Author ORCID Identifier

https://orcid.org/0009-0008-3396-2712

Semester

Summer

Date of Graduation

2024

Document Type

Thesis

Degree Type

MS

College

Statler College of Engineering and Mineral Resources

Department

Lane Department of Computer Science and Electrical Engineering

Committee Chair

Gianfranco Doretto, Ph.D

Committee Co-Chair

Donald Adjeroh, Ph.D

Committee Member

Donald Adjeroh, Ph.D

Committee Member

Nima Karimian, Ph.D

Abstract

In recent years, there has been an upward trend of utilizing deep learning to automate cell segmentation processes. As global storage capacities grow exponentially, so have microscopy data collections become larger and more frequent. To benefit from them, accurate and precise quantitative analysis tools like cell instance segmentation have become necessary. However, the highly variable nature of these data collections necessitates retraining segmentation models to maintain high accuracy on new data collections. This process is time-consuming and labor-intensive since a user must annotate much of the new data, usually under the supervision of a medical professional. The problem is further exacerbated when segmenting cells with elongated and non-convex morphology, like bacteria cells.

In mitigating these concerns, we propose reducing the amount of annotation and compute power needed to retrain the model by introducing a few-shot domain adaptation approach that requires the annotation of only one to five cells of the new data. First, we rely on a robust and precise segmentation method trained on extensive source data and capable of handling highly diverse cell morphologies. Second, we take a few-shot learning approach, where given a target dataset that is distributed differently from source data, we require a user to label only a minimal amount of target data. We then set up a contrastive prediction task by introducing new losses that pull the representation of positive samples in a target domain closer to samples of the same class in a source domain while simultaneously pushing them apart from negative source samples using kernels as a similarity measure. Furthermore, we comprehensively studied the best kernel composition method for combining kernels defined on two inhomogeneous pairs of quantities. Our approach quickly adapts the model to maintain high accuracy, and our results show a significant boost in accuracy after adaptation to very challenging bacteria datasets.

Comments

Thank you God!

Share

COinS