Dear all,
you are cordially invited to join the master's thesis defense of David Hartmann on 26.11.21 at 10:00. You can join via Zoom as follows:
https://uni-bamberg.zoom.us/j/99191342076
Meeting-ID: 991 9134 2076
Passwort: 7Y^b@?
David Hartmann will defend his thesis about Assessing uncertainty for CNN classification -- A case study with histopathology data.
He conducted his research in cooperation with Fraunhofer IIS.
Abstract:
In the context of digital histopatholgy, convolutional neural networks (CNNs)
are used to classify hematoxylin and eosin stained tissue samples according
to the types of tissue contained within them. Two factors that may hinder
network model performance are blurry sections as well as parts of the sample
which contain less actual tissue but debris due to damage caused during the
preparation process. This thesis aims to assess the influence those two factors
have on CNN performance and to link them to model uncertainty by designing
ways to automatically identify their presence.
For this, a blur detection CNN was trained and evaluated, which reached an
accuracy score of 99.89%. Additionally, an algorithm was developed and tested,
which analyzes the standard deviation of foreground pixel positions in image
patches extracted from a histological slide image in order to categorize them
with regards to the amount of tissue they contain. This approach achieved an
accuracy of 95%. Both factors were found to have a negative influence on tissue
type classification, CNN model performance and uncertainty. In the case of
the latter, however, the amount of tissue present was found to be much more
significant than that of the presence of blur. Finally, a number of possible ways
to communicate these findings to an end user by using overlays in a slide viewer
are suggested.
are used to classify hematoxylin and eosin stained tissue samples according
to the types of tissue contained within them. Two factors that may hinder
network model performance are blurry sections as well as parts of the sample
which contain less actual tissue but debris due to damage caused during the
preparation process. This thesis aims to assess the influence those two factors
have on CNN performance and to link them to model uncertainty by designing
ways to automatically identify their presence.
For this, a blur detection CNN was trained and evaluated, which reached an
accuracy score of 99.89%. Additionally, an algorithm was developed and tested,
which analyzes the standard deviation of foreground pixel positions in image
patches extracted from a histological slide image in order to categorize them
with regards to the amount of tissue they contain. This approach achieved an
accuracy of 95%. Both factors were found to have a negative influence on tissue
type classification, CNN model performance and uncertainty. In the case of
the latter, however, the amount of tissue present was found to be much more
significant than that of the presence of blur. Finally, a number of possible ways
to communicate these findings to an end user by using overlays in a slide viewer
are suggested.