Towards Self-Explainability of Deep Neural Networks with Heatmap Captioning and Large-Language Models