Abstract
Medical image segmentation often struggles with ambiguous tissue boundaries where manual "hard" labels (0 or 1) fail to capture inherent uncertainty. This paper introduces a method to "soften" these labels using superpixels and distance functions, allowing models to learn from probabilistic boundaries and achieving superior accuracy on both 2D and 3D medical datasets.
The Challenge of Hard Labels in Medical Imaging
In medical imaging, "ground truth" annotations are rarely definitive. Experts segmenting organs or lesions in MRI or OCT scans face boundaries blurred by the partial volume effect—where a single pixel encompasses multiple tissue types—or low image contrast.
Standard machine learning approaches rely on "hard labels," forcing a binary classification (e.g., 100% tumor vs. 100% healthy). When boundaries are ambiguous, these hard labels introduce noise and uncertainty into training, often resulting in poor generalization and "jittery" segmentation edges.
Superpixel-Guided Label Softening
To address this, the authors propose a Superpixel-Guided Label Softening strategy. This method leverages the image's local structure to guide the softening process, ensuring that uncertainty is modeled in a context-aware manner.
The Pipeline
- Superpixel Generation: The SLIC algorithm groups pixels into "superpixels"—locally homogeneous clusters that respect natural edges.
- Identifying Uncertainty: Superpixels that intersect an annotation boundary are flagged as zones of high uncertainty.
- Label Softening via Distance: For pixels within these uncertain zones, a signed distance to the annotation boundary is calculated and converted into a probability value (0 to 1). A pixel exactly on the boundary is assigned a value of 0.5.
- Joint Training: A U-Net variant is trained using a combination of standard losses (Cross-Entropy and Dice) and a Kullback-Leibler (KL) divergence loss to align model predictions with these soft labels.
Experimental Evaluation
The method was validated on two distinct medical datasets:
- MRBrainS18: A 3D multi-sequence brain MRI dataset for segmenting gray matter, white matter, and cerebrospinal fluid.
- OCT-DME: A 2D Optical Coherence Tomography dataset for segmenting retinal layers and fluid regions in patients with diabetic macular edema.
Key Results
The proposed method consistently outperformed baseline "hard label" models and traditional softening techniques (like Gaussian blur):
- MRBrainS18 (3D): Achieved a Dice score of 85.63% and significantly improved the 95th percentile Hausdorff distance (HD95), reducing it from 3.52mm to 3.11mm.
- OCT-DME (2D): Improved the Dice score from 82.50% to 83.94%.
Qualitative analysis shows that this method produces smoother, more anatomically plausible boundaries compared to the jagged results of baseline models.
Discussion and Future Work
The core innovation of this research is using superpixels to ensure softening is driven by image content. By following actual image gradients, the softening respects underlying anatomy. However, the method is sensitive to the scale of superpixels and the weight of the KL loss ($\beta$). Future work may explore automated hyperparameter optimization or deeper integration of the softening process into neural architectures.
Summary updated