Physics World published a feature on this project.

An automated Convolutional Neural Network (CNN)-based model designed for the diagnosis of malaria from microscopic blood smear images. The primary goal is to reduce the reliance on trained microscopists by leveraging modern deep-learning techniques for accurate and efficient diagnosis. The research employs various strategies, including knowledge distillation, data augmentation, and the use of Autoencoders for feature extraction, followed by classification through Support Vector Machines (SVM) or K-Nearest Neighbors (KNN). These methods are integrated into three distinct training procedures—general training, distillation training, and autoencoder training—to optimize the model's accuracy and inference speed. The resulting model achieves a remarkable accuracy of 99.23%, requiring minimal computational resources, specifically just over 4600 floating point operations per inference.

To validate the model's practical efficiency, it was deployed in various environments, including mobile phones and a server-backed web application. The data gathered from these deployments show that the model can perform inference in less than 1 second per sample in both offline (mobile) and online (web) modes. This quick response time, coupled with high accuracy, demonstrates the model's potential for real-world application in malaria diagnosis, particularly in resource-limited settings where access to trained personnel and sophisticated equipment may be limited. The study highlights the viability of deploying such automated systems to enhance the accessibility and efficiency of malaria diagnosis, making it a promising tool for widespread use.