KAN-Mamba FusionNet: Redefining Medical Image Segmentation with Non-Linear Modeling

2024-11-18

arXiv.org

AI Summary

Medical image segmentation faces the challenge of simultaneously modeling nonlinear features and capturing long-range dependencies: Kolmogorov–Arnold Networks (KANs) excel at nonlinear representation but lack global contextual modeling, whereas Mamba architectures effectively model long-range dependencies yet underperform in local nonlinear fitting. To address this, we propose KAMBA, a novel fusion module that synergistically integrates KANs and Mamba via a learnable gating mechanism to jointly optimize nonlinear feature representation and long-range contextual awareness, further enhanced by multi-scale feature interaction. Evaluated on three benchmark medical datasets—BUSI, Kvasir-Seg, and GlaS—KAMBA achieves state-of-the-art performance in both IoU and F1 scores. Ablation studies confirm its superior robustness in segmenting small lesions and ambiguous boundaries. This work establishes a new paradigm for accurate segmentation of complex pathological regions.

Technology Category

• Computer Vision: Segmentation
• Machine Learning: Mixture of Experts (MoE)
• Search and Optimization: Learning to Search

Application Category

• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Search and Retrieval-Augmented AI: Retrieval-Augmented Generation (RAG) and multi-modal RAG
• Graph Algorithms and Modeling for the Web: Graph embeddings and representation learning for Web-related graphs

Abstract

Medical image segmentation is essential for applications like robotic surgeries, disease diagnosis, and treatment planning. Recently, various deep-learning models have been proposed to enhance medical image segmentation. One promising approach utilizes Kolmogorov-Arnold Networks (KANs), which better capture non-linearity in input data. However, they are unable to effectively capture long-range dependencies, which are required to accurately segment complex medical images and, by that, improve diagnostic accuracy in clinical settings. Neural networks such as Mamba can handle long-range dependencies. However, they have a limited ability to accurately capture non-linearities in the images as compared to KANs. Thus, we propose a novel architecture, the KAN-Mamba FusionNet, which improves segmentation accuracy by effectively capturing the non-linearities from input and handling long-range dependencies with the newly proposed KAMBA block. We evaluated the proposed KAN-Mamba FusionNet on three distinct medical image segmentation datasets: BUSI, Kvasir-Seg, and GlaS - and found it consistently outperforms state-of-the-art methods in IoU and F1 scores. Further, we examined the effects of various components and assessed their contributions to the overall model performance via ablation studies. The findings highlight the effectiveness of this methodology for reliable medical image segmentation, providing a unique approach to address intricate visual data issues in healthcare.

Problem

• Improves medical image segmentation accuracy
• Combines non-linear modeling and long-range dependencies
• Outperforms existing methods on diverse datasets

Innovation

• Combines KANs and Mamba for segmentation
• Uses KAMBA block for long-range dependencies
• Improves accuracy with non-linear modeling