-oriented volumes are concatenated directly to form the larger volumes, following the same concatenation scheme without extracting random patches. Fig. 3 shows the steps for applying the proposed preprocessing methods.

Fig. 3.

Preprocessing pipeline for multi-sequence MRI data. Columns (from left to right) represent the full results of skull stripping, reorientation, cropping the foreground (using FLAIR as reference), random patch extraction, and the multi-sequence volume construction, with the channel dimension representing different sequences

Methods

In this study, we adapted the original UNETR and SwinUNETR-v2 architectures by modifying only the input and output shapes. Specifically, UNETR was evaluated with both single-sequence and multi-sequence inputs according to the experiment, while SwinUNETR-v2 was implemented exclusively with a multi-sequence input—both producing a 3D segmentation output.

UNETR

We adopted the original UNETR architecture  [60] with minimal modifications to suit our experimental needs. UNETR departs from conventional CNN-based U-Net designs by employing a Vision Transformer (ViT) encoder tailored for volumetric medical image segmentation. In this encoder, a 3D input volume (e.g., , where for the single-sequence FLAIR case or for the multi-sequence case) is partitioned into non-overlapping 3D patches of size . Each patch is flattened, linearly projected into an embedding space of dimension , and enriched with learnable positional encodings, forming a sequence of patch tokens.

This token sequence is processed by a stack of transformer blocks with multi-head self-attention (MHSA) and feed-forward layers (FFL), enabling the model to capture long-range spatial dependencies across the entire brain volume. Such global context is particularly important for multiple sclerosis lesions, which may appear as small, scattered regions but with distributed spatial patterns. The encoded features from several transformer layers are reshaped back into volumetric feature maps and fused with a convolutional decoder via skip connections at multiple resolutions, thereby combining global contextual information with local spatial detail, illustrated in Fig. 4.

Fig. 4.

Overview of the UNETR architecture: source  [60]. A transformer is used to extract sequence representations from various layers and then use skip connections to merge them with the decoder. The output sizes are specified for patch resolution p = 16 and embedding size K = 768

In our study, we evaluated UNETR in two configurations: a single-sequence setup using only the 3D FLAIR volume, and a multi-sequence setup where the four MRI sequences (T1-w, T1-w contrast, FLAIR, and T2-w) are concatenated along the channel axis to form a four-channel 3D input volume. In both cases, UNETR produces a 3D lesion segmentation mask, allowing us to directly assess the benefit of multi-sequence information for MS lesion delineation.

SwinUNETR-v2

SwinUNETR-v2  [61] builds on the SwinUNETR  [62] family by combining hierarchical Swin Transformer encoders with stage-wise residual convolutional blocks for 3D medical image segmentation. Unlike UNETR, which uses a flat ViT encoder with global self-attention over all patches, SwinUNETR-v2 adopts a hierarchical architecture with shifted-window attention. This design processes the volume at multiple resolutions using local self-attention windows, significantly improving computational efficiency while retaining the ability to aggregate global context through window shifting and multi-stage feature fusion.

A key component of SwinUNETR-v2 is the introduction of residual convolutional (ResConv) blocks at the beginning of each encoder stage. These convolutional layers provide strong inductive biases for local feature extraction, such as edge and texture patterns, and regularize the transformer representations before attention is applied. The combination of ResConv blocks and Swin Transformer stages enables the model to effectively capture both fine-grained local details and long-range dependencies, which is essential for accurately segmenting small, irregular MS lesions in brain MRI.

In our implementation, we use a multi-sequence 3D input volume of size , where the four MRI sequences are concatenated along the channel axis. The encoder processes this input through four successive stages with decreasing spatial resolution and increasing channel depth. The corresponding decoder mirrors this hierarchy and integrates skip connections from each encoder stage to recover detailed spatial information. We experimented with different patch sizes and configuration settings, and selected the setup that achieved the best trade-off between segmentation accuracy and computational cost on the 3D-MR-MS dataset. This multi-sequence SwinUNETR-v2 configuration consistently outperformed the UNETR baselines in terms of Dice, sensitivity, and precision for MS lesion segmentation. A full description of the proposed method is shown in Fig. 5.

Fig. 5.

SwinUNETR-v2 overview  [61]

Experimental setup and implementation details

The PyTorch framework was used to implement the UNETR and SwinUNETR-v2 models. Every experiment was conducted using a P100 GPU. Both models were trained from scratch with randomly initialized weights, without using any pretrained checkpoints, ensuring that the reported performance reflects training exclusively on the 3D-MR-MS dataset  [56, 57]. Both were trained with a batch size of one, using the AdamW optimizer  [63] with a weight decay of 1e-5 over 12,000 iterations and an initial learning rate of 1e-5. All previously mentioned data augmentation techniques were performed during the training.

Two methods were encountered to calculate sensitivity and precision metrics: volume-wise and set-wise, where the first one calculates metrics for each volume independently then the average of them was obtained. Meanwhile, the second way is treating each validation set as one large volume where all True Positives (TPs), True Negatives (TNs), False Positives (FPs), and False Negatives (FNs) of all volumes were added together to obtain one final value for sensitivity and one for precision.

With L = 12 layers and an embedding size of K = 768, the proposed transformer-based encoder for the UNETR model adheres to the ViT-B16  [15] design. Three different patch (feature) sizes P were experimented with values 8, 16, and 32, and the best one in practice was 16. Meanwhile, SwinUNETR-v2, on the other hand, we used the same settings and hyperparameters of the original work  [61], but for the patch (feature) size parameter, there were two values that were experimented 24 and 48. It was found that a patch size of 48 performs better than 24 by an average DSC of 0.7%. This setting provided richer multi-sequence feature representation and more accurate lesion boundary delineation, yielding a better overall sensitivity–precision balance in our validation folds. For all experiments, we employed a five-fold cross-validation method with a ratio of 80:20 (i.e., 24 volumes for training and 6 for validation) to validate and ensure the efficiency of the proposed method from any data perspective.

To mitigate overfitting, we used complementary controls at the split, data, and optimization levels rather than relying on a single mechanism. First, we enforced patient-level five-fold cross-validation (80:20), so no patient appears in both training and validation within a fold, which reduces data leakage risk. Second, as detailed in Sect. 3.2, random 3D patch extraction converts each training volume into a large pool of candidate samples; because patch centers are randomly redrawn across epochs and across 24 training subjects per fold, exact coordinate repetition is uncommon in practice, yielding high sample diversity with substantial non-redundant content. Balanced 50% foreground / 50% background sampling and on-the-fly geometric/intensity augmentation (rotation, flipping, and intensity shift) further reduce sampling bias and improve variability.

At the optimization level, AdamW with weight decay (1e-5) regularizes parameter growth throughout training. Finally, reporting fold-wise and mean results across all five folds provides a more reliable estimate of generalization than a single split and allows us to verify that performance remains stable across different patient subsets.

Loss function

Both loss techniques—soft Dice loss  [64] and voxel-wise cross-entropy loss—were combined using fixed weights and :

1

where the Dice loss is defined as:

2

and the cross-entropy loss as:

3

Here, represents the ground truth segmentation mask, the predicted probability output, the number of voxels, and the number of classes. This weighted combination balances regional overlap optimization (Dice) with voxel-wise probabilistic discrimination (cross-entropy), providing robust training for the highly imbalanced MS lesion segmentation task.

Experiments and results

This section presents a comprehensive evaluation of the proposed methodologies for MS lesion segmentation. We begin by detailing the evaluation metrics used to quantify segmentation performance. The subsequent subsections outline a series of controlled experiments designed to assess the impact of enhanced preprocessing, data augmentation, combined weighted sum losses, and multi-sequence input on segmentation accuracy. We then evaluate the performance gains achieved using advanced architectures, particularly the transition from single-sequence UNETR to multi-sequence UNETR and finally to the multi-sequence SwinUNETR-v2, demonstrating consistent improvements across all metrics. Lastly, we benchmark the best-performing model against several state-of-the-art segmentation methods under a standardized multi-sequence experimental setup, highlighting its superior performance and robustness in delineating MS lesions.

Evaluation metrics

The models were evaluated through the comparison of the predicted segmentation masks to the true ones, which, as stated earlier, have been selected by more than one expert as consensus ground truth.

Three complementary metrics were selected: Dice Similarity Coefficient (DSC), Sensitivity (True Positive Rate/Recall), and Precision (Positive Predictive Value). This choice is motivated by the highly imbalanced nature of the MS lesion segmentation task, where lesion voxels constitute only a small fraction of the total brain volume. Under such imbalance, global metrics such as pixel-wise accuracy and specificity are misleading: a trivial model that labels the entire volume as background would achieve very high accuracy and specificity, yet detect no lesions whatsoever. DSC, sensitivity, and precision instead focus exclusively on the lesion (positive) class and therefore provide clinically meaningful assessments of segmentation quality.

4

TP, FN, and FP represent the number of True Positive, False Negative, and False Positive voxels, respectively. DSC is the harmonic mean of sensitivity and precision (equivalent to the F1 score for binary segmentation) and is the standard benchmark metric in medical image segmentation, providing a single balanced measure of overlap between predicted and ground truth lesion masks. It is widely adopted in MS lesion segmentation literature, facilitating direct comparison with published works.

Sensitivity (True Positive Rate/Recall) represents the proportion of actual lesion voxels that are correctly detected, and can be computed as follows:

5

High sensitivity is especially important in a clinical context because false negatives (missed lesions) can lead to under-estimation of lesion burden, potentially delaying diagnosis or misguiding treatment decisions. Sensitivity therefore captures the model’s ability to minimize lesion omissions.

Precision (Positive Predictive Value) measures the proportion of predicted lesion voxels that are truly lesions, and can be obtained as follows:

6

Precision reflects the model’s tendency toward over-segmentation: low precision indicates a high rate of false positives, which can cause spurious lesion detections and inflate estimated lesion volumes in longitudinal monitoring. Together, sensitivity and precision capture the clinical trade-off between lesion detection completeness and segmentation specificity, complementing the holistic overlap measure provided by DSC. In addition, ROC-AUC is reported as a threshold-independent measure of discrimination to further assess lesion-versus-background separability.

Experimental results and comparative analysis

This sub-section provides a complete and comprehensive analysis in a set of different aspects as well as a brief discussion of all implemented methods to compare them using metrics DSC, sensitivity, and specificity.

Enhanced methodologies and results

In this study, we implemented two key modifications to our previous approach for MS lesion segmentation using the single-sequence UNETR version with 3D FLAIR modality input. We maintained the same model architecture, hyperparameters, imaging sequences-modalities, GPU setup, and number of iterations as in our prior research  [59] to ensure a controlled comparison. The primary modifications were:

1. Revised random patch extraction strategy: In our previous research, random patches were extracted with a 100% probability of having a lesion foreground at the patch center. In the current work, we adopt a balanced 50–50% sampling scheme where each patch has a 50% chance of lesion foreground at the center and 50% chance of non-lesion background at the center (detailed in Sect. 3.2). This reduces sampling bias and improves the model’s ability to distinguish lesions from non-lesion regions.

2. Weighted loss function: Unlike our prior work, which used only Dice loss, we now employ a weighted combination of soft Dice and cross-entropy losses from Sect. 3.5 with and . This formulation leverages Dice loss for regional overlap optimization and cross-entropy for voxel-wise probabilistic discrimination, yielding more stable training under class imbalance.

As summarized in Table 1, these two methodological enhancements achieved significantly higher DSC, sensitivity, and precision across all five validation folds compared to the baseline, confirming their effectiveness in improving lesion segmentation accuracy.

Table 1.

Comparison of performance metrics between the previous and proposed methodologies for MS lesion segmentation (%)

Validation set	Our Previous Methodology  [59]			The Newly Proposed Methodology
	DSC	Sensitivity	Precision	DSC	Sensitivity	Precision
Fold 1	76.3	54.1	60.7	81.4	70.8	58.7
Fold 2	81.7	62.4	68.1	86.5	77.1	70.6
Fold 3	80.2	60.9	64.6	82.8	62.9	73.0
Fold 4	81.1	63.8	66.3	84.9	74.7	67.7
Fold 5	71.1	43.1	50.9	73.7	44.9	58.3
Mean	78.1	56.9	62.1	81.9	66.1	65.7

Impact of multi-sequence input on segmentation performance

Building on the conclusions from the previous subsection, we conducted an additional experiment incorporating multi-sequence input consisting of FLAIR, T1-w, T1-wKS, and T2-w modalities. This multi-sequence input was used to train the UNETR model, extending the original single-sequence UNETR framework to learn jointly from all four sequences. This approach significantly enhanced the model’s ability to extract and integrate complementary information across sequences, leading to more precise lesion characterization.

In order to determine the effect of this improvement, we compared the performance of the single-sequence UNETR and multi-sequence UNETR models using both volume-wise and set-wise evaluation metrics. As shown in Table 2, the multi-sequence UNETR consistently outperforms the single-sequence UNETR across all folds in terms of DSC, sensitivity, and precision in the volume-wise comparison. This indicates the model’s improved ability to delineate lesions at the volume level.

Table 2.

Performance comparison of single-sequence versus multi-sequence input for MS lesion segmentation: volume-wise method in (%)

Validation set	Single-Sequence UNETR			Multi-Sequence UNETR
	DSC	Sensitivity	Precision	DSC	Sensitivity	Precision
Fold 1	81.4	70.8	58.7	85.2	77.0	65.8
Fold 2	86.5	77.1	70.6	87.6	79.4	72.3
Fold 3	82.8	62.9	73.0	86.8	77.5	71.1
Fold 4	84.9	74.7	67.7	87.0	77.3	72.0
Fold 5	73.7	44.9	58.3	79.1	57.6	61.1
Mean	81.9	66.1	65.7	85.1	73.8	68.5

Similarly, the results in Table 3 demonstrate that the multi-sequence UNETR also achieves superior results in the set-wise evaluation, with notable improvements in sensitivity and precision, suggesting better lesion detection and fewer false positives as well as false negatives. These consistent gains across both evaluation methods highlight the effectiveness of the multi-sequence input strategy and confirm the superiority of the multi-sequence UNETR over its single-sequence counterpart in capturing complex MS lesion characteristics and improving segmentation outcomes.

Table 3.

Performance comparison of single-sequence versus multi-sequence input for MS lesion segmentation: set-wise method in (%)

Validation set	Single-Sequence UNETR			Multi-Sequence UNETR
	DSC	Sensitivity	Precision	DSC	Sensitivity	Precision
Fold 1	81.4	84.6	67.0	85.2	85.1	72.7
Fold 2	86.5	86.8	77.1	87.6	86.4	80.2
Fold 3	82.8	74.2	80.9	86.8	85.8	78.8
Fold 4	84.9	86.3	76.1	87.0	85.2	79.9
Fold 5	73.7	62.4	71.6	79.1	70.4	75.1
Mean	81.9	78.9	74.5	85.1	82.6	77.3

Advancing multiple sclerosis lesions segmentation with multi-sequence SwinUNETR-v2

Building upon the findings from the previous subsection, we conducted an additional experiment to evaluate the performance of the multi-sequence SwinUNETR-v2 architecture compared to the previously adopted multi-sequence UNETR model. All imaging sequences (FLAIR, T1-w, T1-wKS, and T2-w), hyperparameters, number of iterations, GPU setup, and other experimental configurations were kept consistent to ensure a fair comparison.

The primary difference in this experiment lies in the network design. While the multi-sequence UNETR architecture had already demonstrated notable improvements over the single-sequence UNETR in multi-sequence settings, our evaluation reveals that the multi-sequence SwinUNETR-v2 architecture achieves even more substantial gains in segmentation performance.

As shown in Table 4, which presents volume-wise evaluation results, the multi-sequence SwinUNETR-v2 clearly outperforms multi-sequence UNETR across all validation folds in terms of DSC, sensitivity, and precision. These results underscore the model’s superior volume-level lesion delineation capabilities.

Table 4.

Performance comparison of multi-sequence UNETR versus multi-sequence SwinUNETR-v2 for MS lesion segmentation: volume-wise method in (%)

Validation set	Multi-Sequence UNETR			Multi-Sequence SwinUNETR-v2
	DSC	Sensitivity	Precision	DSC	Sensitivity	Precision
Fold 1	85.2	77.0	65.8	87.9	86.5	69.4
Fold 2	87.6	79.4	72.3	90.7	81.9	81.5
Fold 3	86.8	77.5	71.1	88.5	79.9	74.9
Fold 4	87.0	77.3	72.0	89.6	82.1	77.1
Fold 5	79.1	57.6	61.1	84.8	70.2	69.7
Mean	85.1	73.8	68.5	88.3	80.1	74.5

Furthermore, the set-wise evaluation shown in Table 5 reinforces this trend, with multi-sequence SwinUNETR-v2 consistently achieving higher scores across all metrics. The improvements are particularly evident in DSC, sensitivity, and precision, indicating enhanced lesion detection and reduced false positives and false negatives.

Table 5.

Performance comparison of multi-sequence UNETR versus multi-sequence SwinUNETR-v2 for MS lesion segmentation: set-wise method in (%)

Validation set	Multi-Sequence UNETR			Multi-Sequence SwinUNETR-v2
	DSC	Sensitivity	Precision	DSC	Sensitivity	Precision
Fold 1	85.2	85.1	72.7	87.9	86.7	80.9
Fold 2	87.6	86.4	80.2	90.7	86.6	85.8
Fold 3	86.8	85.8	78.8	88.5	86.2	85.2
Fold 4	87.0	85.2	79.9	89.6	88.3	82.7
Fold 5	79.1	70.4	75.1	84.8	80.0	79.4
Mean	85.1	82.6	77.3	88.3	85.6	82.8

Sensitivity analyses for multi-sequence SwinUNETR-v2

To further validate the design choices of our multi-sequence SwinUNETR-v2 framework, we conducted three ablation studies addressing the contribution of individual MRI sequences, patch sampling ratios, and loss function weights. All experiments used identical training settings and five-fold cross-validation as in the main results.

Modality contribution analysis: We performed a leave-one-modality-out ablation to quantify each sequence’s impact. Results in Table 6 show that FLAIR provides the strongest contribution (largest drop when removed), consistent with its role in lesion conspicuity, while T1-w, T1-wKS, and T2-w provide complementary structural information.

Table 6.

Ablation study of individual MRI modality contribution to multi-sequence SwinUNETR-v2 (mean over five folds in (%))

Configuration	DSC	Volume-Wise		Set-Wise
		Sensitivity	Precision	Sensitivity	Precision
Full model	88.3	80.1	74.5	85.6	82.8
w/o T1-w	86.6	78.5	72.2	84.8	80.5
w/o T1-wKS	86.8	74.9	73.3	82.8	80.9
w/o FLAIR	78.5	59.1	57.6	69.4	66.3
w/o T2-w	86.5	77.1	68.9	83.8	76.8

Patch sampling sensitivity: We tested alternative foreground/background ratios (40/60, 60/40) against the baseline 50/50. Table 7 shows the 50/50 ratio achieves optimal balance for small lesion detection.

Table 7.

Impact of patch sampling ratios on SwinUNETR-v2 (mean over five folds in (%))

Configuration	DSC	Volume-Wise		Set-Wise
		Sensitivity	Precision	Sensitivity	Precision
40/60 FG/BG	86.8	79.4	70.8	86.1	80.4
50/50 FG/BG	88.3	80.1	74.5	85.6	82.8
60/40 FG/BG	87.0	75.3	74.3	83.9	81.7

Loss weight validation: We tested alternative weight combinations (, ) against the baseline (, ). Table 8 confirms the selected weights optimize DSC and sensitivity under class imbalance.

Table 8.

Impact of loss function weights on SwinUNETR-v2 (mean over five folds in (%))

Configuration	DSC	Volume-Wise		Set-Wise
		Sensitivity	Precision	Sensitivity	Precision
,	87.0	78.8	71.8	85.2	81.1
,	88.3	80.1	74.5	85.6	82.8
,	86.9	78.9	70.9	85.6	80.1
,	87.1	79.6	70.8	86.6	79.9
,	87.0	82.5	73.0	87.3	80.7

These ablation studies provide quantitative validation of our design choices: FLAIR contributes most critically to lesion conspicuity (9.8% DSC degradation without it, vs. 1.5–1.8% for other modalities); 50/50 foreground/background sampling maximizes DSC by 1.3–1.5% relative to alternatives; and , optimizes overlap/discrimination balance (1.2–1.4% DSC superiority). Collectively, these findings demonstrate the framework’s sensitivity to modality composition, sampling equilibrium, and loss weighting, establishing the multi-sequence SwinUNETR-v2 configuration as empirically optimal for MS lesion segmentation under data imbalance and limited cohort size.

Table 9 summarizes the progressive improvements across experiments. From single-sequence baseline to multi-sequence SwinUNETR-v2, we achieved +10.2% DSC (78.188.3%), volume-wise sensitivity +23.2% (56.980.1%), volume-wise precision +12.4% (62.174.5%), set-wise sensitivity +6.7% (78.985.6%), and set-wise precision +8.3% (74.582.8%). This comprehensive gain profile (+6.7–23.2% across metrics/evaluation modes) underscores the synergistic impact of multi-sequence fusion, balanced sampling, and weighted loss optimization, establishing a robust foundation for external benchmarking.

Table 9.

Summarized average scores obtained in the previous three experiments using both techniques, volume-wise and set-wise in (%)

Experiment	DSC	Volume-Wise		Set-Wise
		Sensitivity	Precision	Sensitivity	Precision
Old Methodology + Single-Sequence UNETR (FLAIR)	78.1	56.9	62.1	-	-
New Methodology + Single-Sequence UNETR (FLAIR)	81.9	66.1	65.7	78.9	74.5
New Methodology + Multi-Sequence UNETR	85.1	73.8	68.5	82.6	77.3
New Methodology + Multi-Sequence Swin-UNETR-v2	88.3	80.1	74.5	85.6	82.8

To provide a threshold-independent evaluation, Fig. 6 reports ROC analyses across all five folds for both volume-wise and set-wise configurations, together with AUC heatmaps summarizing per-volume and per-fold behavior. The model achieved a mean ROC-AUC of 86.56% in the volume-wise setting and 90.80% in the set-wise setting, indicating robust discriminative performance with improved stability when predictions are aggregated at the set level.

Fig. 6.

ROC analysis of multi-sequence SwinUNETR-v2 over five-fold cross-validation. The top row shows fold-wise volume-wise ROC curves, the bottom row shows fold-wise set-wise ROC curves, and the right-side heatmaps summarize AUC values across volumes and folds. Mean ROC-AUC reaches 86.56% (volume-wise) and 90.80% (set-wise), confirming strong lesion-versus-background discrimination under both evaluation protocols

Comparative evaluation of SwinUNETR-v2 with established methodologies

In this section, we extend our investigation by comparing the performance of multi-sequence SwinUNETR-v2 with other well-established methodologies for MS lesion segmentation. Our aim is to evaluate whether integrating the multi-sequence concept into SwinUNETR-v2 within the proposed methodology offers advantages over previously reported approaches.

• Selection of comparative methodologies

The decision to use multi-sequence SwinUNETR-v2 in this comparative analysis is based on our sequential experimental findings. In Sect. 4.2.1, we introduced a novel methodology and demonstrated its superiority over our previously implemented approach. In Sect. 4.2.2, we evaluated the impact of single-sequence versus multi-sequence concepts, concluding that the multi-sequence framework significantly enhances segmentation performance. This led us to Sect. 4.2.3, where we compared multi-sequence UNETR with multi-sequence SwinUNETR-v2, ultimately demonstrating that multi-sequence Swin-UNETR-v2 provides superior segmentation accuracy. Consequently, in this section, we extend our analysis by benchmarking multi-sequence SwinUNETR-v2 against other well-established methodologies in the field.

To ensure a meaningful comparison, we selected AttentionUNet  [65], DenseResidualUNet  [47], Fully Convolutional Neural Network (FCNN)  [66], SegResNet  [67], and nnUNet-v2  [68] as benchmark methodologies. The rationale behind choosing these models is not merely their widespread use, but rather the fact that they have been previously implemented and presented in peer-reviewed research for medical image segmentation; note that SegResNet was originally introduced for brain tumor segmentation and nnUNet-v2 represents a strong self-configuring baseline widely adopted in biomedical segmentation benchmarks. Our selection was guided by the findings of our literature review  [69], a comprehensive review paper that examined the most prominent literature on MS lesions segmentation and classification over the past five years. This review highlighted these methodologies as among the most effective and recent approaches explored in the field.

Since these models were originally designed and evaluated for MS lesions segmentation, we aimed to assess their performance under a standardized experimental setup. To ensure fairness, all methodologies were tested on the same dataset and implemented using the multi-sequence concept, thereby maintaining consistency in data representation and experimental conditions.

• Experimental setup

The experimental setup preserves the same dataset, patient-level five-fold split protocol, and evaluation definitions used in our previous analyses. To maintain methodological fidelity, method-specific training and preprocessing components were retained where they are integral to the original framework. Specifically:

1. AttentionUNet, FCNN, and DenseResidualUNet were implemented using their original configurations as proposed by their respective authors.

2. SegResNet, in contrast, was trained using the same loss function, optimizer, and preprocessing techniques as those employed in SwinUNETR-v2, ensuring a direct performance comparison.

3. nnUNet-v2 was implemented using its official planning, preprocessing, training, and inference pipeline while preserving the same patient-level five-fold split protocol and evaluation definitions used throughout this work.

All models were trained and tested using the same dataset under the multi-sequence concept, with hyperparameters optimized based on their respective original implementations.

• Analysis of results

The performance comparison across the selected methodologies is conducted based on DSC, sensitivity, and precision, which have been consistently used throughout the previous experiments. For each methodology, we report the mean values computed over the five-fold cross-validation, ensuring a reliable assessment of segmentation performance.

Table 10 provides a quantitative comparison of six multi-sequence based methodologies for segmentation of MS lesions, reporting their average DSC, sensitivity, and precision over five cross-validation folds in both volume-wise and set-wise evaluations. Notably, multi-sequence SwinUNETR-v2 achieves the highest DSC (88.3%) and the best sensitivity in both evaluation modes, indicating more reliable lesion detection and fewer missed lesion voxels across folds. Although nnUNetv2 attains slightly higher precision in both evaluation schemes, the absolute gains are limited (0.5% volume-wise and 0.3% set-wise), whereas SwinUNETR-v2 shows clearer improvements in sensitivity (4.1% volume-wise and 1.9% set-wise). This pattern suggests that SwinUNETR-v2 provides a stronger overall detection-oriented profile for MS lesion segmentation, where reducing false negatives is clinically important.

Table 10.

Quantitative comparison of multiple methodologies for MS lesion segmentation using the multi-sequence concept. Results are reported as the mean values over five cross-validation folds for both volume-wise and set-wise evaluations, based on DSC, sensitivity, and precision in (%)

Methodology	DSC	Volume-Wise		Set-Wise
		Sensitivity	Precision	Sensitivity	Precision
SegResNet	85.6	72.6	72.2	83.0	78.6
FCNN	85.0	74.9	67.2	84.8	75.2
AttentionUNet	86.2	76.7	69.5	85.3	76.0
DenseResidualUNet	81.6	67.0	62.1	78.0	71.0
nnUNet-v2	88.1	76.0	75.0	83.7	83.1
SwinUNETR-v2	88.3	80.1	74.5	85.6	82.8

Figure 7 illustrates the trade-off between computational cost (FLOPs) and segmentation accuracy (DSC) across different methodologies. All FLOPs and parameter count for each model were pre-calculated using the fvcore  [70] library from Facebook Research, ensuring consistent measurement and enabling a fair comparison of computational demands. While models such as SegResNet (16.24 GFLOPs, 85.8% DSC) and FCNN (21.48 GFLOPs, 85% DSC) prioritize efficiency, they do not achieve top segmentation accuracy. nnUNet-v2 attains near-top DSC (88.1%) but at the highest computational cost (545.41 GFLOPs), whereas SwinUNETR-v2 achieves the best DSC (88.3%) with substantially lower cost (348.93 GFLOPs), indicating a stronger accuracy-efficiency balance.

Fig. 7.

Comparison of computational cost (GFLOPs) versus segmentation accuracy (DSC) for multiple sclerosis lesions segmentation methods. The figure shows that low-cost models such as SegResNet and FCNN have lower DSC, while nnUnet-v2 reaches near-top DSC at the highest GFLOPs. SwinUNETR-v2 achieves the highest DSC (88.3%) with lower computational cost than nnUnet-v2, supporting a favorable accuracy-efficiency trade-off

In contrast, multi-sequence DenseResidualUNet obtains the lowest DSC (81%), rendering it less favorable for this task. Meanwhile, multi-sequence SegResNet and multi-sequence FCNN yield moderate performance, and multi-sequence AttentionUNet shows slightly better DSC but does not surpass SwinUNETR-v2. Overall, these findings underscore the effectiveness of multi-sequence SwinUNETR-v2 for MS lesions segmentation, particularly when high DSC and robust lesions delineation are paramount for clinical applications.

Figure 8 presents a qualitative comparison of the segmentation results obtained by the proposed multi-sequence SwinUNETR-v2 model using three different patient samples (three rows). Row 1 is dominated by larger and medium lesions; Row 2 represents a mixed lesion scenario containing relatively large, medium, and small lesions; and Row 3 is mainly composed of small lesions. Across all three samples, the predicted masks remain closely aligned with the ground truth. In particular, the third row demonstrates strong delineation of small lesions, which is a clinically crucial challenge in MS identification and segmentation. In each row, column (a) displays the skull-stripped FLAIR slice, column (b) the ground truth mask, column (c) the predicted mask, and column (d) the error map with false positives in red and false negatives in green. This figure complements the quantitative trends in Table 10 and provides visual evidence of robust lesion-boundary quality across lesion-size regimes.

Fig. 8.

Qualitative comparison on three representative patient samples (rows): row 1 is dominated by large/medium lesions, row 2 includes mixed lesion sizes (large, medium, and small), and row 3 is dominated by small lesions. Columns show (a) skull-stripped FLAIR slice, (b) ground truth MS lesion mask, (c) predicted mask from multi-sequence SwinUNETR-v2, and (d) error map with false positives in red and false negatives in green

Beyond score-level comparison, two mechanistic factors help explain the observed gains. First, using FLAIR as the reference for foreground cropping is advantageous because FLAIR suppresses CSF signal and increases lesion conspicuity relative to surrounding tissue, yielding a more stable brain-versus-background boundary for downstream multi-sequence alignment and patch extraction. This reduces the chance that non-brain voxels dominate sampled context and improves consistency of coordinates propagated to the other sequences. Second, SwinUNETR-v2 is better matched to small-lesion detection than UNETR because its hierarchical shifted-window attention and stage-wise residual convolution blocks preserve fine local boundaries while still aggregating broader context across stages. In contrast, the flatter UNETR tokenization path is more prone to attenuating tiny lesion cues when global representations dominate. The combined local-detail bias and multi-scale decoding in SwinUNETR-v2 is consistent with the higher sensitivity observed in Table 4 and  5 and with the stronger small-lesion delineation visible in Row 3 of Fig. 8.

MS lesion segmentation is a high-stakes task, and the proposed framework presents several limitations and potential risks. First, multi-sequence 3D SwinUNETR-v2 has substantial computational overhead (348.93 GFLOPs), which increases training/inference time and may limit practical use in resource-constrained environments or time-critical workflows.

Second, there are potential points of failure across the pipeline. The method is sensitive to preprocessing quality (skull stripping, registration/reorientation, foreground cropping, and patch extraction); errors in these stages can propagate to the final mask and increase false positives or false negatives. The framework also assumes all four MRI sequences are available and of adequate quality, whereas missing or corrupted modalities in real clinical practice can degrade performance and robustness.

Third, there is a measurable trade-off between segmentation accuracy and computational efficiency. As shown in Fig. 7, SwinUNETR-v2 attains the highest DSC (88.3%) with lower compute than nnUNet-v2 (348.93 vs. 545.41 GFLOPs), but it remains considerably more expensive than lightweight alternatives such as SegResNet and FCNN. Therefore, model selection may depend on deployment priorities: maximizing lesion detection performance versus minimizing computational cost.

From a clinical decision-support perspective, the model is intended to function as an assistive tool rather than a standalone diagnostic system. Its practical added value is to provide faster, more reproducible lesion delineation and lesion-burden quantification that can support radiologist review, follow-up comparison, and treatment-response tracking. However, error modes must be considered explicitly in clinical use: false positives may overestimate lesion burden and could trigger unnecessary follow-up or overtreatment, whereas false negatives may miss active or small lesions and could delay escalation of therapy. Therefore, model outputs should be interpreted with expert oversight, ideally with uncertainty flags and structured quality checks before being used for downstream clinical decisions.

Finally, despite strong five-fold cross-validation results, the dataset is relatively small (30 subjects) and from a single public source. External generalization across centers, scanners, and acquisition protocols is not yet fully established. Future work should prioritize multi-center external validation, robustness under missing-modality scenarios, uncertainty-aware quality control, and computational optimization (e.g., pruning, distillation, or efficient variants) to reduce runtime cost while preserving segmentation quality.

In future work, we will further optimize this accuracy-efficiency balance to improve real-world clinical deployability.

Conclusion

This work demonstrates that multi-sequence MRI fusion with the SwinUNETR-v2 backbone, augmented by balanced patch sampling and weighted loss optimization, provides strong MS lesion segmentation performance. Achieving a mean DSC of 88.3% on the 3D-MR-MS dataset, with improvements ranging from 0.2% to 6.7% over the compared methods, the framework effectively leverages complementary information across T1-w, contrast enhanced T1-w, FLAIR, and T2-w sequences to delineate lesions with high fidelity.

While computational demands remain elevated compared to single-sequence baselines, the accuracy gains indicate potential utility for precision diagnostics where lesion volumetry informs treatment planning. However, clinical adoption should be considered preliminary until external multi-center studies confirm robustness across scanners, protocols, and patient populations. Current limitations include dataset scale and inference efficiency; future work will target computational optimization (e.g., pruning and knowledge distillation) alongside external validation.

Clinical significance: By reducing false negatives in small/peripheral lesions and minimizing over-segmentation, our approach enhances prognostic accuracy and supports personalized MS management protocols.

Author contributions

R.M.T. is the main contributor. He conducted all the surveys and experiments, wrote the manuscript text and references, and prepared Figures 1, 2, 3, 4, 5, 6, 7, and 8 and Tables 1, 2, 3, 4, 5, 6, 7, 8, 9, and 10. D.K., M.A.B., and H.A.S. supervised the work, providing guidance on experimental design, technical structure, and the overall presentation of the manuscript. All authors reviewed and approved the final manuscript.

Funding

Open access funding provided by The Science, Technology & Innovation Funding Authority (STDF) in cooperation with The Egyptian Knowledge Bank (EKB).

Data availability

The datasets analysed during the current study are publicly available on the Laboratory of Imaging Technologies (LIT) project website at https://lit.fe.uni-lj.si/en/research/resources/3D-MR-MS/.

Declarations

Ethics approval and consent to participate

All methods were carried out in accordance with relevant guidelines and regulations. This study exclusively used the fully anonymized, publicly available 3D MR-MS database from the Laboratory of Imaging Technologies (LIT). Ethical approval for the original data collection was obtained from the National Medical Ethics Committee of Slovenia (approval reference: 47/03/14), and written informed consent was acquired from all participants. As a retrospective analysis of pre-existing de-identified data, no additional institutional ethics approval was required for this study.

Consent for publication

Not Applicable. This study uses fully anonymized, de-identified data, and no identifying images or personal clinical details of participants are presented.

Competing interests

The authors declare no competing interests.