Abstract
Cardiovascular diseases remain the leading cause of mortality, which determines the relevance of developing robust methods for cardiac image analysis and patient monitoring. This paper presents and validates a prototype pipeline for automated preprocessing and segmentation of biomedical images, developed within the framework of a scientific project (IRN: AP05132044). The experimental protocol is implemented on the Heart Database dataset (18 patients, 3D+t MRI (three-dimensional with time dimension)), using expert endocardial masks and 36 slices (diastole/systole, 2 slices per patient). Performance evaluation was conducted using PSNR, SSIM, Dice, and IoU metrics. A comparative analysis was performed across five preprocessing modes: none, Gaussian, wavelet, NLM, and hybrid (wavelet + NLM + CLAHE). The results show that the NLM method achieves the best denoising performance (PSNR = 26.07±0.33 dB vs. 22.85±0.21 dB for the baseline, p = 1.46×10⁻¹¹, Wilcoxon test), while the highest segmentation accuracy is obtained with hybrid preprocessing (Dice = 0.681±0.176 vs. 0.636±0.153 without preprocessing, p = 0.018). An analysis of challenging cases revealed that segmentation performance is strongly influenced by endocardial contrast and geometric complexity across different phases of the cardiac cycle. The obtained results demonstrate a statistically significant impact of the preprocessing stage on downstream segmentation quality and confirm the feasibility of the integrated “preprocessing-segmentation-evaluation” pipeline as an engineering foundation for clinical frameworks. The proposed approach is designed for reproducible performance under real-world data heterogeneity and can serve as a basis for further integration into clinical information systems.