When pharmaceutical companies develop new drugs or vaccines there are large amounts of data in the form of images that need to be analysed, and any automation of that process is helpful to reduce time. These analyses could be for example concentration, decomposition, or classification, and essential to all these is high-quality particle localisation and segmentation. Therefore this thesis will focus on semantic segmentation of images of particles and viruses from a TEM microscope. Various kinds of CNNs have been shown to give promising results in this area, however, there is still a need for improvement of these methods. Therefore it is interesting to see if combining a modern CNN with a graph-based model would improve its performance. This thesis proposes a combination of a CNN, a modified version of the U-net, and a graph-based method, the CRF-RNN. The CRF-RNN is appended to the modified U-net and they are merged to create an end-to-end trainable segmentation network. This is then compared to using the standalone modified U-net to see if the CRF-RNN improves the accuracy. Various loss functions, activation functions, and dropout rates are then tested to see which gives the best results under the given conditions. For both models, the best out of the tested hyperparameters were the dice loss and no dropout layers at all. For the standalone modified U-net the optimal activation function was the sigmoid function. However, for the network with the CRF-RNN addition, both softmax and sigmoid were good in different aspects. Experiments done show that the model with the CRF-RNN addition performs slightly better than the model without, based on both measured metrics and visual inspection of the predicted outputs. Therefore it can be concluded that the CRF-RNN does improve the network it is attached to in this case. There is still much that could improve the networks though, for example parameter tuning of more hyperparameters using a GA.
| Identifer | oai:union.ndltd.org:UPSALLA1/oai:DiVA.org:mdh-59368 |
| Date | January 2022 |
| Creators | Svens, Lisa |
| Publisher | Mälardalens universitet, Akademin för innovation, design och teknik |
| Source Sets | DiVA Archive at Upsalla University |
| Language | English |
| Detected Language | English |
| Type | Student thesis, info:eu-repo/semantics/bachelorThesis, text |
| Format | application/pdf |
| Rights | info:eu-repo/semantics/openAccess |
Page generated in 0.0093 seconds