In recent years, the use of deep learning models to detect malignant skin diseases is becoming increasingly prevalent. Studies have shown that deep learning models may outperform dermatologists in diagnosing skin diseases. However, the scarcity of skin disease images with diverse skin tones often leads to inaccurate diagnoses for people with darker skin tones, resulting in deaths that could have been prevented with early treatment. In the literature, researchers have attempted to solve this problem by synthesizing images of skin diseases on darker skin colors from those on lighter skin colors. This has resulted in an improvement of detection accuracy, but there is concern over realism, especially since the same skin disease can appear in a different color or shape on alternate skin tones. In this paper, I focus on improving the quality of the image dataset to distinguish between malignant and benign diseases for diverse skin tones. This is achieved by adopting a novel procedure utilizing Gradient-Weighted Class Activation Mapping (Grad-CAM) to generate additional images to incorporate into the initial dataset. For the models trained by the raw and refined datasets, the average detection accuracies between the 900th to 1000th epoch reached 52.25% and 64.85% with standard deviations of 3.23% and 3.93%, respectively. Improvement of detection accuracy despite limited data highlights the proposed procedure’s ability to bring fast, equitable, and inclusive healthcare a step closer, potentially saving lives.