Segmenting Small Stroke Lesions with Novel Labeling Strategies

TL;DR

The paper tackles the challenge of segmenting small stroke lesions in MRI, where conventional Dice-based metrics bias performance toward larger lesions. It introduces two labeling strategies, Multi-Size Labeling (MSL) and Distance-Based Labeling (DBL), that recast voxel labeling into multi-class tasks to emphasize small lesions and boundaries, complemented by an ensemble and postprocessing workflow. On the ATLAS v2.0 dataset, the MSL+DBL ensemble consistently improves recall, F1, and Dice scores over the top baseline, with a single MSL model also surpassing the baseline on mini-lesions. The methods are architecture-agnostic and code is publicly available, enabling broader adoption for precise small-lesion segmentation in clinical stroke imaging.

Abstract

Deep neural networks have demonstrated exceptional efficacy in stroke lesion segmentation. However, the delineation of small lesions, critical for stroke diagnosis, remains a challenge. In this study, we propose two straightforward yet powerful approaches that can be seamlessly integrated into a variety of networks: Multi-Size Labeling (MSL) and Distance-Based Labeling (DBL), with the aim of enhancing the segmentation accuracy of small lesions. MSL divides lesion masks into various categories based on lesion volume while DBL emphasizes the lesion boundaries. Experimental evaluations on the Anatomical Tracings of Lesions After Stroke (ATLAS) v2.0 dataset showcase that an ensemble of MSL and DBL achieves consistently better or equal performance on recall (3.6% and 3.7%), F1 (2.4% and 1.5%), and Dice scores (1.3% and 0.0%) compared to the top-1 winner of the 2022 MICCAI ATLAS Challenge on both the subset only containing small lesions and the entire dataset, respectively. Notably, on the mini-lesion subset, a single MSL model surpasses the previous best ensemble strategy, with enhancements of 1.0% and 0.3% on F1 and Dice scores, respectively. Our code is available at: https://github.com/nadluru/StrokeLesSeg.

Figures (7)

• Figure 1: (a) Overview of our method. We first propose Multi-Size Labeling (MSL) and Distance-Based Labeling (DBL). These two labeling strategies accentuate small lesions and lesion boundaries, respectively. Then, we ensemble models trained with MSL and DBL. Following, post-processing is applied to generate the final segmentation mask. (b) Proposed ensemble strategy. A linear interpolation between MSL and DBL results is applied for small lesions to generate ensemble results. For large lesions, we exclusively rely on DBL results. (c) Proposed postprocessing strategy. The postprocessing pipeline is designed to enhance segmentation accuracy by filtering out small lesions whose maximum probability falls below a predefined threshold.
• Figure 2: Visualization of segmentation results. The green contours represent the ground truth, while the red contours depict the predicted lesions. These segmentation results indicate that our methods accurately label more lesions that are missed, i.e., false negative, by the baselines.
• Figure 3: Ablation studies of mixing rate and postprocessing threshold. For Dice and F1 scores, higher is better. A mixing rate of 0.8 with a post-processing threshold of 0.75 is found as the optimal configuration.
• Figure 4: (a) Lesion distribution of MSL categories over 655 scans in the ATLAS v2.0 dataset. The volume thresholds of each category are listed on the horizontal axis. (b) Average lesion voxels of DBL categories in each training scan in the ATLAS v2.0 dataset. Lesion voxels with a distance not larger than 2 to the non-lesion region are defined as the boundary region.
• Figure 5: Lesion distribution of MSL with (a) 5 categories and (b) 3 categories over 655 scans in the ATLAS v2.0 dataset. The volume thresholds of each category are listed on the horizontal axis.