Diagnosis of skin cancer in the early stage plays a vital role in increasing its survival rate. Advancement in artificial intelligence and deep learning technologies provides the path for easing the detection of the disease using medical image analysis. This paper studies several deep neural network (DNN) models and their comparative performance. The study was performed on three datasets HAM10000, ISIC 2020, and ISIC 2019. Due to variations in the structure of the datasets utilized, such as differing numbers of class labels and types of skin cancer, a modified dataset was curated. This modified dataset ensured uniformity by providing common class labels across all datasets, thereby enabling consistent training for the models across the entire dataset. The resultant dataset consists of 68,472 dermoscopic images. The transfer learning approach was applied to three modified models: DenseNet201, EfficientNetV2M, and ConvNeXtBase. Fine-tuning methods were employed to increase the performance of models on the modified dataset, while data augmentation methods were applied to improve class balancing. The model’s layered architecture classified the images into seven classes which were then aggregated into binary classes. The final classification was done using an ensemble of the three models used. The results favored the ensembling of the models along with the superior individual model performance of DenseNet201. Results show that ensembling the models achieved an overall accuracy of 96.5% in multiclass classification and 98.7% in binary classification.