Translation:
Computed tomography (CT) plays a critical role in medical imaging
as it assists doctors in identifying brain disorders such as hemorrhages
and malignant tumors. This thesis explores the potential of deep learning models as a tool for the detection of brain diseases
in medical imaging science. By first creating a convolutional neural
network model that can identify brain hemorrhage and then transfer it
to the neuromorphic processor Akida AKD1000, enabled the use of "spiking" neural networks and relearning of the model on the hardware (on-edge). In a process called few-shot learning, the model was trained
to also be able to identify brain tumors with only a minimal number of additional computed tomography images. Furthermore, it was investigated how the parameters
which is used in the relearning on the neuromorphic hardware affected
the classification accuracy. It was shown that the choice of parameters and their interaction introduced a trade-off when it came to the accuracy of
the bleeding and tumor classification models, but an optimal constellation
of parameters could be extracted. These results are intended to serve as
a basis for future work in medical imaging with neuromorphic hardware, especially in the area of few-shot learning and on-edge learning.
Classification models like these have the potential to streamline diagnosis
of brain diseases and thus improve the accuracy of diagnosis of
brain diseases, as well as relieve medical professionals.