Browsing by Author "Lehtonen, Oskari"

Filter results by typing the first few letters
Now showing 1 - 2 of 2
  • Fully Convolutional Neural Networks for Nuclei Segmentation and Type Classification
    (2021-06-14) Lehtonen, Oskari
     Perustieteiden korkeakoulu | Master's thesis
    The examination of histological microscopic images is the golden standard in the diagnosis and grading of cancer. The conventional approach to perform diagnosis and grading is a low throughput manual examination of the histological images by a pathologist. This maunal assessment is labor-intensive and suffers from intra- and inter-observer variability which can lead to sub-optimal care for the patients. However, over the past years, advances in digital histopathology and deep learning have enabled the automated digital analysis of histological images. Often, a crucial prequisite in the digital analysis of the histological images is the segmentation and cell type classification of the cell nuclei. The most well known challenge in nuclei segmentation is the separation of clustered nuclei that is especially apparent in cancer cells that often form irregularly bounded clusters of many overlapping cells. Furthermore, the fact that different cell types display extensive intra-class variability complicates the accurate cell type classification of the nuclei that is crucial in the downstream analyses. Many methods have been developed to address nuclei segmentation and cell type classification as two separate problems, but recently a new class of fully convolutional multi-task neural networks have shown great promise in solving both the nuclei segmentation and type classification simultaneously. This thesis introduces a comprehensive benchmarking of the different components that are relevant for multi-task networks that are performing simultaneous nuclei segmentation and type classification. In total, eight ablation studies are conducted for distinct components. For each of these components, several different alternatives are tested to determine the optimal combination that leads to the best end result. The results show that the model performance can be increased over 7\% by choosing the right compnents.
  • Sytometria ja Bayes-verkot solunsisäisen viestinnän mallintamisessa
    (2017-09-15) Lehtonen, Oskari
     Sähkötekniikan korkeakoulu | Bachelor's thesis