Edge-Preserving Probabilistic Downsampling for Reliable Medical Segmentation in Resource-Constrained Environments

Downsampling, often necessitated due to resource constraints such as limited memory, development timelines, or the need to accelerate learning from large volumes of unlabeled medical data, results in the loss of fine-grained details, including small objects and thin class boundaries. When applied to ground truth (GT) labels without proper care, downsampling can significantly impair the segmentation network's ability to accurately interpret and predict detailed features. This, in turn, leads to a degraded performance compared to networks trained on higher-resolution labels. This situation exemplifies the trade-off between computational efficiency and segmentation accuracy, with higher downsampling factors further impairing performance. Preserving critical information during label downsampling is particularly crucial for medical image segmentation. This study introduces a novel approach, Edge-Preserving Probabilistic Downsampling (EPD), designed to retain critical details and bridge the performance gap between networks trained on original and downsampled resolutions. It leverages class uncertainty within a local window to produce soft labels, with the window size determining the downsampling factor. This approach enables segmentation networks to maintain high-quality predictions at lower resolutions. Beyond preserving edge details more effectively than conventional nearest-neighbor downsampling, the proposed algorithm, when adapted for images, exhibits notable performance improvements over conventional bilinear interpolation. Experimental results indicate EPD improves Intersection over Union (IoU) for a multi-class abdominal CT dataset by 2.54%, 10.66%, and 16.31% when downsampling to ½, ¼, and ⅛, respectively, compared to conventional interpolation methods.