Moderní metody automatické segmentace obrazu pro detekci srdečních struktur z ultrazvukových dat

Abstract

Cardiac ultrasound imaging is an essential tool for the diagnosis of cardiovascular diseases. However, accurate interpretation of echocardiographic images requires expert knowledge and is often time-consuming. The aim of this thesis was to design and experimentally validate modern deep learning methods for automatic segmentation of the left ventricle from 2D echocardiographic data using U-Net and U-Net++ architectures.

The study was conducted using the publicly available CAMUS dataset, specifically 500 images from the 2CH projection. The data were split into 70% training, 20% validation, and 10% test sets. Several network configurations were tested, varying primarily in dropout rate (0–30%) and learning rate 10^-4 to 10^-5. The training process lasted up to 50 epochs with Binary Cross-Entropy as the loss function and early stopping applied with a patience of 10 epochs.

Model performance was evaluated using the Dice score and Jaccard index. The best results were achieved with the U-Net++ architecture, reaching a Dice score of 0.93 and a Jaccard index of 0.87 on the test set. To support deployment and visualization, a web-based graphical user interface (GUI) was implemented using React, FastAPI, and PyTorch. Although the calculation of cardiac parameters (EDV, ESV, SV, EF) was performed using reference masks from the dataset, the workflow demonstrates the clinical potential of automated segmentation tools.

The results confirm that deep neural networks can serve as effective assistive tools for the analysis of echocardiographic data. Future work may include direct computation of cardiac parameters from model outputs, real-time integration, or the adoption of more advanced architectures for video sequence analysis.