| Summary: | Chemotherapy is a type of cancer treatment that uses drugs to kill the cancer cells. In order to evaluate the drug performance, we need cell death assays. Today, most cell death assays are based on fluorometric detection, which can cause phototoxic effects. To overcome the problem of phototoxicity, this thesis aims at examining new approaches that are not relying on fluorometric detection. Using deep learning for cell death classification from phase contrast images, we examine how well semi-supervised learning performs compared to supervised learning. Besides, we investigate if model generalization depends on the cell type: A549, AU565, and HeLa. Using SimCLR for self-supervised pre-training, we examine if our model can distinguish between dead and live cells (binary problem), and further whether the cells died from apoptosis or other cause (3-class problem). With supervised learning, our model achieves an accuracy of 76.1% and 53.6% on the binary and 3-class problem respectively. We demonstrate that the accuracy increases to 84.3% and 65.6% on respective problem using semi-supervised learning. The model generalizes most to HeLa and least to AU565, due to differences in cell type size. Nevertheless, semi-supervised learning can improve model generalization. Given that semi-supervised learning outperforms supervised learning, we conclude that there is a great potential of semi-supervised learning for cell death assays. However, we experience that the applied data augmentation in SimCLR is insufficient on phase contrast images. Future work should therefore focus on finding proper data augmentation.
|
|---|
