Pseudo-RGB Slice Stacking in 2D ResUNet for High-Sensitivity Multiple Sclerosis Lesion Segmentation

Background: Multiple sclerosis (MS) is a chronic autoimmune demyelinating disease of the central nervous system, affecting more than 2.8 million individuals worldwide. Automated segmentation of white matter lesions on fluid-attenuated inversion recovery (FLAIR) magnetic resonance images (MRI) is essential for reproducible diagnosis and treatment monitoring, yet remains challenging due to extreme class imbalance, high lesion load variability, and poor contrast at lesion boundaries. Method: We propose a 2D U-Net with a ResNet50 encoder that exploits ImageNet-pretrained representations through a novel pseudo-RGB input strategy: three consecutive FLAIR slices centred on the target slice are stacked channel-wise to form a three-channel input, recovering inter-slice spatial context while enabling direct reuse of pretrained convolutional weights without modality-specific pretraining. A two-phase transfer-learning protocol first optimises only the decoder with the encoder frozen, then fine-tunes the upper encoder blocks at a reduced learning rate. Test-time augmentation (TTA) averaging over horizontal-flip and vertical-flip transformations further improves prediction robustness. Results: Evaluation on the held-out test set of the MSLesSeg2024 benchmark (12 patients, approximately 1650 axial slices) shows that the proposed model achieves a Dice similarity coefficient (DSC) of 0.714 (95% confidence interval (CI): 0.6845-0.7194), intersection-over-union (IoU) of 0.6571 (95% CI: 0.6272-0.6632), and area under the receiver operating characteristic (ROC) curve (AUC) of 0.9628 (95% CI: 0.9422-0.9793). Critically, the model records the lowest false-negative pixel count per slice (FNV = 33.6 px/slice) across all ablation conditions, indicating superior sensitivity to lesion tissue; this is a property of direct clinical relevance for MS monitoring, where missed lesions carry the greatest diagnostic risk. A patient-matched comparison against a 3D nnU-Net baseline shows statistically comparable DSC (0.714 vs. 0.726; paired Wilcoxon p=0.68-0.79, not significant) alongside a substantially higher pixel-level AUC for the proposed model (0.963 vs. 0.773) and a true volumetric Hausdorff distance gap smaller than an earlier estimate (12.5 mm vs. 10.7 mm), showing an honest mixed-strengths result rather than an outright improvement. A supplementary ablation further shows that replicating a single FLAIR slice across all three channels significantly outperforms the pseudo-RGB adjacent-slice encoding (p<0.001), indicating that the anticipated inter-slice-context benefit did not materialise here (see Discussion section). Cross-dataset evaluation on the independent MSSEG 2016 benchmark confirms generalisability: zero-shot transfer achieves DSC = 0.6562, recovering to DSC = 0.7046 after brief fine-tuning (30 epochs), within 1.5 percentage points of in-domain performance. Conclusions: The present study reveals that an optimized, lightweight 2D pipeline can rival the segmentation overlap of context-aware 3D baselines on specific datasets, doing so with a significantly reduced computational footprint. Given its strong pixel-wise discrimination, this methodology offers an effective and practical tool for routine automated MS lesion assessment in clinical settings.