Cross-modality guided Image Enhancement

Abstract

The quality of medical images is a crucial factor that affects the performance of several image analysis tasks. Low contrast and noise are among the widely investigated distortions in medical image enhancement problems. In this thesis, the approaches to improve the contrast of medical images and reduce the noise have been proposed by particularly investigating how the cross-modal guidance from another medical image impacts the enhancement. We are particularly interested in enhancing Computed Tomography (CT) and Magnetic Resonance Imaging (MRI) which are widely used in both diagnosis and therapy planning. The first section of the thesis focuses on contrast enhancement and the second section focuses ondenoising. This dissertation presents our research work supported by six original publications (including five published papers and one accepted for publication). First, in the context of cross-modality guided contrast enhancement, two traditional global enhancement approaches are proposed to improve the contrast of CT images of the human liver using corresponding MR images. The first approach uses context-aware two-dimensional histogram specification (HS) and morphological operations. The objective of this scheme is to improve the visibility of the organ’s anatomy to facilitate the tasks of surgeons and radiologists. The second uses 2D-HS followed by an optimization scheme to minimize the artifacts associated with histogram-based methods and simultaneously preserves the structure of the image during enhancement. In this approach, the enhanced images are analyzed from two perspectives (contrast enhancement and improvement in tumor segmentation). Both techniques have been validated on multi-modal data acquired from a hospital in Norway. Furthermore, an acceleration scheme was proposed by parallelizing the steps involved in the proposed CE approach which drastically reduced the execution time of the algorithm. The third method uses deep learning to improve the contrast of medical images using guidance from multi-modal MR images. Cycle-GAN (Generative Adversarial Network) was applied for this purpose where the corresponding high-contrast image from another modality was used as ground truth as opposed to using manually enhanced ground truth/ referenceimage. Secondly, noise is another artifact that affects the visual quality of medical images. It not only hampers the visibility of structures for clinicians who inspect these images to thoroughly understand the organ’s morphology; but it also affects the subsequent image analysis tasks. It is therefore imperative to remove noise and improve the perceptual quality of medical images. Different kinds of noise contaminate medical images. In this thesis, we proposed a method to denoise T1- weighted (T1-w) MR images contaminated with Rician noise. We exploited the complementarity-aware information in better perceptual quality multi-modal medical images for denoising purpose. In particular, the role of deep learning approach was investigated in this regard. The features from dual images were combined in a hierarchical manner to extract rich features, which are later combined in a systematic way as opposed to simple feature concatenation. The performance was validated on two public datasets both from a qualitative and quantitative perspective. Moreover, the comparison was done with single image denoising schemes on varying levels of noise.