Abstract
Training deep learning models for medical imaging typically necessitates thousands of hand-labeled images, a process that is both resource-intensive and time-consuming. This paper presents a novel self-supervised, few-shot learning framework designed to segment ischemic stroke lesions using as little as one annotated patient sample. By leveraging domain-specific parametric maps to automate the generation of training data, the proposed method addresses the data scarcity bottleneck inherent in clinical neuroimaging.
Introduction
Ischemic stroke remains a leading cause of global mortality and long-term disability. Rapid and precise identification of the infarct core—the area of the brain where blood flow is critically obstructed—is essential for determining treatment eligibility and predicting patient outcomes. While deep learning has demonstrated significant potential in medical image analysis, standard supervised models are notoriously "data-hungry," requiring massive datasets of brain scans manually delineated by expert neuroradiologists.
In the clinical domain, such high-quality annotated data is scarce. Furthermore, stroke lesions exhibit high variability in size, shape, and location, making generalization difficult for models trained on limited datasets. This research explores Few-Shot Segmentation (FSS), a paradigm where a model learns to segment novel images after being exposed to only a minimal number of labeled examples.
Methodology
The researchers developed a framework based on the ADNet architecture, introducing a specialized self-supervised task to pre-train the model before it encounters any manual labels.
1. Self-Supervision via Supervoxels
To facilitate training without manual annotations, the model generates its own "pseudolabels." This is achieved by grouping similar neighboring pixels into 3D clusters known as supervoxels. During the self-supervised phase, the model selects a specific supervoxel as a "target" and attempts to identify the corresponding shape in a different slice of the same brain volume, thereby learning spatial and structural representations.
2. Integration of Parametric Maps (PMs)
A primary innovation of this work is the use of Parametric Maps (PMs) for supervoxel generation. Standard self-supervised models often rely on raw CT scans, which can be homogeneous and lack the contrast necessary to define distinct physiological boundaries. By using PMs—color-coded images such as Cerebral Blood Flow (CBF) or Time-to-Maximum (Tmax)—the model leverages data that already highlights physiological changes. These maps act as a form of domain knowledge, providing a clearer signal for the self-supervised task than raw intensity data.
3. Prototypical Learning
The model employs a prototypical learning strategy. Given a "support" image (the single labeled example), the network extracts a feature "prototype" representing the lesion. This prototype is then compared against a "query" image (the target for segmentation). The model classifies pixels in the query image based on their similarity to the lesion prototype in the feature space.
Experimental Setup
The study utilized a dataset of 152 Computed Tomography Perfusion (CTP) scans acquired from Stavanger University Hospital between 2014 and 2020. The data was categorized into three subsets:
- CTP-ALL: The complete dataset of 152 subjects.
- CTP-LVO: A subset of 77 patients with Large Vessel Occlusion, characterized by more severe strokes and larger, more visible lesions.
- WIS: A control group of 15 patients exhibiting stroke-like symptoms but no confirmed ischemic stroke.
Results and Discussion
The proposed method demonstrated that domain-specific guidance significantly enhances few-shot performance.
Performance Metrics
Using only one annotated patient for training, the model achieved an average Dice score of 0.58 for LVO patients. The Dice score, which measures the overlap between the predicted segmentation and the expert ground truth, indicated that the model significantly outperformed two primary baselines: a "CTP-Baseline" (using raw data for self-supervision) and a "PMs-Baseline" (using PMs as input but not for pseudolabel generation).
Key Insights
- Domain Knowledge as a Catalyst: The results suggest that using processed physiological maps to guide self-supervision is more effective than relying on raw data. PMs provide a structural "shortcut" that allows the model to grasp relevant features more efficiently.
- Sensitivity to Severity: The model performed substantially better on Large Vessel Occlusions (LVO) than on minor strokes. Larger lesions provide a more robust signal for the few-shot prototype, whereas smaller lesions are easily lost in the noise of healthy tissue.
- Hyperparameter Robustness: The model’s performance remained stable across different supervoxel scales (controlled by the parameter $\rho$), suggesting that the method is not overly sensitive to specific hyperparameter configurations.
Limitations and Future Work
Despite the promising results in 1-shot learning, a performance gap remains when compared to fully supervised models, which have achieved Dice scores of approximately 0.85 on similar datasets using hundreds of labeled images.
Qualitative analysis revealed a tendency toward under-segmentation, where the model is overly conservative in its lesion predictions. Furthermore, the model struggled with Non-LVO patients, where lesions are smaller and less distinct. Future research will focus on refining supervoxel generation techniques and exploring the integration of multi-modal data (such as MRI) to improve sensitivity to smaller, more subtle ischemic changes. This work represents a critical first step toward deploying AI in clinical settings where labeled data is a luxury rather than a given.
Summary updated