Congratulation from Internet Conference!
Hello
BPG-BASED COMPRESSION ANALYSIS FOR DENTAL IMAGES
27.02.2026 13:21
[1. Information systems and technologies]
Author: Adamovych Andrii, PhD Student, National Aerospace University «KhAI» Kharkiv, Ukraine; Oleksandr Arkhipov, PhD Student, National Aerospace University «KhAI» Kharkiv, Ukraine; Sergii Kryvenko, Doctor of Science, Senior Researcher, National Aerospace University «KhAI» Kharkiv, Ukraine; Iryna Shulga, Candidate of Science, Associate Professor, National Aerospace University «KhAI» Kharkiv, Ukraine; Volodymyr Lukin, Doctor of Science, Professor, National Aerospace University «KhAI» Kharkiv, Ukraine
Introduction. Medical imaging is developing quickly and produces a huge amount of valuable data for different areas of diagnostics [1]. This leads to the necessity to compress the obtained images, where lossless compression is practically useless due to the low compression ratio (CR) attained [2]. Consequently, lossy compression has to be applied. However, this should be “careful” compression, providing preservation of valuable diagnostic information simultaneously with a rather large CR and fast processing. The latter means that it is not possible to carry out iterative compression to provide a desired quality or ensure the absence of just noticeable differences [3]. This is especially important for modern coders such as the Better Portable Graphics (BPG) coder (bellard.org) which, on the one hand, produces better performance compared to JPEG and JPEG2000 [2], but, on the other hand, needs more time for carrying out compression and decompression. These very obstacles explain why our attention has been paid to the BPG coder with application to a specific type of medical images – dental ones.
The paper's goal is to study the main performance characteristics of the BPG coder in more detail with application to dental images produced by the Morita system, with special attention to the dependence of CR on the parameter Q that controls compression for the BPG coder.
Materials and methods. Our study is performed on Morita system images produced in two modes, for which image size and noise characteristics are slightly different. For these images, we analyze 512×512 pixel fragments of different complexity. This is convenient from two viewpoints: 1) such image size allows visualizing an original fragment and the result of its lossy compression on a computer monitor for detecting possible differences (distortions); 2) such image size allows for comparing the coder performance in terms of provided CR and quality.
The considered image fragments are compressed and then decompressed, followed by both quantitative and visual analysis to understand the dependence of quality and CR on Q and fragment complexity. Examples of fragments are given in Fig. 1. As seen, the noise in images acquired in mode 2 is slightly more intensive.
a b Fig. 1. Examples of image fragments for Mode 1 with a simple structure image (a) and Mode 2 with a complex structure image (b)
Results and discussion. Let us first consider a standard rate/distortion curve, namely, the dependence of the peak signal-to-noise ratio (PSNR) on Q. The plots for Mode 1 for two fragments are presented in Fig. 2,a and the plots for Mode 2 are given in Fig. 2,b. As seen, PSNR values for small Q (Q < 10) are very high and guarantee that introduced distortions are invisible (this happens with high probability if PSNR exceeds 40 dB).
Then, for 9 < Q < 30, PSNR values decrease almost linearly with a reduction of about 1.4 dB for every Q increase by 1 (Q for the BPG coder are integers from 0 to 51). Distortions are not seen either (except for some particular cases [2]). After this, for Q > 29, the curves start to behave more individually depending on fragment complexity. It is considered [2] that the use of Q = 28 practically guarantees that the introduced distortions are invisible.
Fig. 3 presents the dependences of the CR on Q for the same fragments as the data in Fig. 2. They show that these dependences are monotonously increasing and the CR can reach very large values of several hundreds and even exceed 1000. However, for Q ≤ 30, CR values are considerably smaller and do not exceed 50. Meanwhile, the CR for different fragments depending on Q and image fragment complexity might differ by several (up to 5) times. This might be important if the requirement is to provide a CR not smaller than a given threshold.
Notably, it is difficult to analyze CR vs Q dependences in a representation such as in Fig. 3 (for Q < 25 it seems that the CRs for the two curves are equal to each other although this is not true).
Then, an analysis of the dependences of lg(CR) can be reasonable.
The corresponding curves are presented in Fig. 4. Analysis of the data in Fig. 4,a shows that the difference in CR values can be very large, especially in the range 20 < Q < 45. For Mode 2, CR values are, in general, smaller than in the first mode. This can be associated with the more intensive noise typical of Mode 2.
Fig. 2. Dependences of PSNR on Q for two image fragments of Mode 1 (a) and Mode 2 (b)
Fig. 3. Dependences of CR on Q for two fragments of Mode 1 (a) and Mode 2 (b)
Fig. 4. Dependences of lg(CR) on Q for two fragments of Mode 1 (a) and Mode 2 (b)
Analysis of lg(CR(Q)) shows that these dependences are quite smooth. Moreover, most likely, they can be well described by third-order polynomials for the entire range of Q variation and second-order polynomials for 10 < Q < 30, which is of primary interest for the practice of dental image compression. Therefore, procedures for providing a desired CR using a preliminary CR estimation for three or four CR values in a properly set Q seem possible based on the CR representation as 10f(Q), where f(Q) denotes the corresponding polynomial.
Conclusions. Specific features of the main dependences of BPG-based compression of dental images were considered. It is shown that they are quite typical for lossy compression. Meanwhile, there are two positive features: a) the almost linear behavior of PSNR(Q) and the almost exponential behavior of CR(Q) in the main interval of Q variation for visually lossless compression.
References:
1. Paul Suetens, Fundamentals of Medical Imaging, Cambridge University Press, 3rd Edition, 2017, ISBN 978-1107159785; https://doi.org/10.1017/9781316671849
2. S. Krivenko, V. Lukin, O. Krylova, L. Kryvenko, K. Eguiazarian, A Fast Method of Visually Lossless Compression of Dental Images, Applied. Sciences, 2021, 11(1), 135; https://doi.org/10.3390/app11010135
3. Stojanović, N.; Bondžulić, B.; Lukin, V.; Bujaković, D.; Kryvenko, S.; Ieremeiev, O. Compression Ratio as Picture-Wise Just Noticeable Difference Predictor. Mathematics 2025, 13, 1445. https://doi.org/10.3390/math13091445
Ця робота ліцензується відповідно до Creative Commons Attribution 4.0 International License
Знайшли помилку? Виділіть помилковий текст мишкою і натисніть Ctrl + Enter
Another articles in this section
Сonferences
Conference 2026
Conference 2025
Conference 2024
Conference 2023
Conference 2022
Conference 2021
