Medical Image Object Detection via Layout-Aware Convolution and Optimal Transport Collaboration

To address the dual challenges of limited local feature representation in conventional convolutional networks and the loss of spatial context in knowledge distillation for medical image object detection, we propose an innovative framework that integrates layout-aware convolution with optimal transport-based distillation. A layout-aware convolution module (LAConv) is designed, incorporating a parameter decomposition strategy to generate spatially adaptive dynamic kernels, which effectively capture fine-grained features and stable anatomical priors in medical images. Additionally, we introduce an optimal transport-based knowledge distillation method (OTKD) using Sinkhorn divergence. By constructing batch-level transport cost matrices, this method models the global distribution discrepancy between teacher and student models, achieving precise alignment of spatial structure information. Extensive experiments on the ChestX-Det and Kvasir-SEG datasets demonstrate the effectiveness of our approach, offering a solution that balances accuracy and efficiency for precise medical image detection.