A Hybrid Attention U-Net and ResNet101-Based Deep Learning Approach for Left Atrium-Ventricle Segmentation in Echocardiography Images Ekokardiyografi Görüntülerinde Sol Atriyum-Ventrikül Segmentasyonu için Hibrit Attention U-Net ve ResNet101 Tabanli Derin Ögrenme Yaklasimi
34th Signal Processing and Communications Applications Conference, SIU 2026, İstanbul, Turkey, 7 - 10 July 2026, (Full Text)
- Publication Type: Conference Paper / Full Text
- Doi Number: 10.1109/siu71813.2026.11637006
- City: İstanbul
- Country: Turkey
- Keywords: artificial intelligence, cardiovascular diseases, deep learning, echocardiography, medical image processnig
- Akdeniz University Affiliated: Yes
Abstract
Cardiovascular diseases are among the leading causes of mortality worldwide; therefore, improving early diagnosis processes is of paramount importance. Although echocardiography is a widely used imaging modality for assessing the morphological structure of the heart, diagnostic accuracy is often limited by subjectivity in clinical evaluations and class imbalance issues in automated image analysis. This study proposes a hybrid deep learning approach that integrates the Attention U-Net and ResNet101 architectures. Utilising the CAMUS dataset, this approach segments the left atrium-ventricle region to enable the automatic and objective analysis of echocardiography images. Specifically, the proposed method focuses on left ventricular segmentation utilizing the CAMUS dataset. The model offers a more robust architecture than existing methods for both extracting deep features and preserving spatial resolution. Experimental results demonstrate that the proposed hybrid model outperforms conventional methods, achieving a superior Dice score of 95.51%. This high accuracy not only highlights the effective management of the class imbalance problem but also reinforces the potential of artificial intelligence-based approaches in facilitating early diagnosis.