CAENet: Contrast adaptively enhanced network for medical image segmentation based on a differentiable pooling function

Li, Shengke, Feng, Yue, Xu, Hong ORCID: 0000-0002-2968-9804, Miao, Yuan ORCID: 0000-0002-6712-3465, Lin, Zhuosheng, Liu, Huilin, Xu, Ying and Li, Fufeng (2023) CAENet: Contrast adaptively enhanced network for medical image segmentation based on a differentiable pooling function. Computers in Biology and Medicine, 167. ISSN 0010-4825


Pixel differences between classes with low contrast in medical image semantic segmentation tasks often lead to confusion in category classification, posing a typical challenge for recognition of small targets. To address this challenge, we propose a Contrastive Adaptive Augmented Semantic Segmentation Network with a differentiable pooling function. Firstly, an Adaptive Contrast Augmentation module is constructed to automatically extract local high-frequency information, thereby enhancing image details and accentuating the differences between classes. Subsequently, the Frequency-Efficient Channel Attention mechanism is designed to select useful features in the encoding phase, where multifrequency information is employed to extract channel features. One-dimensional convolutional cross-channel interactions are adopted to reduce model complexity. Finally, a differentiable approximation of max pooling is introduced in order to replace standard max pooling, strengthening the connectivity between neurons and reducing information loss caused by downsampling. We evaluated the effectiveness of our proposed method through several ablation experiments and comparison experiments under homogeneous conditions. The experimental results demonstrate that our method competes favorably with other state-of-the-art networks on five medical image datasets, including four public medical image datasets and one clinical image dataset. It can be effectively applied to medical image segmentation.

Dimensions Badge

Altmetric Badge

Item type Article
DOI 10.1016/j.compbiomed.2023.107578
Official URL
Subjects Current > FOR (2020) Classification > 4603 Computer vision and multimedia computation
Current > Division/Research > Institute for Sustainable Industries and Liveable Cities
Keywords medical image; semantic segmentation; differentiable pooling function; channel attention; deep supervision
Download/View statistics View download statistics for this item

Search Google Scholar

Repository staff login