MicroView AI: A Low-Cost Urine Microscopy Analysis System Based on Vision-Language Models

MicroView AI is a urine sediment microscopy analysis system based on Raspberry Pi and large vision-language models (VLMs). It aims to provide low-cost, efficient medical testing tools for resource-limited areas. Developed by undergraduate students from the University of Manila in the Philippines, this project demonstrates the innovative application of multimodal AI in medical diagnosis. Its core is to achieve local inference through edge computing, addressing the pain points of traditional urine analysis that relies on professionals and expensive equipment.

Urine analysis is a basic clinical testing method, but traditional sediment microscopy is time-consuming and relies on the operator's experience. In areas with scarce medical resources, the lack of professionals and automated equipment limits diagnostic capabilities. MicroView AI addresses this pain point by using VLM technology to develop a low-cost, portable solution, embodying the combination of academic research and practical application.

Hardware Platform: Raspberry Pi

Raspberry Pi is chosen as the core computing platform, with advantages including affordability, small size, low power consumption, and a rich ecosystem. It upgrades the microscope to a digital imaging system by connecting a custom optical module.

Core AI: Large Vision-Language Models

VLMs have advantages such as multimodal understanding (processing images and text simultaneously), zero-shot/few-shot learning (reducing the need for labeled data), and interpretable output (natural language reports).

Software Architecture

Modular design: Image acquisition layer, preprocessing module, AI inference engine, result generation layer, and user interface.

Model Optimization

Fine-tuning VLMs for medical scenarios, including domain adaptation (fine-tuning on urine image datasets), prompt engineering (guiding attention to medical features), and multi-scale analysis (combining different magnification levels).

Edge Computing Deployment

Local inference without cloud dependency, available offline, data securely stored locally, suitable for remote areas.

User Interaction Design

Simple and intuitive interface, auxiliary decision-making (AI suggestions + final diagnosis by doctors), built-in image quality detection.

Auxiliary Diagnostic Capability

Identify and classify urine components: cells (red/white blood cells, etc.), casts (transparent/granular, etc.), crystals (calcium oxalate, etc.), microorganisms (bacteria, etc.), and other components, with automatic counting and abnormal marking.

Applicable Scenarios

Primary healthcare institutions, resource-limited areas, mobile healthcare, medical education, home monitoring (requires doctor guidance).

Unstable Image Quality

Challenge: Image issues caused by differences in microscope equipment; Solution: Built-in quality assessment algorithm + image enhancement preprocessing.

Computing Resource Constraints

Challenge: Limited performance of Raspberry Pi; Solution: Model quantization/knowledge distillation compression, optimized inference pipeline, optional cloud collaboration.

Scarce Labeled Data

Challenge: Difficulty in obtaining medically labeled data; Solution: VLM few-shot learning + active learning strategies, collaborating with medical institutions to share data.

MicroView AI is released as open source, lowering technical barriers for secondary development, promoting academic exchanges, empowering global health, and narrowing the gap in medical diagnostic capabilities. Developers hope to continuously improve it through the open source community, promote deployment to institutions in need, and improve health equity.

Future Directions

1. Multimodal fusion (integrating dry chemical analysis data); 2. Cloud collaboration (assisting with complex cases); 3. Extending to other body fluid analyses; 4. Establishing data standards and quality control systems; 5. Conducting clinical trials to verify accuracy.

Conclusion

The project combines open source hardware, edge computing, and VLMs to achieve the functions of traditionally expensive equipment at low cost, demonstrating the potential of AI to improve diagnostic capabilities in resource-limited areas, which is an embodiment of technology for good.