• AF
• atrial fibrillation
• Attention U-Net
• deep learning
• fibrillazione atriale
• MRI
• segmentation
• segmentazione
• U-Net
• UNetR

Atrial fibrillation (AF) is the most common form of arrhythmia, significantly impacting global healthcare due to its association with high morbidity and mortality rates. Diagnosing AF accurately is challenging due to the com-plex anatomical features of the left atrium (LA) and its variability among patients. The emergence of magnetic resonance imaging (MRI), particularly late gadolinium-enhanced (LGE) MRI, has greatly enhanced the capacity to visualize and assess LA morphology, scar tissue, and fibrosis. LGE-MRI is instrumental in noninvasively mapping atrial anatomy, thus allowing for improved AF treatment planning and outcomes. However, identifying AF from 3D MRI (magnetic resonance imaging) images involves complex segmentation tasks that necessitate advanced computational techniques to delineate anatomical structures like the LA cavity and atrial walls accurately.

Image segmentation enables systems to identify and analyze objects, boundaries, and structures within an image, supporting more informed decision-making across multiple domains. By breaking down complex images into manageable parts, image segmentation allows computers to "understand" visual data in a manner similar to human perception.
Recent developments in deep learning, particularly convolutional neu-ral networks, have proven highly effective for segmentation tasks in medical imaging. These advancements offer solutions for accurately segmenting complex anatomical structures, such as LA, which is crucial for understand-ing cardiac health. Applying deep learning to 3D MRI segmentation enhances AF recognition and provides clinicians with detailed, accurate anatomical information. This enables personalized treatment planning, with deep learn-ing models allowing the precise delineation of structures that were previously challenging to analyze. The integration of CNN-based segmentation in clin-ical workflows supports improved diagnostic accuracy and patient outcomes by enabling efficient analysis of complex imaging data.
Deep learning techniques, particularly those utilizing CNNs and variants, have become essential for medical image segmentation tasks due to their ability to learn and represent complex features. In LA segmentation for AF detection, CNN-based architectures, like UNet, demonstrate good accuracy, especially when extended to three-dimensional data. By learning spatial rela-tionships and anatomical structures, these networks can effectively highlight areas of fibrosis, a key indicator for AF.

Traditional methods for segmenting the LA and identifying fibrosis often depend on manual or semi-automated techniques, which are labor-intensive and prone to variability due to observer bias. Segmenting these com-plex anatomical features is essential for identifying fibrosis regions, critical for planning effective AF ablation therapy. Consequently, there is a pressing need for a robust, automated solution to perform segmentation with high precision, reducing the reliance on manual intervention and improving diag-nostic consistency.

The idea for this thesis is inspired by the Medical Image Computing and Computer-Assisted Intervention (MICCAI) Society’s annual challenge. MIC-CAI, a prestigious organization, brings together researchers and clinicians to advance computational techniques in medical imaging and diagnostics. Through organized competitions, MICCAI offers datasets and benchmarks in critical fields such as MRI-based atrial segmentation, fibrosis detection, and more. By using the dataset provided by MICCAI for atrial segmen-tation, this research builds upon the challenge’s efforts to improve atrial fibrillation recognition and segmentation through deep learning techniques, thus advancing the clinical applications of MRI data.

This research aims to develop a deep learning-based segmentation frame-work specifically tailored for 3D MRI images of the LA, to enhance AF recognition through improved fibrosis detection and delineation. The spe-cific objectives are as follows:

1. To develop and evaluate various state-of-the-art 3D CNN (convolutional neural network) architectures for the automated segmentation of LA structures in MRI scans, experimenting with architectural and logical variations to optimize segmentation accuracy and performance.
2. To evaluate the accuracy of the proposed model in differentiating LA cavity and fibrosis regions.
3. To assess the potential clinical impact of the segmentation framework on the diagnosis and treatment of AF.