Deep unsupervised learning for Microscopy-Based Malaria detection

TL;DR

This paper presents an unsupervised deep learning approach using a modified U-net and Mahalanobis distance for malaria detection in microscopy images, aiming to reduce manual labeling and improve diagnostic efficiency.

Contribution

It introduces a novel unsupervised workflow combining cell boundary detection and malaria identification, eliminating the need for manual annotations.

Findings

Effective cell segmentation without manual labels

Accurate malaria detection in microscopy images

Potential for automated diagnosis workflows

Abstract

Malaria, a mosquito-borne disease caused by a parasite, kills over 1 million people globally each year. People, if left untreated, may develop severe complications, leading to death. Effective and accurate diagnosis is important for the management and control of malaria. Our research focuses on utilizing machine learning to improve the efficiency in Malaria diagnosis. We utilize a modified U-net architecture, as an unsupervised learning model, to conduct cell boundary detection. The blood cells infected by malaria are then identified in chromatic space by a Mahalanobis distance algorithm. Both the cell segmentation and Malaria detection process often requires intensive manual label, which we hope to eliminate via the unsupervised workflow.