Off-campus Michigan Tech users: To download campus access theses or dissertations, please use the following button to log in with your Michigan Tech ID and password: log in to proxy server

Non-Michigan Tech users: Please talk to your librarian about requesting this thesis or dissertation through interlibrary loan.

Date of Award


Document Type

Campus Access Master's Thesis

Degree Name

Master of Science in Mechanical Engineering (MS)

Administrative Home Department

Department of Mechanical Engineering-Engineering Mechanics

Advisor 1

Susanta Ghosh

Committee Member 1

Shiva Rudraraju

Committee Member 2

Benjamin W. Ong


Breast cancer is the most common cause of cancer in women. Histopathological imaging data can provide important information on cancer since it preserves the underlying tissue architecture in the preparation process. Accurate and automated classification of breast tissue into malignant or healthy from histological images can be used for the diagnosis of breast cancer. However, publicly available labeled histopathological datasets are limited in size and also biased. For such datasets, existing machine learning classifiers have shown limited success. The goal of the present work is to develop a classification technique using machine learning, which can overcome the challenge posed by small and biased datasets. This technique will reduce the inaccuracy of the image analysis, and quantify the uncertainty in their prediction.

The field of computer vision and neural networks are aimed at improving the accuracy of image analysis by various network architectures on algorithms. Plain feed-forward neural networks have been successfully used for pattern recognition, but their performance is good only when the data is sufficiently large. When the data size is limited, neural networks yield erroneous or overfitted results, since they don't take the uncertainty of the dataset into account. These feed-forward neural networks learn their weights as point estimates or a deterministic value. Whereas, a Bayesian neural network learns a probability distribution on the weights. The loss function used in Bayesian neural networks is known as a Variational Free Energy (VFE) or Evidence Lower Bound (ELBO) which is to be optimized. Since the Bayesian approach provides probability distributions on the weights of the neural networks, it is possible to calculate the variance of the predictive posterior probability distribution, which is the sum of aleatoric and the epistemic uncertainty. Uncertainty quantification, along with the point estimate, leads to a more informed decision, improved accuracy, and reduced overfitting. In critical applications, especially medical-imaging applications, uncertainty quantification can potentially reduce the unexpected and incorrect results due to the poor decisions. In this thesis, we have worked on the application of Bayesian neural networks on publicly available histopathological images for the detection of breast cancer and uncertainty quantification of the prediction. We have demonstrated that using the Bayesian CNN, the false-negative predictions can be reduced remarkably, by almost 22%. We have found that the predictions associated with higher epistemic uncertainties have features of both the classes. These findings should improve the state of the art machine learning-based biomedical imaging.