The Adaptive Neuron Network Expansion for Medical Image Segmentation

Many deep learning architectures struggle to combine the benefits of local feature extraction and global contextual awareness, limiting their effectiveness in complex segmentation tasks (i.e., medical image segmentation). In this work, we introduce ANNE: The Adaptive Neuron Network Expansion for 2D Medical Image Segmentation. ANNE is an efficient yet computationally effective deep learning architecture, inspired by the Remez and Progressive Expansion Neurons algorithms, designed to approximate complex functions with fewer trainable parameters. Our model utilizes Mamba and Adaptable Progressive Expansion Neurons-based techniques to better leverage the combined benefits of local and global features. Experimental results demonstrate competitive performance in various medical image segmentation tasks, including retinal, polyp, and skin lesion segmentation, while achieving a significantly reduced number of trainable parameters compared to existing deep learning models.