<title>Edge-localized image sharpening via reassignment with application to computed tomography</title>

Timothy D. Dorney, Srikrishna Bhashyam, Andrew Doran, Hyeokho Choi, Patrick Flandrin, Richard G. BaraniukProceedings of SPIE, the International Society for Optical Engineering/Proceedings of SPIE 2000 · DOI

TL;DR

This paper introduces a novel post-reconstruction image sharpening technique for computed tomography (CT) that specifically targets and enhances edge features. By leveraging the reassignment principle and wavelets, the algorithm effectively sharpens blurred edges caused by CT reconstruction limitations without altering important low-frequency textural details, leading to significantly improved image quality.

What Problem Does This Paper Address?

Computed tomography (CT) reconstruction, while powerful, faces practical limitations that blur image edges. These limitations stem from factors like a finite number of projection angles, non-uniform sampling grids, and the need to band-limit the data to avoid aliasing artifacts. This results in a loss of high-frequency information, particularly at edges, which are crucial for anatomical interpretation. Traditional filtering methods often struggle to sharpen these edges without also degrading smooth regions or introducing unwanted artifacts. The authors aim to solve this by developing a method that precisely targets and sharpens only the edge features, preserving the integrity of other image content.

Methods

The core of the proposed method is an edge-localized image sharpening algorithm that combines the reassignment principle with wavelet transforms.

  1. Undecimated Discrete Wavelet Transform (UDWT): The algorithm first decomposes the CT reconstructed image into high-pass (HP) and low-pass (LP) components using the UDWT. The UDWT is chosen for its ability to handle shift and rotation invariance to some extent and its integral edge detection property. The Haar wavelet is used for its simplicity and efficiency.
  2. Edge Detection: The high-pass (HP) coefficients are used to identify edge locations. Local maxima in the absolute value of the HP coefficients, exceeding a defined threshold, indicate the presence of an edge.
  3. Additive Reassignment: Instead of simply relocating coefficients, the paper employs an additive reassignment strategy. At identified edge locations (local maxima/minima in HP coefficients), adjacent HP coefficients are added to the central edge coefficient. This process is designed to rebuild the edge and can cause the central point to exceed its original HP value, which is necessary for sharpening.
  4. LP Signal Adjustment: Crucially, modifications to the HP signal necessitate corresponding adjustments to the LP signal to maintain consistency. The LP samples adjacent to the identified edge transition point are adjusted based on the HP reassignment. This ensures that the overall signal slope is made steeper, enhancing the edge.
  5. Two-Stage Processing: The reassignment is applied first along the rows of the image to sharpen vertical and diagonal edges, followed by an inverse UDWT. Then, the process is repeated along the columns to sharpen horizontal and the remaining diagonal edge components, followed by another inverse UDWT. This two-stage approach ensures comprehensive edge sharpening.

Datasets

The paper uses a thoracic cavity phantom for simulations. This phantom is described as a 512x512 pixel image containing three main anatomical structures: lungs, heart, and spinal column. The phantom is designed to simulate realistic anatomical variations and movements. No real-world medical datasets are used for evaluation in this paper.

Key Insights and Ideas

  • Edge-centric sharpening is key: The paper highlights that the majority of reconstruction errors in CT are concentrated at edges. This insight justifies focusing sharpening efforts solely on these regions.
  • Reassignment for localization: The reassignment principle, traditionally used for time-frequency analysis, is ingeniously adapted for spatial image sharpening. Its ability to shift features to their "center of mass" is leveraged to precisely target edge locations.
  • Additive reassignment preserves energy and enhances edges: By adding coefficients rather than just relocating them, the algorithm can push edge values beyond their original limits, effectively sharpening them while conserving energy.
  • Wavelets as edge detectors: The UDWT's inherent ability to highlight discontinuities makes it an ideal tool for identifying the precise locations where sharpening is needed.
  • Preserving smooth regions is critical: The authors emphasize that altering smooth regions is unacceptable, as these might contain subtle but important diagnostic information (e.g., tumors). Their method successfully achieves this by localizing the sharpening.

Key Results

The paper demonstrates a significant reduction in squared error (SE), particularly in edge regions.

  • Overall SE reduction: The total SE for the reconstructed image was reduced from 51.91 to 33.84, a decrease of 18.07 (34.8%).
  • Edge region improvement: The SE in edge regions was reduced by 18.01, representing a 57.0% decrease.
  • Smooth region preservation: The SE in smooth regions was altered by only 0.06 (0.3%), indicating that these areas were largely unaffected by the sharpening process.
  • Qualitative results: Figure 6 visually confirms the effectiveness of the algorithm, showing a sharpened image where edges are clearly defined, while smooth regions appear unchanged, and Moiré patterns from the original reconstruction are still visible in smooth areas.

Limitations and Future Work

  • Shift and Rotation Invariance: While the UDWT helps, the paper acknowledges that wavelets are not perfectly shift and rotation invariant. The authors address this by using small image area operations (2x2 pixels) and applying the reassignment sequentially on rows and columns. However, perfect invariance for all orientations might still be a challenge.
  • Threshold Selection: The algorithm relies on a user-defined threshold (TOL) for edge detection. The optimal choice of this threshold might depend on the specific image characteristics and could require tuning.
  • Computational Cost: While not explicitly detailed as a limitation, applying UDWT and reassignment twice (rows and columns) likely increases computational complexity compared to simpler filtering methods.

Future work could explore more robust methods for handling shift and rotation invariance, investigate adaptive thresholding techniques, and potentially optimize the computational efficiency of the algorithm. The authors also mention the possibility of studying region growth and body shift, suggesting potential applications beyond static image sharpening.

Summary updated

Abstract

Traditional filtering methods operate on the entire signal or image. In some applications, however, errors are concentrated in specific regions or features. A prime example is images generated using computed tomography. Practical implementations limit the amount of high frequency content in the reconstructed image, and consequently, edges are blurred. We introduce a new post-reconstruction edge enhancement algorithm, based on the reassignment principle and wavelets, that localizes its sharpening exclusively to edge features. Our method enhances edges without disturbing the low frequency textural details.
BibTeX
@inproceedings{dorney2000lt,
  author = {Timothy D. Dorney and Srikrishna Bhashyam and Andrew Doran and Hyeokho Choi and Patrick Flandrin and Richard G. Baraniuk},
  title = {{\<title\>Edge-localized image sharpening via reassignment with application to computed tomography\</title\>}},
  booktitle = {Proceedings of SPIE, the International Society for Optical Engineering/Proceedings of SPIE},
  year = {2000},
  doi = {10.1117/12.402443},
}
Figure