An SSD Multi-box Approach to Automated Malaria Diagnosis on Thick Blood Smears
Keywords
No Thumbnail Available
Authors
Issue Date
2018-08-03
Language
en
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
In 2017, the World Health Organization (WHO) reported 216 million cases of malaria worldwide (91 coun-
tries), an increase of about 5 million cases over the year 2015. An early diagnosis is the best prevention against
malaria, and microscopy remains to date the gold standard for malaria diagnosis. Yet, in most a ected regions,
shortage in expertise and costly laboratory equipment grant limited access to an adequate diagnosis. This
thesis introduces a high-performing, state-of-the-art, real-time object detector (SSD Multi-box) to the prob-
lem of malaria detection. Using deep convolutional networks, it responds to the presence of trophozoites of
P. falciparum and white blood cells in Field-stained thick blood smears. The detection of white blood cells
allows for parasitaemia estimation, a requirement to assess the severity of the infection, and in turn, the type of
treatment. SSD Multi-box localizes and recognizes P. falciparum (recall: 87.52%; precision: 90.72%) and white
blood cells (recall: 97.54%; precision: 81.27%) on a single feed-forward pass of network at a rate of 9 FPS. Our
model outperforms results from similar multi-class studies, yet o ering unprecedented time performance even
on low-res image sets, such as the one used here. Transfer learning did not show great improvement, yet results
suggest that transferring features extractors from early layers can be bene cial.
Description
Citation
Faculty
Faculteit der Sociale Wetenschappen