Trustworthiness and Interpretability in Deep Learning for Cardiac Ultrasound

Abstract

In recent years, deep learning techniques have achieved significant progress in medical image segmentation, especially for heart imaging. However, most applications often overlook key aspects such as model uncertainty, interpretability, and model architecture, which can pose challenges in real-time applications. This paper presents an Uncertainty-Aware and Interpretable (UAI) lightweight deep network that integrates Bayesian optimization for hyperparameter tuning. The proposed model combines real-time segmentation with uncertainty quantification and backward image reconstruction to provide interpretable results. Bayesian optimization is leveraged to maximize performance and reduce computational burden by focusing on a small-sized Convolutional Neural Network (CNN). A modified loss function incorporates epsiloninsensitive uncertainty loss with adjusted confidence interval widths, segmentation, and reconstruction metrics, ensuring that segmentation accuracy is maintained even in the presence of uncertainty, while also ensuring that data still holds meaning despite potential loss of information during forward mapping in CNN. The model is evaluated on a set of ultrasound heart images with four subsets, including different views and phases of the heart cycle (i.e., 2-Chamber (<tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{2 C H}$</tex>) and 4-Chamber (<tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{4 C H}$</tex>) views, as well as End-Diastole (ED) and End-Systole (ES) phases). Overall, UAI-CNN achieves a mean of 0.9609 in training and 0.9661 in testing for segmentation metrics, representing an improvement of approximately 4.23 % in training and 4.17 % in testing compared to UA-CNN, and an improvement of 20.47 % in training and 5.76% in testing compared to traditional CNN. These results demonstrate improved interpretability and certainty in segmentation while maintaining computational efficiency, making the model suitable for real-time clinical applications.