Lightweight Mamba Networks for Edge-Based Medical Image Classification and Diagnosis

Abstract

Medical image classification has advanced rapidly through large-scale deep learning, yet translating these models into point-of-care diagnostics remains difficult because of privacy, latency, connectivity, and computational constraints. Edge-based deployment offers a possible solution by moving inference and selective learning closer to the clinical site, but conventional convolutional and attention-based architectures often carry assumptions that are poorly aligned with constrained hardware, heterogeneous imaging environments, and regulated healthcare workflows. Selective state space models, particularly lightweight Mamba networks, provide a promising alternative because they combine linear-time sequence processing with input-dependent state transitions that can capture long-range dependencies without the quadratic cost of self-attention. This paper examines the system-level implications of adopting lightweight Mamba networks for edge-based medical image classification and diagnosis. It addresses architectural trade-offs in tokenization, state dimensioning, scan direction, and hybrid feature extraction. It further analyzes the integration of such networks into federated learning, differential privacy, robustness engineering, and clinical governance. Rather than focusing narrowly on accuracy, the discussion emphasizes structural trade-offs across energy budgets, hardware heterogeneity, adversarial exposure, fairness, auditability, and lifecycle management. The paper argues that lightweight Mamba architectures should be understood as socio-technical components embedded within distributed diagnostic infrastructures. Their clinical value depends not only on algorithmic performance but also on how effectively they are governed, monitored, and aligned with delivery systems, regulatory expectations, and sustainability goals.