| dc.description.abstract | A novel federated computational workflow for analyzing digital palynological slide images is implemented in this thesis. The slide data files, typically exceeding 3GB, present significant data mobility and computation challenges. The novel distributed computational framework is implemented to address privacy concerns and the challenges associated with moving large data. The idea is to move computational to the data location, optimally utilizing local computational capacity and reducing data movement.
Trained deep-learning models deployed in a containerized environment leveraging the Docker technology are integrated in the workflow with a user-friendly interface, and users can run processes with the trained models.\\
The workflow processes include reading slide image files, generating tiled images, and identifying and removing undesirable tiles such as blank tiles. Object detection with the watershed segmentation algorithm identifies tiles with potential microfossils. The identified dinoflagellates are classified with a trained convolution neural network (CNN) model. The classification results are sent to the host and shared with the users.
The federated computational approach effectively addresses the challenges related to moving and handling large palynological slide images, creating a more efficient, scalable, and distributed pipeline. Collaborative efforts involving domain experts for model training with more annotated slide images will improve the effectiveness of the workflow. | |
