Anatomical landmark detection leveraging implicit and explicit information sharing techniques
Year: 2021
Type: dissertation
Abstract: Anatomical landmark detection is an essential step in various medical imaging processes, including morphological analysis, inter-/intra-subject registration, and, fundamentally, anatomy orientation. Deep reinforcement learning (DRL) has shown promise in replacing heuristic methods and classical image processing approaches to landmark detection. In this work, we propose multiple extensions of a multi-agent deep q-network approach to anatomical landmark detection. We first improve the localization of high confidence primary landmarks by searching in the physical space coordinate system of the image rather than voxel space. Second, when detecting a large number of landmarks, we decompose the detection process into two stages to compensate for the memory limitations induced by detecting a large number of landmarks.
DOI: 10.17077/etd.005911
Authors: Alexander Blair Powers, Mathews Jacob, Hans J Johnson, Stephen Baek
Language: English
Topic: Medical Imaging and Analysis
Subfield: Biomedical Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W4386657078