ISLA: A U-Net for MRI-based acute ischemic stroke lesion segmentation with deep supervision, attention, domain adaptation, and ensemble learning

TL;DR

ISLA addresses AIS lesion segmentation in diffusion MRI by systematically optimizing a 3D U-Net-based framework across loss functions, deep supervision, residual vs standard blocks, and attention modules. It further enhances generalization through unsupervised domain adaptation using a Mean Teacher setup and improves robustness via ensemble learning. On internal validation, deep supervision and attention yield the best base-model performance (ISLA-B), while combining multiple base models (ISLA-ENS) delivers the strongest external-test results, outperforming DAGMNet and DeepISLES across metrics and lesion sizes. The approach demonstrates strong cross-site robustness and artifact tolerance, with publicly available code and models to support reproducibility and multi-center deployment. Overall, ISLA advances AIS lesion segmentation by combining targeted architectural choices with domain adaptation and ensembling to achieve reliable, clinically applicable performance.

Abstract

Accurate delineation of acute ischemic stroke lesions in MRI is a key component of stroke diagnosis and management. In recent years, deep learning models have been successfully applied to the automatic segmentation of such lesions. While most proposed architectures are based on the U-Net framework, they primarily differ in their choice of loss functions and in the use of deep supervision, residual connections, and attention mechanisms. Moreover, many implementations are not publicly available, and the optimal configuration for acute ischemic stroke (AIS) lesion segmentation remains unclear. In this work, we introduce ISLA (Ischemic Stroke Lesion Analyzer), a new deep learning model for AIS lesion segmentation from diffusion MRI, trained on three multicenter databases totaling more than 1500 AIS participants. Through systematic optimization of the loss function, convolutional architecture, deep supervision, and attention mechanisms, we developed a robust segmentation framework. We further investigated unsupervised domain adaptation to improve generalization to an external clinical dataset. ISLA outperformed two state-of-the-art approaches for AIS lesion segmentation on an external test set. Codes and trained models will be made publicly available to facilitate reuse and reproducibility.

Figures (12)

• Figure 1: Proportions of stroke lesions across brain regions and datasets. The manual lesion delineations were registered in a common space (Section \ref{['meth:preprocessing']}) and the reference trace diffusion-weighted image is shown in background. The proportions masks were smoothed with a gaussian filter (FWHM = 4 mm) and values below 0.02 were masked.
• Figure 1: Illustration of the implemented attention gate modules.$x_d$ denotes the decoder feature maps, $x_e$ denotes the encoder feature maps, and $x_m$ the output modulated feature maps.. $Conv[k_1 \times k_2 \times k_3, s_1 \times s_2 \times s_3, f]$ represents a convolution with $f$ filters, kernel size $k_1 \times k_2 \times k_3$, and stride $s_1 \times s_2 \times s_3$. $TLU[u_1 \times u_2 \times u_3]$ denotes trilinear upsampling with factors $u_1 \times u_2 \times u_3$.
• Figure 1: Cases from the test set with DAGMNet falses positives in the posterior region of the brain. Each column represents a different participant. The trace diffusion-weighted image and the segmentation masks were registered in a common space (Section \ref{['meth:preprocessing']}).
• Figure 1: Illustration of attention coefficients computed by ISLA-B during inference on a test-set case. The coefficients were extracted from the AGs modules. att$_i$ denotes the coefficients at the i-th U-Net level (from higher to lower resolution). Trilinear upsampling was applied to make attention maps match the original image resolution (not for att$_1$).
• Figure 1: An illustrative case of DeepISLES and ISLA-ENS segmentations from the test set. Axial slices are shown from the volume corresponding to the median difference in DSC between the two models across the test set. The trace diffusion-weighted image and the segmentation masks were registered in a common space (Section \ref{['meth:preprocessing']}).