SparseQ: Jensen-Shannon Divergence Guided Query Sparsified-Attention with Convolutional Distilling for Medical Image Segmentation.
Authors
Abstract
Accurate medical image segmentation is essential for clinical diagnosis and treatment planning. Although Convolutional Neural Networks (CNNs) excel at capturing local spatial details, they struggle to model long-range dependencies across anatomical regions, which is critical for lesion and tumor delineation. Transformer-based models address this through self-attention but incur prohibitive computational costs that scale quadratically with the number of tokens, limiting deployment in resource-constrained environments. We propose SparseQ, a hybrid framework that combines efficient sparse attention with convolutional feature refinement. The core innovation introduces a Jensen-Shannon Divergence (JSD)-motivated sparsity criterion, efficiently approximated via logit-dispersion scoring, to selectively compute dominant query-key interactions, reducing attention complexity from $\mathcal {O}(L^{2})$ to $\mathcal {O}(L \ln L)$ while preserving segmentation accuracy. A complementary convolutional distilling module with progressively dilated convolutions expands the receptive field without parameter inflation, enabling multi-scale feature learning. The decoder integrates unified content queries with high-resolution pixel embeddings for precise boundary delineation. Extensive experiments on BraTS2023 and BraTS2019 (multi-region tumor segmentation) and ISLES-2022 (binary ischemic lesion segmentation) demonstrate that SparseQ consistently outperforms strong CNN, Transformer, and hybrid baselines, achieving a mean Dice of 85.8% and a mean HD95 of 3.28, while reducing training time and GPU memory. These results validate that principled, content-adaptive sparsity can enable accurate and efficient medical image segmentation suitable for practical use.