Clustering-Based Dual Deep Learning Architecture for Detecting Red Blood Cells in Malaria Diagnostic Smears

Yasmin M. Kassim, Kannappan Palaniappan, Feng Yang, Mahdieh Poostchi, Nila Palaniappan, Richard James Maude, Sameer K. Antani, Stefan Jäger

Abstract: Computer-assisted algorithms have become amainstay of biomedical applications to improve accuracyand reproducibility of repetitive tasks like manual segmen-tation and annotation. We propose a novel pipeline for redblood cell detection and counting in thin blood smear mi-croscopy images, named RBCNet, using a dual deep learn-ing architecture. RBCNet consists of a U-Net first stage forcell-clusteror superpixel segmentation, followed by a sec-ond refinement stage Faster R-CNN for detecting small cellobjects within the connected component clusters. RBCNetuses cell clustering instead of region proposals, which isrobust to cell fragmentation, is highly scalable for detectingsmall objects or fine scale morphological structures in verylarge images, can be trained using non-overlapping tiles,and during inference is adaptive to the scale of cell-clusterswith a low memory footprint. We tested our method on anarchived collection of human malaria smears with nearly 200,000 labeled cells across 965 images from 193 patients,acquired in Bangladesh, with each patient contributing fiveimages. Cell detection accuracy using RBCNet was higherthan 97%. The novel dual cascade RBCNet architectureprovides more accurate cell detections because the fore-ground cell-cluster masks from U-Net adaptively guide the detection stage, resulting in a notably higher true positiveand lower false alarm rates, compared to traditional andother deep learning methods. The RBCNet pipeline imple-ments a crucial step towards automated malaria diagnosis.