Efficient Medical Image Segmentation Using Probabilistic KNN Label Downsampling

Deep learning-based medical image segmentation has advanced diagnostic precision and treatment planning. However, training on high-dimensional data remains computationally challenging due to substantial memory and processing demands. Downsampling is a widely employed strategy that reduces memory requirements and accelerates training processes to mitigate these issues. Conventionally, nearest neighbor (NN) interpolation has been utilized to downsample ground truth labels. However, this approach often leads to loss of class information and can detrimentally impact segmentation performance compared to training on the original high-dimensional data. This study proposes a Probabilistic K-Nearest Neighbors (PKNN) downsampling method that effectively preserves class details over NN interpolation. Evaluations at half and quarter resolutions with varying K values demonstrate that PKNN consistently outperforms NN interpolation on the KNUH Abdominal CT and CVC-ClinicDB datasets, improving Intersection over Union (IoU) by up to 2.29% and 2.88%, respectively. PKNN's performance closely approximates models trained on full-resolution data, confirming its suitability for maintaining segmentation accuracy despite reduced resolution.