In Search of the Monolayer - Utilizing Deep Learning to efficiently locate the Monolayer on Peripheral Blood Films Abstract

A Peripheral Blood Film (PBF) examination with differential count (DIFF) is a common blood test, yielding quantitative and qualitative information about the blood cells of the patient important for diagnostics. The test relies on locating the monolayer, an area on the PBF slide where the cells are located in one layer. This masters thesis explores the possibilities of adopting Machine Learning (ML) to locate the monolayer on a PBF slide in a commercially available microscopic system for medical analysis. Over two million microscopic images from 972 PBF slides was collected through the system, to be used as data. Several deep learning models including Convolutional Neural Networks (CNNs), Long-Short-Term-Memory Networks (LSTMs) and Transformer encoders were trained to traverse the slide. A regression model constructed using classical machine learning techniques such as Support Vector Regression (SVR), Gradient Boosting (GB) and Random Forests (RF) were developed to be able to predict a good starting position for the search using overview images of the PBF slides. The results showed a great improvement in the efficiency of the search. A combination of using a SVR to find a start position and an LSTM model to traverse the slide produced the best result with a time reduction of over 50% with higher average quality of the monolayers found, compared to the current implemented solution. We also provide an in-depth analysis showing that our models remain robust with good results across a wide range of different PBF types. This study highlights a new use case of machine learning in automated hematology systems.