Research Perspective: Efficient Cervical Cytology at the Edge¶
1. Introduction: The Need for Edge AI in Pathology¶
Cervical cancer screening is most critical in low-and-middle-income regions where access to high-end cloud computing or powerful GPU clusters is limited. To democratize AI-assisted screening, the underlying models must be computationally efficient without sacrificing diagnostic accuracy.
This project focuses on On-Device or Edge AI capabilities, simulating a system that could be embedded directly into a digital microscope or a local clinic's workstation.
2. Model Selection: Why MobileViT-S?¶
We selected MobileViT-S (Small) as the backbone for cell classification. This decision represents a strategic trade-off between model size, inference latency, and accuracy.
2.1 Key Architectural Advantages¶
MobileViT combines the best of two worlds: * Inductive Bias of CNNs: It uses convolutions to efficiently process local features (edges, textures). * Global Context of Transformers: It allows pixels to "see" distinct parts of the image, capturing global shape information crucial for differentiating subtle cellular anomalies.
2.2 Efficiency by the Numbers¶
- Parameters: ~5.6 Million
- Model Size: < 25 MB (FP32)
- Inference Speed: Capable of real-time processing on standard mobile-grade CPUs or entry-level GPUs.
compared to standard architectures: * ResNet-50: ~25M params (4x larger) * ViT-Base: ~86M params (15x larger)
Despite being significantly smaller, MobileViT approaches the accuracy of these larger models for this task.
3. Performance & Clinical Relevance¶
High Sensitivity for Screening¶
In our evaluation (see README metrics), the model achieves: * 92.6% Overall Accuracy * 98% Recall on Dyskeratotic cells (High-grade lesions) * 93% Recall on Koilocytotic cells (Low-grade lesions)
The Edge Advantage¶
- Privacy: Data does not need to leave the clinic.
- Latency: No network round-trip time; results are available immediately as the slide is scanned.
- Cost: Runs on affordable hardware (e.g., Jetson Nano, Raspberry Pi with accelerator, or standard office laptops).
4. Conclusion¶
The choice of MobileViT-S proves that high-performance medical AI does not require massive compute resources. By optimizing for efficiency (~5.6M params), we enable a scalable, deployable solution for automated cervical cytology screening.