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.

  1. 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.

  2. 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.

  3. 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:

  1. 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.

  2. 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:

  1. 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.

  2. For Patients:

    • Enables early-stage malaria detection, facilitating timely treatment and reducing the likelihood of severe symptoms or death.

  3. 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.