Automated parasite detection in thick blood smears, A deep learning approach
Brown Bag Lecture by Feng Yang, Ph.D. | 5/15/2018 11:00AM – 12PM | 7th Floor Conference Room, Bldg 38A
We investigate the problem of automated parasite detection in thick blood smear images for malaria detection. We propose a framework consisting of two steps: screening and prediction. First, we use an intensity-based greedy method to perform a fast screening to generate parasite candidate regions of interest. Then, we train a customized CNN model to classify the preselected candidates as parasite or background. Our CNN model consists of seven convolutional layers, three max-pooling layers, three fully connected layers, and a softmax layer. Experimental results based on a dataset including 120 patients via patient-level five-fold cross-validation demonstrates the effectiveness of the customized CNN model in discriminating between positive and negative patches in terms of the following performance indicators: accuracy (92.87%±0.67%), AUC (98.16%±0.40%), sensitivity (91.98%±1.99%), specificity (93.77±1.06%), precision (93.68±0.91%), and negative prediction (92.16%±1.76%). Linear regression between the predicted parasite number and the annotated parasites in ground truth shows that correlation coefficient above 0.98 is obtained on both image level and patient level.
Dr. Yang is a Visiting Scientist at CEB working with Dr. Jaeger. She received her PhD from National Institute of Applied Science (INSA Lyon) in France in 2011, and her B.S. and M.S. degrees from Northwestern Polytechnical University in China in 2005 and 2007, respectively. She is also an associate professor in Beijing Jiaotong University, Beijing, China. Her research includes deep learning based medical image analysis and traditional medical image processing and analysis.