Degree

Doctor of Philosophy (PhD)

Department

Electrical and Computer Engineering

Document Type

Dissertation

Abstract

The growing convenience of complex biomedical data begins new roads for better disease detection and functional identification via artificial intelligence (AI). Nevertheless, conventional analysis methods often rely on basic metrics that drop sensitive biotic differences, and various AI systems are difficult to infer, limiting their clinical reliability and practical use. There is a growing need for explainable, physiologically relevant computational models that can extract key biomarkers from diverse biomedical data sources. This dissertation addresses this problem by obtaining explainable machine learning and deep learning procedures for studying biomedical signals and optical imaging data.

This dissertation is divided into two parts; in the first part, Hypopharyngeal cancer (HPC) is detected in a murine model using the NIR-I imaging modality, which is integrated with interpretable machine learning. Two methods are used to analyze the NIR-I data: first, a tumor-to-background ratio (TBR)-based method, which is a commonly used method in NIR-I imaging, and second, an AI-based method. AI-based method performs better compared to the TBR-based method for NIR-I data, showing the significance of interpretable AI for detection of subtle patterns that represent the dynamic tumor microenvironment.

The second work explores human brain responses using electroencephalography (EEG) when music is used as a stimulus to study neurophysiological pattern changes during auditory processing. With signal preprocessing, spectral and temporal feature extraction combined with traditional machine learning provides a path to understanding changing brain states when different music stimuli are used rather than standard EEG analysis based on final decisions.

Lastly, idiopathic absence seizure (IAS) is detected with a proposed explainable convolutional channel ranking (ECCR) from EEG. This method obtained a deep interpretation of layers of deep learning for ranking EEG channels, which represents the physiological meaning of the high-ranked channels involved in IAS detection. This method identifies neural regions associated with the IAS and are relevant to seizures.

In general, all investigations across different datasets are based on a unified explainable AI framework that involves NIR imaging and human brain signals represented with electroencephalography. By centering on interpretability, biological significance, and profound biomarker findings, this work moved one step towards an advanced and transparent AI framework in healthcare.

Date

7-17-2026

Committee Chair

Jian Xu

LSU Acknowledgement

1

LSU Accessibility Acknowledgment

1

Download

Share

COinS