Analyzing white blood cells using deep learning techniques

• Main Authors: Neelakantan, Suraj, Kalidindi, Sai Sushanth Varma
• Format: Others
• Language: English
• Published: Högskolan i Halmstad, CAISR Centrum för tillämpade intelligenta system (IS-lab) 2020
• Subjects: White blood cells; Deep learning; Convolutional neural networks; Classification; Medical device technology; Medical Equipment Engineering; Medicinsk apparatteknik
• Online Access: http://urn.kb.se/resolve?urn=urn:nbn:se:hh:diva-43283

The field of hematology involves the analysis of blood and its components like platelets, red blood cells, white blood cells. The outcome of this analysis can be vital in determining the condition of the human body and it is important to obtain accurate results. A deep learning algorithm scans over the given input data for unique features and learns them. Then it identifies these features and correlates them to give the result. This can save a significant amount of time and manual work. In contrast, a traditional machine learning algorithm requires the developer to carry-out the feature engineering. This thesis involves the analysis of white blood cells (WBC) using deep learning techniques. In collaboration with a hematology company HemoCue AB based in Angelholm, we will be developing deep learning algorithms for the analysis of white blood cells in the HemoCue R WBC DIFF System. Predominantly, there are two stages in this thesis. The first stage is white blood cell identification, which is used to calculate the number of white blood cells in the given blood sample. The next stage is to identify the different types of white blood cells with which the concentration of each type of WBC in the given blood sample is calculated. We have explored different classification approaches like ’one vs all’ and ’4-class classifier’, and have developed two CNN architectural designs i.e. ’multi-input’ and ’multi-channel’. On comparing the performance of all these design approaches, a final integrated model is put forth for the analysis of WBCs in the company’s device. The proposed ’one vs all’ classification approach combined with a 3-class CNN classifier has yielded very promising results with a combined accuracy 95.45% in WBC identification and 90.49% in WBC differential classification.