Fully Convolutional Model for Variable Bit Length and Lossy High Density Compression of Mammograms

Aupendu Kar, Sri Phani Krishna Karri, Nirmalya Ghosh, Ramanathan Sethuraman, Debdoot SheetWorkshops @ CVPR 2018

TL;DR

This paper introduces a novel fully convolutional autoencoder (CAE) for high-density compression of mammograms, achieving significant compression ratios while preserving diagnostically relevant features. By combining a deep learning-based compressor with adaptive arithmetic coding, the method rivals JPEG and JPEG2000 at extreme compression levels, showing promise for efficient medical image archival and transmission.

What Problem Does This Paper Address?

Medical imaging, particularly mammography, generates vast amounts of data. The need for high-resolution digital detectors, coupled with trends towards personalized medicine and lifelong archival of medical records, creates a demand for extremely high-density image compression. Traditional compression methods like JPEG and JPEG2000, while effective for general images, can introduce artifacts that may obscure subtle diagnostic features in medical images, especially at very low bit rates. This paper aims to bridge this gap by developing a compression technique specifically tailored for mammograms, focusing on preserving critical diagnostic information at unprecedented compression ratios.

Methods

The core of the proposed solution is a Fully Convolutional Autoencoder (CAE). This architecture consists of two main parts:

  • Compressor-Decompressor (CAE): This is a deep neural network trained as an autoencoder. The "compressor" (encoder) uses a series of "down-convolution" blocks to extract key features and reduce the spatial dimensions of the input mammogram. The "decompressor" (decoder) uses corresponding "up-convolution" blocks to reconstruct the image from the compressed representation. The network is designed to learn to preserve diagnostically relevant features.
  • Adaptive Arithmetic Encoding: After the CAE generates a latent code tensor, this phase comes into play during deployment (not training). The integer-valued latent code tensor is linearized and then compressed losslessly using adaptive arithmetic coding. This step leverages the spatial redundancy inherent in mammograms to further pack the code, enabling variable bit-length compression and achieving very high compression densities.

The CAE architecture employs specific convolutional layers with varying strides and activations (ReLU, ClippedReLU) to progressively downsample and then upsample the image. The latent code tensor is represented using integer values derived from the floating-point output of the compressor, with specialized operations (float2int, int2float) to handle the non-differentiable nature of quantization during training.

Datasets

The authors utilized two publicly available digital mammogram datasets for training and evaluation:

  • CBIS-DDSM: Mammograms in this dataset are encoded at 16 bpp.
  • Dream: Mammograms in this dataset are encoded at 12 bpp.

For training, a large number of patches were extracted from these datasets. For testing, a subset of images was randomly selected. The paper also mentions evaluating the model's domain adaptability by training on one dataset and testing on the other.

Key Insights and Ideas

  • Deep Learning for Medical Image Compression: The paper demonstrates the efficacy of deep learning, specifically CAEs, in achieving high-density compression for medical images, outperforming traditional methods at extreme compression levels.
  • Feature Preservation over Pixel Accuracy: The focus is on preserving diagnostically relevant features rather than achieving perfect pixel-wise reconstruction, which is crucial for medical applications.
  • Adaptive Arithmetic Coding Synergy: Combining the CAE's learned feature representation with adaptive arithmetic coding is key to achieving both high compression ratios and variable bit-length output.
  • Flat Quality Curve: The CAE exhibits a remarkably flat image quality (PSNR/SSIM) curve across a wide range of bit rates, suggesting robustness and consistent performance.
  • Reduced Artifacts: Unlike JPEG, the CAE approach avoids introducing visual distortions, which is a significant advantage for medical image interpretation.

Key Results

  • High Compression Ratios: The proposed CAE method achieves compression factors of >300x (resulting in <0.04 bpp) while maintaining diagnostic relevance.
  • Performance vs. Benchmarks: At high compression factors (<0.1 bpp), the CAE method rivals and often outperforms JPEG and JPEG2000 in terms of PSNR and SSIM. For instance, a visual Turing test showed that radiologists could identify uncompressed images from CAE-compressed versions 50% of the time, comparable to JPEG2000 (52%).
  • Domain Adaptability: The model shows promising cross-dataset generalizability, performing well even when trained on one dataset and tested on another.
  • Parameter Efficiency: The compressor has 268,368 learnable parameters, and the decompressor has 454,724, indicating a relatively efficient model size.

Limitations and Future Work

  • Clinical Usability Study: While a visual Turing test with radiologists was conducted, a more extensive clinical usability study would be beneficial to fully validate the diagnostic utility of the compressed images.
  • Artifacts in Traditional Methods: The paper highlights the artifacts produced by JPEG, but a deeper analysis of how the CAE's learned features specifically mitigate these issues could be explored further.
  • Real-time Performance: While training time is mentioned, the real-time performance of the compression and decompression pipeline on various hardware platforms could be a subject for future work.
  • Broader Medical Imaging Modalities: The current work focuses on mammograms. Extending this approach to other medical imaging modalities (e.g., CT, MRI) could be a valuable future direction.

Summary updated

BibTeX
@inproceedings{kar2018fully,
  author = {Aupendu Kar and Sri Phani Krishna Karri and Nirmalya Ghosh and Ramanathan Sethuraman and Debdoot Sheet},
  title = {{Fully Convolutional Model for Variable Bit Length and Lossy High Density Compression of Mammograms}},
  booktitle = {Workshops @ CVPR},
  year = {2018},
}
Figure