BME PhD Preliminary Exam Announcement for Mohseu Rashid Subah (R. Surowiec, advisor) Everyone is invited to attend the public presentation beginning at 2:00 PM. Title: Advancing Multi-Modal Representation Learning in Biomedical Imaging for Segmentation and Disease Characterization Date: December 12th, 2025 Time: 2:00 PM Location: MRGN 129 Microsoft Teams Link: Mohseu Rashid Subah Preliminary Exam Meeting Link (Teams)<https://teams.microsoft.com/l/meetup-join/19%3ameeting_YWEyMTVlZDQtNzdkYi00NGU4LWJmYTItNmM1NTliZTkxZDZk%40thread.v2/0?context=%7b%22Tid%22%3a%224130bd39-7c53-419c-b1e5-8758d6d63f21%22%2c%22Oid%22%3a%227e74c2d8-de69-4d71-9c6b-5fed3736d6e9%22%7d> Committee Members: Dr. Rachel Surowiec (Chair), Dr. Craig Goergen, Dr. Young Kim and Dr. Christopher Newman Abstract: With the advancement of artificial intelligence (AI) in computer vision, deep learning-based architectures have enabled precise analysis of medical images for tasks such as segmentation, classification, reconstruction, and generation, to support downstream clinical applications. Despite this progress, existing AI models often require substantial expert knowledge, depend heavily on large-scale labeled datasets, cannot integrate multimodal information effectively, and lack generalization across domains. These limitations highlight the need for label-efficient, multimodal deep learning methods capable of robust segmentation and disease characterization across diverse biomedical imaging modalities. We hypothesize that integrating domain adaptation, semi-supervised, few-shot, and multimodal learning can yield generalizable image representations that enable accurate segmentation and classification with substantially fewer labels. To address this research gap, we will first develop an AI framework that combines task-specific segmentation with radiomics feature extraction to characterize osteoporosis from high-resolution CT scans, validating the effectiveness of machine and deep learning methodologies in a clinically relevant setting. Next, to reduce the dependence on costly manual annotations and improve cross-modality performance, we will design a mixed-domain semi-supervised segmentation pipeline that leverages state-of-the-art domain adaptation strategies and foundation model priors. Finally, to further enhance the representation capacity of standard deep learning models for label-efficient medical image classification, we will explore few-shot learning approaches built on multimodal vision-language models. Successful completion of the proposed work will demonstrate the feasibility of label-efficient, domain-agnostic, and multi-modal representation learning for segmentation and disease characterization in biomedical images. -- Bmeroundtable-list mailing list Bmeroundtable-list@ecn.purdue.edu https://engineering.purdue.edu/ECN/mailman/listinfo/bmeroundtable-list