Classification of Skin Lesions by Incorporating Drop-Block and Batch Normalization Layers in Representative CNN Models.

Due to increased exposure to severe environmental elements and an aging population, the prevalence of skin cancer has been progressively rising in recent decades. When the type of skin infection is not identified in the early stages, it can lead to skin cancer and even death. Low-level visual characteristics directly connected to color and shape have traditionally been employed in studies on skin-lesion identification. The performance features of deep-based models profoundly made them applicable in diverse domains. Indeed, balancing the size of the deep networks with limited and unbalanced training data to meet the real-time demand in a computationally limited platform is an ongoing challenge. Often fine-tuning the last layers is performed to tackle unbalanced and shortage of data. However in the current study, layers like global average pooling, drop-block, and batch normalization are incorporated to the base models to classify unbalanced skin-lesion data. This boosts the performance of the model compared to fine-tuning with drop-out. The effectiveness of the incorporated layers are evaluated on three popular representative models having different depths. Comparisons are provided with the results obtained with conventional fine-tuning and a four-layer CNN scratch model. The best classification accuracy obtained with the proposed approach on DenseNet-121 model is 91.54%, and the average macro-F1 score obtained is 0.774. [ABSTRACT FROM AUTHOR]