|
|
Deep Learning Model Uncertainty in Medical Image Analysis
Drevický, Dušan ; Kolář, Martin (oponent) ; Kodym, Oldřich (vedoucí práce)
This thesis deals with quantifying uncertainty in the predictions of deep learning models. While they achieve state of the art results in many areas of computer vision, their outputs are usually deterministic and provide by themselves little information about how certain the model is about its prediction. This is important especially in the domain of medical image analysis where mistakes are costly and the ability to filter uncertain predictions would allow a supervising physician to review the relevant cases. This work applies several different uncertainty measures developed in recent research to deep learning models trained on a cephalometric landmark localization task. They are then evaluated and compared in a set of experiments which aim to determine whether each of the uncertainty measures provides us with useful information about the model's confidence in its predictions.
|
|
|
Deep Learning Model Uncertainty in Medical Image Analysis
Drevický, Dušan ; Kolář, Martin (oponent) ; Kodym, Oldřich (vedoucí práce)
This thesis deals with quantifying uncertainty in the predictions of deep learning models. While they achieve state of the art results in many areas of computer vision, their outputs are usually deterministic and provide by themselves little information about how certain the model is about its prediction. This is important especially in the domain of medical image analysis where mistakes are costly and the ability to filter uncertain predictions would allow a supervising physician to review the relevant cases. This work applies several different uncertainty measures developed in recent research to deep learning models trained on a cephalometric landmark localization task. They are then evaluated and compared in a set of experiments which aim to determine whether each of the uncertainty measures provides us with useful information about the model's confidence in its predictions.
|
|
|
Segmentation of cells from microscopic images
Lašan, Michal ; Soukup, Jindřich (vedoucí práce) ; Blažek, Jan (oponent)
V této práci prezentujeme novou metodu na automatickou segmentaci savčích rakovinových buněk z časozběrných snímků pořízených mikroskopem na bázi fázového kontrastu. Tato metoda je sledem kroků založených na základních technikách z oblasti zpracování obrazu, matematické morfologie a teorie grafů. Její hlavní myšlenkou je využít přítomnosti halo artefaktů kolem buněk, díky nimž jsou hranice mezi buňkama světlejší než zbytek obrázku. Navazuje na metodu navrženou Jindřichem Soukupem, která umí oddělovat buňky od pozadí. Srovnáme tuto metodu s watershed - věřejně dostupným algoritmem z odvětví matematické morfologie. Jako referenci použijeme segmentaci lidským expertem. Prezentována metoda je implementována v MATLABu a Javě s jednoduchým a intuitivním rozhraním. Také připojujeme přímočarý editor segmentace, pomocí něhož užívatel může napravit nepřesnosti segmentace, nebo dokonce vytvořit svou vlastní manuální segmentaci. Powered by TCPDF (www.tcpdf.org)
|
|
|
Approximating Probability Densities by Mixtures of Gaussian Dependence Trees
Grim, Jiří
Considering the probabilistic approach to practical problems we are increasingly confronted with the need to estimate unknown multivariate probability density functions from large high-dimensional databases produced by electronic devices. The underlying densities are usually strongly multimodal and therefore mixtures of unimodal density functions suggest themselves as a suitable approximation tool. In this respect the product mixture models are preferable because they can be efficiently estimated from data by means of EM algorithm and have some advantageous properties. However, in some cases the simplicity of product components could appear too restrictive and a natural idea is to use a more complex mixture of dependence-tree densities. The dependence tree densities can explicitly describe the statistical relationships between pairs of variables at the level of individual components and therefore the approximation power of the resulting mixture may essentially increase.
|
host ::
RSS.