In recent years, skin cancer has been on the rise. It is visible, but distinguishing between a skin lesion that is malignant melanoma and a regular skin lesion is challenging. To address this issue, we propose combining models using CNN architectures to extract features and a mechanism, namely Soft Attention (SA), to enhance the performance of these original CNN models. The results show that the performance of CNN models combined with SA is improved when experimenting with two datasets - HAM-10000 dataset and integrated dataset (including HAM-10000, ISIC (2019), DERMNET, and PAD-UFES-20). On the HAM-10000 dataset, the model with Soft Attention achieves 82% Accuracy (IRv2 + SA), compared to the model without Soft Attention which achieves 69%. On the integrated dataset, our best Accuracy is 81% (IRv2 + SA) which is a 5% increase, compared to the original model (IRv2). Additionally, the Average Precision score (of IRv2 + SA) improves by 31%, compared to baseline on the HAM-10000 dataset. It can be seen that incorporating the Soft Attention mechanism into the CNN architecture is effective on both datasets in improving the performance.