Visually Lossless Compression of Windowed Images

Summary form only given. Visually lossless image compression methods aim to compress images while ensuring that compression distortions are below perceptible levels. In medical imaging applications where high bit-depth images are often displayed on lower bit-depth displays, adjustments of image brightness and contrast during display are very common. In these applications, radiologists often change the display window level and width to view different ranges of intensities within the full range to better visualize diverse types of tissue in the image. However, when an image created to be visually lossless at a particular display setting is manipulated prior to display, compression distortions that were initially invisible may become visible. Similarly, compression artifacts that would be visible in certain window settings can be invisible in others, creating opportunities for the compression algorithm to allow increased compression distortion with corresponding increases in compression ratios. In this work, the effects of window level and window width adjustments on visibility thresholds were investigated. A JPEG2000 based image compression method to achieve visually lossless compression for a given window level and width was then proposed. A validation study was performed to confirm that the images obtained using the proposed method cannot be distinguished from original windowed images. The proposed compression method was also extended to a client-server setting where the server transmits incremental data to the client to ensure visually lossless representation after adjustments to the window level and width are made at the client side. The proposed incremental compression method was compared to a reference compression system where an 8-bit image corresponding to the desired window settings is created from a 12-bit CT image first at the encoder. This image is then compressed to achieve visually lossless compression using the methods described in. When the window settings are updated, a new 8-bit image corresponding to the updated window settings is created and compressed in a visually lossless manner. A comparison of the two methods illustrate that while the reference system is more efficient when the display settings are changed only once, the proposed method is advantageous when the display settings are changed more than once, requiring only 18% of the data transmitted by the reference system at the end of seven window setting adjustments.