Probability Based Bone Marrow Cell Detection

Bone marrow cell images are widely used to di-agnose leukaemia. Accurate detection of cells has great impact on the sound effects of diagnosis. In this paper, we propose a new framework to detect bone marrow cells and to improve the detection results of some specific cell types. We first use detection network to get preliminary bounding boxes and corre-sponding cell types, then we analyze the probability distribution for each bounding box, finding that those easily misclassified bounding boxes' probability distribution obey some specific rules. Therefore, we chose those bounding boxes that satisfied specific probability distribution as easily confused cell pairs, and then use a feature extraction network to classify these cell pairs. We evaluate this framework on the bone marrow cell dataset and compare with the traditional object detection framework.