Automated Leukocyte Differential Count System Using Dual-Stage Convolutional Neural Network

Abstract

Leukocyte or white blood cell differential count is an essential examination modality of hematology laboratory in diagnosis of various blood disorders. However, it requires highly experienced hematologists for correct diagnosis from samples with inter- and intra-sample variations. Due to tedious, time and cost consuming procedure of manual differential count, there has been high demands for development of automated system. In order for it to be applicable in clinical hematology laboratories, an automated system will have to detect and classify leukocytes of different maturation stages, especially in bone marrow aspirate smears. This has been a challenging problem in computer vision, image processing, and machine learning, because of complex nature of bone marrow aspirate smear. The leukocyte has multiple maturation stages, and these maturation stages have small inter-class differences, so it is difficult to differentiate even with expert knowledge. Moreover, a problem of color, shape, and size variations among samples exists and a problem of touching cell due to high leukocyte density of bone marrow aspirate smear exists. In this dissertation, an automated leukocyte differential count system for bone marrow aspirate smear was developed to overcome problems of manual differential count and to fulfill clinically unmet needs. The system should perform the differential count with high accuracy and objectivity, and high throughput and efficiency. Moreover, it should overcome challenges of bone marrow aspirate smear. To this end, a large dataset of bone marrow smear was collected for development of a detection and a classification algorithms. Watershed transformation and saliency map were utilized for single-leukocyte detection, and the dual-stage convolutional neural network that learns global and local features of complex leukocyte maturation stages was proposed for classification. Lastly, a probability guidance algorithm was proposed for integration of detection and classification algorithms. The performance of proposed system was assessed with ten leukocyte maturation stages of myeloid and erythroid series in bone marrow aspirate smears. Total of 200 large (1388 × 1040) digital images of bone marrow aspirate smears and 2,323 small (96 × 96) single leukocyte digital images were collected. The proposed system showed a state-of-the-art performance. It achieved an average detection accuracy of 96.09% and an average classification accuracy of 97.06%, and it was able to differential count 100 leukocytes in 4 to 5 seconds. This proposes a new paradigm in diagnosis of blood disorder and showed a potential of deep learning, especially the convolutional neural network, in medical image processing. The proposed system is expected to increase the total number of analyzed leukocytes in a sample, which will provide more statistically reliable information of a patient for diagnosis.