Subah is a Ph.D. candidate at the Weldon School of Biomedical Engineering and an American Association of University Women (AAUW) International Doctoral Degree Fellow (2025–2026). She received her undergraduate degree in Electrical and Electronic Engineering (EEE) from the Bangladesh University of Engineering and Technology (BUET), with a major in Communication and Signal Processing. Her research investigates label-efficient learning strategies for medical data, where annotations are scarce and expensive. Specifically, she explores semi-supervised and few-shot learning frameworks for segmentation and disease characterization, including the use of foundation models such as DINO, MedSAM, and CLIP, to enable knowledge transfer under limited supervision. Her work spans a range of imaging modalities, including HR-pQCT, MRI, ultrasound, DXA, and OCT. In parallel, she studies domain adaptation using generative adversarial networks to improve generalization across datasets, and complements deep learning with interpretable, classical machine-learning approaches to support clinical relevance and translational impact. In summer 2026, she interned with the Product Development Data Science and Analytics group at Genentech within the Ophthalmology therapeutic area, where she developed a pretrained foundation model-based segmentation algorithm for medical imaging data. When not engaged in research, she enjoys reading and traveling. Our group aims to identify imaging biomarkers of risk and resilience to musculoskeletal decline in aging and fracture-prone conditions. We are advancing radiomics and machine learning techniques to extract critical imaging features, transforming high resolution peripheral quantitative computed tomography (HR-pQCT) images into rich, quantitative datasets. To achieve this, we developed a deep learning-based pipeline for multi-class HR-pQCT segmentation, which separates not only cortical and trabecular bone but also the adjacent soft tissues such as muscle and fat. Building on these segmentations, we extract quantitative radiomic features from each tissue region and use them to train machine learning classifiers for osteoporosis detection, benchmarked against dual-energy X-ray absorptiometry (DXA)-based diagnoses. Notably, features derived from muscles proved especially informative, offering strong discriminative power between osteoporotic and non-osteoporotic patients and, in some cases, exceeding the performance of bone-based radiomics. These findings suggest that soft tissue surrounding bone carries meaningful signals that could help advance future fracture risk and osteoporosis screening tools.