Malaria Detection via SAM-based Red Blood Cell Segmentation and Feature Analysis

Abstract

The diagnosis of malaria is still predominantly based on the manual observation of blood samples through optical microscopy, a time-consuming and tiring task for laboratory staff. As an alternative to this gold standard, previous studies suggest that the application of artificial intelligence can allow for faster and more accurate diagnoses, leading to more effective patient treatment. This study aims to develop a pipeline for the automatic image-based detection of malaria using a traditional machine learning method to classify Plasmodium-infected and uninfected red blood cells. Additionally, the project seeks to identify the most important types of image features that enable differentiation between infected and uninfected cells. The methodology includes the manual annotation of red blood cells from thin blood smear images of malaria patients, segmentation of cells using the novel Segment Anything Model (SAM), application of the Random Forest algorithm to classify segmented cells into infected or uninfected classes using a feature dataset, and calculation of feature importances. Moreover, an object detection model is trained to automate the detection of RBCs in microscopy images. Results of segmentation with SAM show remarkable performance. After training and testing the classification model within an eighty-feature dataset of colour, texture, and morphology, the accuracy and F1-score obtained are 99.5% and 99.2%, respectively. Thus, an accurate malaria diagnosis could be achieved by applying this pipeline based on SAM segmentation and Random Forest classification to analyse thin smears. Regarding feature importance, colour features, specifically green and red histograms, appear to be the most distinctive.