A Machine Learning Algorithm for Biomedical Images Compression Using Orthogonal Transforms

Abstract

Compression methods are increasingly used for medical images for efficient transmission and reduction of storage space. In this work, we proposed a compression scheme for colored biomedical image based on vector quantization and orthogonal transforms. The vector quantization relies on machine learning algorithm (K-Means and Splitting Method). Discrete Walsh Transform (DWaT) and Discrete Chebyshev Transform (DChT) are two orthogonal transforms considered. In a first step, the image is decomposed into sub-blocks, on each sub-block we applied the orthogonal transforms. Machine learning algorithm is used to calculate the centers of clusters and generates the codebook that is used for vector quantization on the transformed image. Huffman encoding is applied to the index resulting from the vector quantization. Parameters Such as Mean Square Error (MSE), Mean Average Error (MAE), PSNR (Peak Signal to Noise Ratio), compression ratio, compression and decompression time are analyzed. We observed that the proposed method achieves excellent performance in image quality with a reduction in storage space. Using the proposed method, we obtained a compression ratio greater than 99.50 percent. For some codebook size, we obtained a MSE and MAE equal to zero. A comparison between DWaT, DChT method and existing literature method is performed. The proposed method is really appropriate for biomedical images which cannot tolerate distortions of the reconstructed image because the slightest information on the image is important for diagnosis.

Artifact-Minimized High-Ratio Image Compression with Preserved Analysis Fidelity

“…The initial learning rate (lr) was set to 1e −2 and adopted a poly decay strategy as the same in Eq. (15), where the maximum epoch count was set to 300. Each epoch took 10 seconds, while the whole training process took 50 minutes.…”

“…2D biomedical image compression typically applies well-established lossy compression algorithms, such as JPEG 13 , JPEG2000 14 , and, more recently, machine learning 15 . For 3D volumetric image compression, video codecs such as H.264 16 , H.265 17 , and AV1 18 , can be adapted by treating one image axis, typically the z axis, as the pseudo-time axis.…”