TL;DR
This paper investigates why JPEG sometimes outperforms JPEG2000 for CT neuro images, attributing it to the sharp skull bone edges. They propose a segmentation-based approach to mitigate these edge effects, demonstrating improved image quality metrics.
What Problem Does This Paper Address?
The increasing volume of medical imaging data necessitates efficient storage and transmission, making lossy compression techniques like JPEG and JPEG2000 essential. While JPEG2000 generally offers better compression performance than JPEG, studies have shown JPEG can be superior for CT neuro images. This paper aims to understand this anomaly, hypothesizing that the high intensity and sharp edges of the skull bone in CT neuro images negatively impact JPEG2000's performance, leading to artifacts in diagnostically important interior regions.
Methods
The authors explore the impact of skull bone intensity on JPEG and JPEG2000 compression. Their core hypothesis is that sharp edges between bone and background/tissue create large wavelet coefficients, which JPEG2000's bitplane coding scheme prioritizes, potentially at the expense of image detail in other regions. To validate this, they propose and test a segmentation-based compression algorithm. This method involves segmenting CT neuro images into three regions (background, skull bone, and interior) using thresholding and morphological operators. Each segmented region is then compressed independently at the same ratio and later merged. They evaluate compression quality using standard metrics like PSNR and more advanced ones like SSIM, MS-SSIM, and IW-SSIM. They also compare 8-bit and 16-bit JPEG2000 compression.
Datasets
The study utilized a database of 105 CT neuro images from 7 studies, each with a slice thickness of 0.1 mm. These images were obtained from the Medical Informatics Research Centre at McMaster (MI-IRCAM), Hamilton, Canada.
Key Insights and Ideas
- Skull Bone Edges are the Culprit: The paper strongly suggests that the sharp, high-intensity edges of the skull bone are the primary reason JPEG2000 can underperform JPEG for CT neuro images. These edges create a disproportionate number of high-energy coefficients that can be poorly handled by JPEG2000's bitplane coding.
- Segmentation as a Solution: By segmenting the image and compressing regions independently, the authors demonstrate a significant improvement in objective quality metrics. This suggests that treating different tissue types (especially bone vs. soft tissue) with tailored compression strategies can be beneficial.
- SSIM is a Better Indicator: The study highlights the superiority of SSIM and its variants (MS-SSIM, IW-SSIM) over traditional PSNR for assessing the perceptual quality of compressed medical images, especially in identifying localized artifacts.
- 16-bit JPEG2000 Offers Better Quality: For CT neuro images, 16-bit JPEG2000 compression generally yields better quality (fewer edge artifacts) than 8-bit compression at the same compression ratio, although it results in larger file sizes.
Key Results
- The proposed segmentation-based compression approach consistently improved quality scores across PSNR, SSIM, MS-SSIM, and IW-SSIM for both JPEG and JPEG2000 compression. For example, in one case, segmentation improved PSNR from 32.1 to 34.5 for JPEG and from 34.2 to 36.5 for JPEG2000.
- The SSIM quality maps visually confirmed that segmentation reduced artifacts, particularly around the skull bone.
- While JPEG2000 generally achieved higher compression ratios for a given quality level on segmented images, the paper notes that JPEG sometimes still performed better overall for unsegmented CT neuro images.
- 16-bit JPEG2000 compression showed fewer "edge artifacts" compared to 8-bit compression, as evidenced by SSIM quality maps.
Limitations and Future Work
- Segmentation Method Not for Production: The authors explicitly state that the segmentation technique used in this study is not intended for general use in compressing images as it is not easily generalizable.
- Future Work: The paper indicates ongoing work to develop more robust and generalizable methods for handling troublesome sharp edges to improve JPEG and JPEG2000 performance. They also note the challenge of comparing 8-bit and 16-bit compression and the lack of established guidelines for 16-bit compression ratios in radiological studies.
Summary updated