The Effect of Partial Fine Tuning on AlexNet for Skin Lesions Classification

Abstract

Melanoma is the deadliest form of skin cancer. The similarity between melanoma and nevus brings a challenge in identifying the presence of melanoma. Hence, automatic recognition will help dermatologists in identifying the lesion. Here, we implement transfer learning on AlexNet to classify skin lesions into melanoma and nevus. Additionally, we make a comparison between full and partial fine-tuning. Full fine-tuning is applied on all layers, while partial is applied on the 13th to 25th layers. Both models are tested using three different datasets, i.e. HAM10000, MSK, and UDA, each containing 400 skin images. The results show that the fully tuning was trained in 82 minutes and achieved 91%, 75.75%, and 84.75% classification accuracy for the test with HAM10000, MSK, and UDA datasets, respectively. Partial tuning was trained in 63 minutes and achieved 92.25%, 77.75%, and 85.75% of classification accuracy for with HAM10000, MSK, and UDA datasets, respectively. These results show that partial fine-tuning gives an insignificant improvement in terms of accuracy, but a quite significant boost in training time.