Visual saliency modelling for medical images

Lou, Jianxun 2024. Visual saliency modelling for medical images. PhD Thesis, Cardiff University. Item availability restricted.

	PDF - Accepted Post-Print Version Restricted to Repository staff only until 15 July 2025 due to copyright restrictions. Available under License Creative Commons Attribution Non-commercial No Derivatives. Download (16MB)
	PDF (Cardiff University Electronic Publication Form) - Supplemental Material Restricted to Repository staff only Download (340kB)

Abstract

Visual attention is a pivotal mechanism of the human visual system (HVS), which allows humans to select and interpret the most relevant information in the visual scene. Visual saliency, a direct manifestation of human observation, effectively reflects the visual attention. In the contexts of medical imaging, the visual saliency generated by radiologists during the interpretation of medical images can reflect their perceptual-cognitive processes involved in diagnostic decision-making. Despite extensive research into visual saliency modelling, most studies focus on non-medical images, leaving persistent challenges in medical image saliency modelling. This thesis investigates computational modelling of visual saliency in medical images. The initial exploration was focused on performing visual saliency modelling that is perceptually more relevant from a generic perspective, leveraging deep learning-based methods. A systematic comparative analysis was then conducted to establish a plausible modelling paradigm for medical image saliency, which involves studying a range of saliency models in the specialised context of screening mammography. Based on these findings, a new visual saliency model was developed to predict radiologists’ visual attention when examining mammogram images, and this model was expanded for time-interval visual saliency prediction for mammogram images. Subsequently, research was expanded to Chest X-ray (CXR) images. A reliable visual saliency dataset for CXR images was established and analysed, and a novel semi-supervised framework for CXR image saliency prediction was devised to mitigate data insufficiency. In this thesis, comprehensive new guidelines were introduced to inform the design of better saliency models for medical images. Furthermore, superior performance visual saliency models and a reliable visual saliency dataset were developed for medical applications. This thesis demonstrates that the visual saliency generated by radiologists while reading medical images can be reliably modelled by computational models, which can benefit from appropriate deep learning-based modelling methods.

Item Type:	Thesis (PhD)
Date Type:	Completion
Status:	Unpublished
Schools:	Computer Science & Informatics
Funders:	China Scholarship Council (CSC)
Date of First Compliant Deposit:	15 July 2024
Date of Acceptance:	8 July 2024
Last Modified:	18 Jul 2024 15:33
URI:	https://orca.cardiff.ac.uk/id/eprint/170495

Actions (repository staff only)

Edit Item