Towards Real-time Detection in Medical Imaging
Sheikh, Tooba Tehreem
Sheikh, Tooba Tehreem
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
Accurate object detection in medical imaging is vital but challenging due to its complexity and the limitations of traditional detectors. This thesis introduces two realtime detection methods tailored for medical imaging: IHA-YOLO and Med-YOLOWorld. IHAYOLO enhances YOLOv10 with selfattention and improved feature blocks, achieving a 2.03% m AP50 gain on five cell detection datasets while maintaining high speed. To extend object detection beyond predefined categories, we introduce Med-YOLOWorld, the first openvocabulary detection in medical imaging using visionlanguage alignment, trained on the diverse Omnis 600K dataset across nine modalities. It outperforms existing closedset detectors by 3 m AP while running in real time. Future directions include improved generalization, openvocabulary test benchmarks, and extending to segmentation tasks.
Citation
Tooba Tehreem Sheikh, “Towards Real-time Detection in Medical Imaging,” Master of Science thesis, Computer Vision, MBZUAI, 2025.
Source
Conference
Keywords
Object Detection, Medical Image Analysis, Open Vocabulary