IVE-MDNet: Intensity Value Estimation Model Combined with a Transfer Learning Approach for Melanoma Skin Cancer Diagnosis

The percentage of people affected by skin cancer has been rising in recent years. Melanoma is identified as the most dangerous and life-threatening among the three types of skin cancer since it causes more deaths than the other two over time. According to the expertise, if the melanoma case is discovered at an early stage, the death rate may be decreased. Due to the skin lesion’s complexity, it is challenging to identify melanoma at an early stage. In this work, an automated assistant system is suggested to help doctors in identifying melanoma effectively at an early stage. Because pixel intensity values include distinctive and useful features in an image, hence the pixel intensity value estimation (IVE) model is embedded with a transfer learning network for efficient Melanoma detection. Four popular transfer learning models have been analyzed to derive the best-performed model in Melanoma detection (MD). Finally, data sensitivity is analyzed on the best model. The experiment shows that overall the best performance in Recall (98.8%), F1-score (99.0%), Accuracy (99.18%), and AUC-ROC curve (97.8%) is achieved by the VGG-16 transfer learning model for 4056 data; we denoted the model as IVE-MDNet model. In the model, the network consists of 13 convolutional layers and five max-pooling layers and the learning weights are obtained using the ‘ImageNet’ dataset. A new sub-layer model is formed, which is combined with the pre-trained network to design the proposed transfer learning approach. Before feeding the image to the model, the artifacts were removed using a pre-processing technique which uses a series of precise procedures.