Comparative analysis of the effectiveness of BPG, AGU, AVIF and HEIF compression methods for medical images corrupted by noise of two types

The subject matter is lossy compression using the BPG, AGU, AVIF, and HEIF encoders for medical images with different levels of visual complexity corrupted by additive Gaussian and Poisson noise. The goal of this study is to compare encoders regarding optimal image compression parameters and select the most suitable metric to determine the optimal operation point. The tasks considered include: selecting 512x512 grayscale test images with various degrees of visual complexity, including visually complex images rich in edges and textures, moderately complex images with edges and textures adjacent to homogeneous areas, and visually simple images consisting mainly of homogeneous areas; establishing image quality assessment metrics and evaluating their effectiveness under different encoder compression parameters; selecting one or more metrics that clearly determine the position of the optimal operation point; providing recommendations based on the results obtained for compressing medical images corrupted by additive white Gaussian and Poisson noises using four encoders to maximize the quality of the restored image to the noise-free original. The employed methods encompass image quality assessment techniques employing MSE, PSNR, and MSSIM metrics, as well as software modeling in Python without using the built-in Poisson noise generator. The results show that optimal operation points (OOPs) can be determined for all these metrics when the quality of the compressed image is better than the quality of the corresponding noisy original image, accompanied by a sufficiently high compression ratio. Moreover, achieving an appropriate balance between the compression ratio and image quality leads to partial noise reduction without noticeable information content distortion in the compressed image. This study emphasizes the importance of using appropriate metrics to assess the quality of compressed medical images and provides insight into the determination of the compression parameter Q to achieve the optimal operation point of the BPG encoder for specific images. However, the position of the OOP and its presence depend not only on the image complexity but also on the chosen encoder. Conclusions. The scientific novelty of the obtained results includes: 1) The consideration of noise models and parameter levels typical for medical imaging, namely, additive Gaussian noise of such intensity that it approximately corresponds to just noticeable differences, and signal-dependent Poisson noise; 2) The analysis of the multi-scale structural similarity index (MS-SSIM), which has not been previously explored in studies on lossy compression of noisy medical images; 3) A detailed examination of AVIF and HEIF coders to determine whether the optimal operating point (OOP) is observed for them and under which noise conditions; 4) The use of a dataset comprising ten medical images of varying visual complexity, with generalized tendencies revealed for different structural types; 5) The identification of the ability of many metrics to exhibit an OOP for images of moderate visual complexity or those dominated by homogeneous areas; 6) For Poisson noise, the demonstration of a dependence between the quality factor Q in the OOP and the average image intensity, which can be practically estimated for a given image; 7) The finding that different encoders require different approaches to determine their respective OOPs due to their distinct compression control parameters; 8) The observation that compression ratios achieved at the OOP are generally high, supporting the feasibility of using the OOP or its neighbourhood in practice.