High-Order and Interactive Perceptual Feature Learning for Medical Image Retargeting

Computed tomography (CT) imaging is a valuable technique for detecting lesions and injuries in internal structures of the human body. Visual retargeting focuses on resizing less significant areas, such as normal organs, either horizontally or vertically, while preserving the dimensions of key regions, such as affected organs. This process enhances the display of CT images, which can be particularly beneficial for applications like medical education, enabling optimal presentation of CT images to students. This study explores an innovative approach to CT image retargeting by effectively fusing gaze shift path (GSP) features using a hypergraph framework to address the complex spatial layout of CT images. To capture both semantically and visually significant elements in each CT image, the BING objectness descriptor is employed to rapidly and accurately localize objects or their components across multiple scales. Each identified object-aware patch is then described using multi-channel low-level features. To model human attention and its perception of important elements within a CT image, a locality-preserved and interactive active learning (LIAL) approach is introduced. This method sequentially generates GSPs for each CT image, which represent human gaze trajectories. Given the high dimensionality of these features, a hypergraph fusion enables the representation of higher-order relationships among features, capturing complex dependencies and interactions between the GSP features. Through this framework, the CT image retargeting process benefits from enhanced feature fusion, leading to more accurate identification of critical areas. Extensive empirical evaluations demonstrate the efficacy of this hypergraph-based method in improving CT image retargeting outcomes, offering significant advancements for medical imaging and educational applications.