Abstract
With the rapidly growing use of digital images in medical archival and communication, image compression technology—especially diagnosis lossless compression—plays an increasingly vital role in medical applications. This paper presents a novel diagnosis lossless compression algorithm specifically for digital mammography. Using Computer-Aided Diagnosis (CAD) technology, mammograms are segmented into the breast region, pectoral muscle, and background. A multiple arbitrary shape Regions of Interest (ROIs) coding framework is then applied: the breast region is compressed losslessly to preserve diagnostic integrity, the pectoral muscle is compressed lossily, and the background is either discarded or compressed with high loss. Experimental results demonstrate that this proposed method offers significant advantages for the storage and transmission of digital mammography.
Introduction
Digital mammography is essential for the early detection of breast cancer, requiring high-resolution imaging to identify subtle features like microcalcifications. However, these high-quality files are large (often exceeding 4MB), creating a significant burden on medical storage and communication systems.
The industry currently faces a "Goldilocks" dilemma: standard lossless compression is medically safe but inefficient (typically maxing out at a 4:1 ratio), while lossy compression (like standard JPEG) can reach 80:1 but risks "smearing" critical diagnostic details. The authors address this by recognizing that a mammogram is not a uniform block of data; it consists of a critical breast region, a less-critical pectoral muscle, and a non-informative background.
The Multiple Arbitrary Shape ROIs Framework
The core of the proposed method is a region-aware compression pipeline that utilizes CAD to segment the image into functional zones. Unlike traditional ROI methods that use simple rectangular boxes, this framework accommodates arbitrary shapes to better match human anatomy.
The process involves three primary stages:
- CAD-Based Segmentation: The system employs iterative thresholding, Watershed methods, and zonal Hough transforms to precisely outline the breast region and pectoral muscle.
- Priority Masking: Regions are assigned priority levels based on clinical value. The breast region receives the highest priority to ensure bit-for-bit accuracy.
- Region-Specific Compression: By treating the image as a collection of anatomical objects rather than a simple grid, the system "spends" the bit budget where it matters most.
Technical Implementation: ISA-DWT and SPIHT
To handle the irregular, curved shapes of anatomical regions, the researchers introduced the Integer-to-integer Shape Adaptive Discrete Wavelet Transform (ISA-DWT). Standard wavelet transforms are designed for rectangular grids; ISA-DWT allows the mathematical transform to adapt to arbitrary boundaries, preventing the artifacts and inefficiencies that occur when trying to fit curved anatomy into square blocks.
The resulting coefficients are processed using a modified Set Partitioning in Hierarchical Trees (SPIHT) algorithm. This allows the breast region to be reconstructed losslessly, while the pectoral muscle is compressed lossily based on a pre-set target distortion constraint.
Key Results
The proposed method was evaluated against the industry-standard JPEG2000 (in lossless mode) using digital mammograms with a resolution of 1760 × 2304 pixels. The results showed a significant leap in efficiency:
- JPEG2000 (Lossless): 2.838:1 compression ratio.
- Proposed Method (IWT-5/3 Wavelet): 6.412:1 compression ratio.
Qualitatively, the reconstructed images were indistinguishable from the originals in the critical breast ROI. By discarding or heavily compressing the "dead weight" of the background, the researchers achieved over twice the efficiency of standard lossless methods without sacrificing diagnostic safety.
Conclusion and Future Work
This anatomy-driven approach offers a promising path for medical image archiving. By moving toward object-based compression, significant storage gains are possible while maintaining the mathematical integrity of diagnostic tissue. Future work will focus on scaling the method for 12-bit and 16-bit modern digital mammography and conducting formal clinical trials with radiologists to ensure visual transitions between regions do not impact diagnostic performance.
Summary updated