Date of Award
Master of Science (MS)
Electrical and Computer Engineering (Holcomb Dept. of)
In order to discover nuclei of cells, biological scientist depend on various types of dyes, and coupled with the overall time involved in using these, it makes this a very inefficient process. Identifying nuclei with Deep Learning techniques can save a lot of time and can yield better accuracy.
Recent developments in Generative Adversarial Networks(GANs) have been shown to produce text, video and realistically looking synthetic images. We use the Conditional Generative Adversarial Networks to produce Nuclei Stained images from label maps using two different Generator Architectures(U-Net and Encoder-Decoder with ResNet) and evaluate the performance of these networks. We also evaluate the classical Segmentation Network as a Generator and further improve it by adding Residual Blocks. Different evaluation measures like Pearson Correlation Coefficient, Mean Squared Error and Structural Similarity have been used along with software Cell Profiler to count the number of cells in the generated images and the corresponding real labels.
This Thesis makes several contributions to the field of Deep Learning and using it for the generation of nuclei stained images which makes it useful to detect various cell lines including cancerous cell lines.
Bhatia, Tanmay, "Semantic Segmentation for Producing Nuclei Stained Images Using Conditional Generative Adversarial Networks" (2019). All Theses. 3213.