Designing Deep Neural High-Density Compression Engines for Radiology Images

Abstract

As a speciality, radiology produces the highest volume of medical images in clinical establishments compared to other commonly employed imaging modalities like digital pathology, ophthalmic imaging, etc. Archiving this massive quantity of images with large file sizes is a major problem since the costs associated with storing medical images continue to rise with an increase in cost of electronic storage devices. One of the possible solutions is to compress them for effective storage. The prime challenge is that each modality is distinctively characterized by dynamic range and resolution of the signal and its spatial and statistical distribution. Such variations in medical images are different from camera-acquired natural scene images. Thus, conventional natural image compression algorithms such as J2K and JPEG often fail to preserve the clinically relevant details present in medical images. We address this challenge by developing a modality-specific compressor and a modality-agnostic generic decompressor implemented using a deep neural network (DNN) and capable of preserving clinically relevant image information. Architecture of the DNN is obtained through design space exploration (DSE) with the objective to feature the least computational complexity at the highest compression and a target high-quality factor, thereby leading to a low power requirement for computation. The neural compressed bitstream is further compressed using the lossless Huffman encoding to obtain a variable bit length and high-density compression ($20\times -400\times$). Experimental validation is performed on X-ray, CT and MRI. Through quantitative measurement and clinical validation with a radiologist in the loop, we experimentally demonstrate our approach’s performance superiority over traditional methods like JPEG and J2K operating at matching compression factors.

Appendices

Appendix A Cost of compute for ${{\texttt{net}_\texttt{C}}}^{\mathrm {(Stage I)}}_{(d)}(\cdot )$

See Appendix Table 11.

Table 11 Number of multiply–accumulate operations and size of output tensor for each layer of the ${{\texttt{net}_\texttt{C}}}^{\mathrm {(Stage I)}}_{(d)}(\cdot )$ seed architecture with 3 digest units and an input size $1\times 128 \times 128$

Appendix B Cost of compute for ${{\texttt{net}_\texttt{D}}}^{\mathrm {(Stage I)}}_{(d)}(\cdot )$

See Appendix Fig. 22 and Table 12.

Fig. 22

Output of generic decompressor (${\texttt{net}_\texttt{D}}_{(d)}^{\mathrm {(Stage III)}}$) for different depths(d) of network. Each row of the figure corresponds to a data set with depth along the columns. The resolution of the images is mentioned within parentheses along each row. The compression factor (CF) of each image is mentioned across the row

Table 12 Number of multiply–accumulate operations and size of output tensor for each layer of the ${{\texttt{net}_\texttt{D}}}^{\mathrm {(Stage I)}}_{(d)}(\cdot )$ seed architecture with 3 digest units and an input size $64\times 8 \times 8$