Automatizovaná analýza krevních buněk pomocí strojového učení

Abstract

This master's thesis deals with the automated analysis of peripheral blood smears using deep learning methods to accelerate and refine hematological diagnostics. In the first phase, the classification of isolated leukocytes into five classes was performed using the GoogLeNet, ResNet-101, and MobileNetV2 convolutional neural networks, utilizing transfer learning and data oversampling techniques. It was experimentally proven that the lightweight MobileNetV2 architecture achieves the best results. The second part of the thesis is dedicated to the spatial localization and detection of blood elements in complex microscopic images using single-stage detectors of the YOLOv8 family. The results confirmed that these lightweight detectors effectively overcome the problems of mechanical cell overlapping and data imbalance. The proposed models thus represent a robust and computationally undemanding foundation for clinical decision support systems (CAD), applicable even on edge devices (Edge AI).