Predicting skin cancer melanoma using stacked convolutional neural networks model

Skin malignant growth has been regarded as the most widely recognized disease in the world and Malignant Melanoma is one of the deadliest diseases of skin cancer. Early prediction can be helpful to avoid the damage of this disease, however, many lab tests are required which are costly and time-consuming. Devising an automatic smart system for predicting the disease accurately and efficiently can be very helpful. Despite previous research efforts for such systems, accuracy and efficiency requirements still demand continual work to improve the performance of such systems. This study proposes a stacked convolutional neural network (CNN) model that can provide higher prediction accuracy compared to other pre-trained CNN variants. Stacked CNN uses the 2D CNN layers sequentially to process the data deeply to make accurate predictions. Data augmentation is performed for minority class data to make training data balanced and avoid the model’s over-fitting. The model is trained on red, green, and blue (RGB) features extracted from the training data. For testing the performance of the proposed approach, two public datasets MINST-HAM10000 and ISIC-2020 are used for training and validation, respectively. The proposed model outperforms other models with 0.96 and 0.73 accuracy scores on the test dataset and validation dataset, respectively. In the end, a statistical T-test is used to show the significance of the proposed approach.