CardioSegNet Meets XAI A Breakthrough in Left Ventricle Delineation within Cardiac Diagnostics

Abstract

To assess heart function, the precise delineation of the left ventricle (LV) using cardiac imaging is essential. While manual delineation offers precision, it is time-consuming and requires specialized expertise. Consequently, the development of automated techniques for efficient and accurate LV segmentation is imperative. In this study, we introduce CardioSegNet, an advanced deep learning CNN model designed with an emphasis on Explainable Artificial Intelligence (XAI). This model integrates multi-scale convolutional layers with attention mechanisms, spatial data, and interpretability features to ensure precise LV segmentation. Utilizing a blend of Dice loss and binary cross-entropy loss with class weight adjustments, CardioSegNet effectively manages the disparity in class distribution between LV and background pixels in the dataset. The model’s encoder-decoder structure, enriched with skip connections, further enhances its segmentation accuracy. We utilized the widely recognized CAMUS dataset, comprising short-axis cine ultrasound images, to evaluate the CardioSegNet’s performance in LV segmentation. Preliminary results indicate a Dice coefficient of 0.83, showcasing its superiority over several existing methodologies. However, it’s essential to note that these outcomes are initial and may evolve with further model optimization and parameter adjustments. These findings underscore the potential of the CardioSegNet model in achieving precise LV segmentation from cardiac images, augmented by its XAI capabilities. This model stands as a promising tool to assist medical professionals in diagnosing and monitoring cardiovascular diseases. Future endeavors might explore the model’s efficacy across diverse datasets and its real-world clinical applicability. In conclusion, CardioSegNet offers a novel approach to LV segmentation in cardiac images, blending multi-scale convolutional layers, attention mechanisms, spatial data, and XAI features, paving the way for transparent, accurate, and efficient cardiac evaluations.

Keywords: CardioSegNet, Ventricle, Delineation