Dermoscopic Skin Lesion Classification using Color Features Based on Fine-Tuning of Resnet50 and Densenet121 Transfer Learning Models

Skin cancer, which stands as the most prevalent form of human malignancy, is primarily identified through visual examination. The procedure for identifying skin lesions typically starts with an initial clinical assessment, which might be followed by dermoscopic analysis, a biopsy, and a subsequent histopathological checkup. Image-based Skin lesion classification presents a significant challenge due to the subtle and intricate variations in their appearance. Deep Convolutional Neural Network (DCNN) show promising results in managing diverse and intricate tasks involving a wide array of finely detailed object categories. Skin lesions Classification is shown in this work initially using a six-layered CNN performed on the skin cancer dataset to classify seven classes of lesions with 10015 images. Exploratory data analysis of the dataset is performed to identify any outliers and to verify the balance of samples in each class. It was observed after 50 epochs, the training accuracy was 98.03% and the testing accuracy was 67.40%. To avoid the overfitting observed in the CNN approach, the pre-trained transfer learning models, Resnet50 and Densenet121 were finetuned with data augmentation to handle the imbalance of samples present in seven classes. The results obtained prove that the Densenet121 model achieves high performance compared with the other Resnet50 model after augmentation.