More About the AI Model
Technical Approach
The foundation of our solution lies in a combination of state-of-the-art AI methodologies designed for accuracy, efficiency, and adaptability in resource-limited settings.
YOLO (You Only Look Once):
Developed by Ultralytics, YOLO is a high-speed object detection framework known for its real-time processing capabilities.
In our model, YOLO accurately identifies and classifies malaria parasites, specifically detecting the trophozoite stage and distinguishing between infected Red Blood Cells (RBCs) and uninfected White Blood Cells (WBCs).
It is particularly well-suited for environments where computational resources are constrained, as it balances speed and precision effectively.
Detection Transformers with Assignment (DETA):
DETA adds another layer of robustness by improving detection accuracy in complex and cluttered images.
By leveraging transformers, it ensures the model performs reliably even in challenging scenarios where parasites may overlap or present atypical patterns.
Smaller YOLO Variant for Deployment:
For the web-based tool, a lightweight version of the YOLO model has been implemented.
This variant prioritizes speed and ease of inference, making it accessible for real-time use in healthcare settings without high-end hardware requirements.
A detailed technical write-up outlining our approach will be made available soon.
Infrastructure of the App
The app infrastructure has been designed to ensure scalability, reliability, and currently most importantly cost-efficiency:
Inference Backend:
Powered by NVIDIA A10G GPU, hosted on Modal.
This setup enables processing of blood slide images for AI-driven malaria detection.
Web Hosting:
Hosted on Render, a platform optimized for deploying scalable web applications.
The current implementation uses free-tier hosting on both Modal and Render, allowing us to provide the service without incurring costs up to a certain amount of usage.
Accessibility:
The infrastructure ensures the model can be used effectively in low-resource settings. However, users are kindly requested not to overload the system to maintain smooth operation.
Current Limitations
While our solution addresses many challenges, certain limitations currently remain:
Throughput Limitations:
Modal gives a free credit of 30$/month of compute, one query takes up about 0.02$. This means we get roughly 1500 detections for free.
Dependency on Compute Resources:
The speed and efficiency of the AI model are influenced by the computational hardware available. While the lightweight YOLO variant mitigates this to an extent, access to GPUs or high-performance CPUs enhances performance significantly.
The goal is to self host this and other developed models, which would mitigate the limitations and the costs. But would require a considerate time investment on our part.
Impact of the Solution
Our model offers transformative benefits for healthcare in resource-limited settings:
For Healthcare Systems:
Automates the diagnostic process, reducing the workload for overburdened technicians.
Improves diagnostic speed, allowing more patients to be examined in less time.
For Patients:
Enables early-stage malaria detection, facilitating timely treatment and reducing the likelihood of severe symptoms or death.
For Malaria Transmission Control:
Early and accurate detection helps interrupt the transmission cycle, supporting broader eradication efforts.
By leveraging AI, this solution enhances healthcare accessibility, efficiency, and effectiveness, making a tangible difference in the fight against malaria.
Looking Ahead
We aim to refine and expand the capabilities of this solution by:
Exploring new deployment methods, such as self hosting.
Updating the model with additional training data to improve accuracy further.
Collaborating with local healthcare systems, hopefully collaborating with Lacuna Fund, to integrate the solution into everyday diagnostics.