Robust Detection and Lossless Compression of the Foreground in Magnetic Resonance Images

Abstract

We present a collection of techniques for robust detection of the foreground (as opposed to background) in a MR volumetric image. A novel voting strategy makes our compressor more reliable. The image in which the background has been assigned a zero value is then losslessly compressed by another collection of techniques including a novel ordering of blocks to exploit an adaptive arithmetic coder. The image is segmented into few classes. Quantized data (represented by an index map and a codebook) and quantization differences are encoded separately. Correlations between slices are reduced by differential coding of the index map for consecutive slices. Correlations in the 3 dimensions are further reduced by an integer wavelet transform and by class-contextual arithmetic encoding of the quantization differences. Our compressor outperforms JPEG-LS, JPEG2000, SPIHT, and 3D-SPIHT.