GHOST MEETS KAN WITH MAMBA: A HYBRID LOCAL-GLOBAL NETWORK FOR LOW-COST MEDICAL IMAGE CLASSIFICATION

Accurate medical image classification demands sensitivity to fine-grained details and long-range context under tight compute budgets. We propose GHOST Meets KAN With Mamba Networks (GhostKMNet), a lightweight and low-cost hybrid that couples Ghost Convolution for cheap local feature extraction, Kolmogorov-Arnold Network (KAN) modules for data-adaptive and nonlinear calibration, and Mobile Mamba blocks for modeling of long-range dependencies. For efficient and low-cost local-global fusion, we propose multiple modules to enhance feature context modeling and cross-scale propagation, and to adaptively fuse the outputs of Mamba and KAN. We evaluate GhostKMNet on 8 diverse medical imaging benchmarks, including hematology, radiology, pathology, endoscopy, and dermatology datasets. The results demonstrate consistent state-of-the-art performance with only 0.2 GFLOPs. The source code and weight will be made publicly available on GitHub after acceptance.