Malaria parasite detection
During my MSc, I worked together with Xavier Liceras and Albert Folch to implement a prototype of a mobile parasite detection application. The app runs on Android devices and is capable of scanning dummy blood slides according to WHO standards within 2 minutes. This project was provided by Momala, a company focused on delivering high-quality malaria diagnosis in low-resource environments.
We wrote the application from the ground up in Java and used Android development tools. The detection model is an SSD with a Mobilenet backbone written in Tensorflow. The pruned, quantized and compressed model was transferred to mobile using TFLite (at the time still in beta). All other visual processing, such as blur detection, duplicate image detection, and focus calibration was done with OpenCV.
In conclusion, the application:
- Scans dummy blood slides in 2 minutes - thereby potentially relieving doctors from strenuous manual counting that can take up to 15 minutes.
- Detects and rejects blurry images - to help prevent further issues with poor data acquisition.
- Exposes an API to accommodate microscope and phone focus calibration.
- Avoids capturing duplicate images - essential for the validity of cell and parasite counts
- Provides a user-friendly interface.
