Abstract
Diffusion-weighted magnetic resonance imaging (DWI) is a critical diagnostic tool for acute ischemic stroke. However, manual segmentation of lesions is time-consuming and subjective, varying significantly between neurologists. Given that "time is brain," rapid and automatic identification of lesions is essential for improving patient prognosis. This paper presents a novel framework for automatic ischemic stroke lesion segmentation on DWI. Our approach introduces three key contributions: a detection and segmentation network (DSN) designed to address data imbalance; a triple-branch DSN architecture for multi-plane feature extraction; and a multi-plane fusion network (MPFN) to refine predictions. Experiments on the ISLES2015 dataset demonstrate the superiority of our method, achieving a Dice score of 62.2% and a sensitivity of 71.7%.
The Challenge of Stroke Segmentation
In the context of acute ischemic stroke, the time from onset to thrombus removal is the most significant predictor of patient outcome. While DWI is the gold standard for diagnosis, automating lesion segmentation faces three major hurdles:
- Severe Data Imbalance: Lesions are often minute compared to the total brain volume (pixel-level imbalance), and many MRI slices contain no lesions at all (slice-level imbalance).
- Image Artifacts: Non-pathological structures can mimic the appearance of stroke lesions on DWI, leading to high false-positive rates.
- Spatial Complexity: Lesions vary wildly in size, shape, and location, making standard 2D models insufficient.
Proposed Framework: DSN and MPFN
To overcome these challenges, the proposed framework utilizes a two-stage process that prioritizes detection before fine-grained segmentation.
Detection and Segmentation Network (DSN)
Instead of an end-to-end segmentation approach, which often fails when the target is small, the DSN first identifies candidate regions (bounding boxes). This "needle in a haystack" strategy focuses the network's attention on relevant areas. The DSN uses a Feature Pyramid Network (FPN) with a ResNet-101 backbone to extract multi-scale features, followed by a U-Net-based segmentation within the detected boxes.
Triple-Branch Multi-Plane Fusion
To capture 3D context without the computational burden of 3D convolutions, the DSN is applied to three orthogonal planes: Axial, Sagittal, and Coronal. The outputs from these branches are then processed by the Multi-Plane Information Fusion Network (MPFN).
The MPFN acts as a spatial refiner. By analyzing the lesion from three perspectives simultaneously, the system can effectively distinguish between true lesions (which persist across planes) and artifacts (which typically appear only in a single view).
Experimental Results and Discussion
The study utilized the ISLES2015 SSIS DWI sequence dataset. A key finding during preprocessing was that standard N4Bias correction actually hindered performance; instead, a "three-window" concatenation method was used to highlight tissue intensities.
Our method significantly outperformed existing baselines:
- Dice Score: 62.2% (vs. 56.31% for the 3D tool DeepMedic and <1% for standard 2D U-Net).
- Sensitivity: 71.7%, demonstrating robust detection of true positive lesions.
The failure of standard 2D models highlights the necessity of the detection-first approach. When the background is overwhelmingly dominant, traditional networks fail to learn lesion features. By forcing the network to detect the lesion first, we solve the pixel-level imbalance problem.
Conclusion
This multi-plane fusion framework provides a robust, "pseudo-3D" solution for stroke segmentation that is more computationally efficient than full 3D convolutions while remaining more accurate than 2D approaches. Future work will explore the integration of attention mechanisms to further mimic the diagnostic reasoning of expert neurologists, potentially increasing the transparency and clinical utility of the model.
Summary updated