Author: Yizheng Chen, Md Tauhidul Islam, Mingjie Li, Lei Xing 👨🔬
Affiliation: Department of Radiation Oncology, Stanford University 🌍
Purpose:
Biomedical image segmentation (BIS) is a cornerstone of medical physics, enabling accurate delineation of anatomical structures and abnormalities, which is critical for diagnosis, treatment planning, and image-guided interventions. However, existing methods often struggle to retain long-range spatial information due to conventional down-sampling techniques that prioritize computational efficiency over information preservation. This study introduces a novel strategy, Stair Pooling, to address these challenges and enhance segmentation accuracy by improving spatial information retention without increasing computational burden.
Methods:
Our Stair Pooling aims to bridge the gap between computational demands and detail preservation, ultimately supporting large-scale medical imaging workflows that require both efficiency and precise anatomical delineation. We utilized three BIS benchmarks: the Synapse multi-organ segmentation dataset (Synapse), the Automated Cardiac Diagnosis Challenge dataset (ACDC), and the Kidney Tumor Segmentation Challenge 2023 dataset (KiTS23). The proposed Stair Pooling moderates the down-sampling process by employing a sequence of small and narrow pooling operations in varied orientations, reducing the dimensionality per 2D pooling step from ½ to ¼, with extensions for 3D pooling to further enhance information preservation. Segmentation performance was assessed using Dice scores, and transfer entropy analysis was conducted to quantitatively evaluate information loss.
Results: The proposed Stair Pooling method demonstrated an average Dice score improvement of 3.8% for both 2D and 3D U-Net models across the Synapse, ACDC, and KiTS23 datasets. By preserving spatial details and long-range information during down-sampling, Stair Pooling significantly improved segmentation accuracy while maintaining the computational efficiency. Transfer entropy analysis further confirmed the reduction in information loss enabled by this approach.
Conclusion: Stair Pooling is a robust strategy that addresses critical limitations in medical image segmentation by enhancing spatial information retention during down-sampling. Its demonstrated improvements across multiple BIS benchmarks highlight its potential to advance medical imaging applications in clinical and research settings.