LGLD-Net: A Breakthrough in Lightweight Global-Local Dual-Branch Neural Network for Medical Microscopic Image Classification

LGLD-Net is a lightweight global-local dual-branch neural network with only 161K parameters. By fusing local texture features and global context information, it achieves an accuracy of 99.69% in acute lymphoblastic leukemia classification tasks and 96.50% in blood cell classification tasks. It can also perform real-time inference at 17.56 FPS on Raspberry Pi 5, providing a new solution for edge deployment of medical microscopic image analysis.

Microscopic image analysis is crucial in medical diagnosis, but traditional deep learning models face two major challenges: First, it is difficult to capture both fine-grained local texture features (such as cell morphology details) and global context information (such as spatial relationships between cells) simultaneously; Second, high-performance models have large parameter counts and high computational costs, making them difficult to deploy on resource-constrained edge devices, which limits the realization of on-site rapid diagnosis.

LGLD-Net addresses the above problems through a dual-branch architecture:

1. Local Feature Aggregator (LFA)：Uses an Involution-based Patch Stem (IPS). Involution shares parameters across channels and is independent in space, making it suitable for capturing cell micro-textures (e.g., nucleus shape, cytoplasm distribution);
2. Global Context Extractor (GCE)：Uses a Fourier Transform Feature Extraction Module (FFM), which converts images to the frequency domain to capture long-range dependencies and global tissue patterns;
3. Adaptive Fusion Module (LGFM)：Dynamically balances the contributions of local and global features, automatically adjusting weights based on the characteristics of input images. The total number of model parameters is 161K, with a computational cost of only 1.88 GFLOPs.

Dataset Performance

• Acute Lymphoblastic Leukemia (ALL) Dataset：Classification accuracy reaches 99.69%;
• Raabin-WBC White Blood Cell Dataset：Multi-class classification accuracy reaches 96.50%.

Edge Device Deployment

Achieves real-time inference speed of 17.56 FPS on Raspberry Pi 5, supporting local rapid acquisition of AI-assisted diagnosis results without relying on the cloud.

1. Balance between Parameter Count and Efficiency：161K parameters are far lower than ResNet-50 (25 million) and EfficientNet-B0 (5.3 million), achieving comparable or better performance with less than 1% of the parameters;
2. Adaptation to Medical Images：Involution captures cell details, while Fourier transform models global tissue patterns, fitting the characteristics of microscopic images;
3. Feasibility of Clinical Deployment：The real-time inference capability of edge devices can support scenarios such as remote area clinics and mobile medical vehicles, improving medical equity.

1. Generalization Ability Verification：Currently only validated on blood-related microscopic images; needs to be extended to other types such as histopathological sections and cytological smears;
2. Interpretability Enhancement：The decision-making process of the adaptive fusion module needs to be strengthened to enhance doctors' trust;
3. Task Expansion：Can explore expansion to object detection (abnormal cell localization) or segmentation tasks (cell boundary outlining).

LGLD-Net represents the "small but refined" design trend in the medical AI field, proving that well-designed architecture can surpass parameter stacking strategies. For researchers, it demonstrates the possibility of fusing traditional signal processing (Fourier transform) with deep learning (Involution); For medical practitioners, it means that lightweight and fast AI-assisted tools will gradually become popular. In the future, they are expected to be deployed on smartphones and portable devices to achieve anytime, anywhere intelligent medical care.