Abstract:
"In recent years application of AI in the medical domain, especially in the medical imaging related tasks has become increasingly common. However, due to the black box nature of these AI models it is crucial to understand the reasoning behind the models predictions to provide trust, transparency, and accountability to the model predictions. To address this problem Explainable AI (XAI) has become a field of interest. Among the medical imaging, for Cardiac MRI (CMRI) are pivotal to diagnosing different cardiac related issues. There have been numerous Deep Neural Network model architectures proposed to accurately segment the different regions in the heart.
However, to provide insights into the model behavior, state of the art XAI techniques such as Gradient Weighted Class Activation Maps (GradCAM), LRP and Feature Ablation study are explored in this research for both Long Axis (LA) CMRI and Short Axis(SA) CMRI segmentation models to visually pinpoint the salient regions for the segmentation prediction. In the initial implementation, GradCAM and Feature Ablation were implemented for both the LA and SA segmentation models. In the initial implementation only the patterns/ focus areas that can be identified during different phases of the heart were analyzed and identified."
