On Completeness-aware Concept-Based Explanations in Deep Neural Networks