Enhancing Ultrasound Imaging with Deep Learning
Khan, Ufaq
Khan, Ufaq
Author
Supervisor
Department
Computer Vision
Embargo End Date
2025-05-30
Type
Thesis
Date
2025
License
Language
English
Collections
Research Projects
Organizational Units
Journal Issue
Abstract
Ultrasound imaging is an essential tool in numerous clinical fields, especially where repeated, realtime assessments are needed, such as prenatal monitoring and cancer detection. Its advantages, including low cost, portability, and lack of ionizing radiation, make it highly practical in large hospitals. This thesis introduces two frameworks, FETR and Ultra Weak, designed to enhance ultrasound image analysis with minimal annotations. FETR uses transformers and Multiple Instance Learning (MIL) for fetal anatomy detection, achieving significant localization improvements on the FPUS23 and Fetal Plane DB datasets. Ultra Weak employs MIL with a Deformable DETR architecture for breast ultrasound lesion detection, adapting to the varied appearance of lesions and outperforming other models in detecting malignancies with limited supervisory data. These methods demonstrate effective strategies to address the challenges of inconsistent image quality and annotation demands in ultrasound imaging.
Citation
Ufaq Khan, “Enhancing Ultrasound Imaging with Deep Learning,” Master of Science thesis, Computer Vision, MBZUAI, 2025.
Source
Conference
Keywords
Ultrasound Imaging, Multiple Instance Learning (MIL), Fetal Anatomy Detection, Breast Lesion Detection, Transformers