Recently, the study on model interpretability has become a hot topic in deep learning research area. Especially in the field of medical imaging, the requirements for safety are extremely high; Moreover, it is very important for the model to be able to explain. However, the existing solutions for left ventricular segmentation by convolutional neural networks are black boxes; explainable CNNs remains a challenge; explainable deep learning models has always been a task often overlooked in the entire data science lifecycle by data scientists or deep learning engineers. Because of very limited medical imaging data, most solutions currently use transfer learning methods to transfer the model which used on large-scale benchmark data sets (such as ImageNet) to fine tune medical imaging models. Consequently, a large amount of useless parameters are generated, resulting in further barrier for the model to provide a convincing explanation. This paper presents a novel method to automatically segment the Left Ventricle in Cardiac MRI by explainable convolutional neural networks with optimized size and parameters by our enhanced Deep Learning GPU Training System. It is very suitable for deployment on mobile devices. We simplify deep learning tasks on DIGITS systems, monitoring performance, and displaying the heat map of each layer of the network with advanced visualizations in real time. Our experiment results demonstrated that the proposed method is feasible and efficient.
An Explainable Convolutional Neural Networks for Automatic Segmentation of the Left Ventricle in Cardiac MRI