[Forschungsseminar-BSV] Research Seminar 'Computer Graphics, Image Processing

[Vanessa Kretzschmar]

Attention:
- online only
- more than one talk (increased duration possible)


I N V I T A T I O N

======================================================================

to the Research Seminar 'Computer Graphics, Image Processing, and
Visualization'

     on Wednesday, November the 30th, 2022, at 01:15 PM,
     online via conf.fmi.uni-leipzig.de/b/van-bwb-jil-vk2

======================================================================

A talk will be given by

     Felix Diez

and is entitled:

     'Using feed forward neural networks to predict stroke outcomes 
based on clinical data'

Abstract:

     Predictions about the outcome after a stroke could help both patients
     as well as doctors. Treatment options could be prepared and ressources
     could be distributed to the patients that need them the most. This 
thesis
     tested if feed forward neural networks could be used to deliver such
     predictions, based on three different datasets. One with 114 patients
     and five attributes, the second one with the same 114 patients but 
seven
     attributes and a third one with 268 patients and seven different 
attributes.
     The prediction variable for the first two, was the modified Ranking 
Scale
     (mRS) at a three month follow-up session, whereas for the third one it
     was the mRS at discharge from the hospital. A secondary objective was
     to determine which of the input attributes contributed most to the
     predictions. Different models were trained and evaluated on their
     respective datasets. During the evaluation, it became obvious that the
     models were not able to predict all seven mRS classes. As a result, it
     was decided to dichotomize the target variable (mRS > 2). Both of the
     models for the first two datasets were able achieve a mean prediction
     accuracy of 0.84 for the dichotomized data. However, the third model
     showed no learning success and no prediction capabilities beyond random
     guessing. After evaluating the contribution of the models attributes,
     it turned out that both models are heavily influenced by the patient’s
     physical state before the three month period. It also made no 
difference
     in what form this information is given (mRS and National Institute of
     Health Stroke Scale (NIHSS) were tested). In conclusion, this thesis
     showed that neural networks can be used to predict dichotomized mRS
     values, given clinical training data. Furthermore, the networks are
     dependent on information about the patient’s physical state at the time
     of discharge in order to predict the follow-up outcome.

======================================================================

Another talk will be given by

     Johanna Soest

and is entitled:

     'Predicting Stroke Outcomes based on medical imaging data using 
Convolutional Neural Networks'

Abstract:

     Convolutional Neural Networks have recently become more popular, 
especially
     in the field of computer vision and medicine. This thesis would like to
     add to this success by implementing a CNN that can predict the physical
     limitations of stroke patients, by combining medical imaging data and
     clinical data. The thesis would like to answer, whether CNNs can be 
used
     to make such prediction given two different data sets. The data sets in
     question consist of medical imaging scans, binary maps showing the 
exact
     position of the affected tissue, structural and functional lesion 
network
     maps as well as clinical data. The model trained on the first data set
     is expected to predict the mRS > 2 at a three-month follow-up session,
     given information about the patient at the point of admission and 
discharge
     from the hospital. The model trained on the second data set is going to
     predict the mRS > 2 at the time of discharge given information 
about the
     patient premorbid and at the time of admission. Both data sets are
     relatively small in size with 114 and 168 patients respectively. This
     became a serious problem as both networks failed to learn from the 
imaging
     data. Although learning success could be achieved from the first data
     set, if given the clinical data with or without imaging data. Possible
     reason as to why the image models failed to learn could be the size of
     the patient pool or the limited complexity of the network, that was
     chosen because of the small data set at hand.

======================================================================

On behalf of Professor Scheuermann all those interested are cordially
invited to attend.

Yours sincerely
Vanessa Kretzschmar
-------------- nächster Teil --------------
Ein Dateianhang mit Binärdaten wurde abgetrennt...
Dateiname   : 20221130T131500_bsv_invitation.ics
Dateityp    : text/calendar
Dateigröße  : 4672 bytes
Beschreibung: nicht verfügbar
URL         : <http://lists.informatik.uni-leipzig.de/pipermail/forschungsseminar-bsv/attachments/20221128/39cc22ed/attachment.ics>