Introducing soft TopK quadratic convolution to improve the performance of neural networks in medical image segmentation.

Authors

Abstract

Medical image segmentation is crucial for computer-aided diagnosis and treatment planning, where accurate organ delineation serves as the foundation for reliable quantitative analysis. Standard convolutional layers primarily extract local features through linear projections within the receptive field, which to some extent limits their ability to model complex nonlinear local responses. To address this, we propose a Soft TopK Quadratic Convolution (STKQC) module as a plug-in enhancement unit for standard convolutional layers. STKQC introduces a parallel quadratic feature modeling branch alongside the traditional linear convolution branch. Unlike explicitly constructing a complete quadratic matrix over the entire local patch, STKQC employs soft-pooling-style neighborhood feature selection through learnable local scale selection and dynamic grid adjustment, and further constructs a lightweight diagonal quadratic response. This approach introduces local second-order nonlinear information while effectively controlling additional parameters and computational overhead. We evaluated STKQC on the MICCAI 2017 cardiac biventricular segmentation dataset and the AMOS22 abdominal multi-organ segmentation dataset. Experimental results demonstrate that integrating STKQC into a typical encoder-decoder network improves model performance in segmentation tasks, validating its effectiveness as an efficient convolutional enhancement module for medical image segmentation.

Tags

Topics