SKIN LESION CLASSIFICATION : TRANSFORMATION-BASED APPROACH TO CONVOLUTIONAL NEURAL NETWORKS

Diagnosing malignant skin lesions early is often the difference between life or death. With the increasing accessibility of deep learning tools that have demonstrated outstanding performance for image classification, it is no surprise that there has been an extensive effort to employ neural networks in the diagnosis of skin lesions. We explore a method of late-fusion of three identical CNN’s models, trained with three different image transformations (RGB, FFT, and HSV) of the same dataset. The resulting fused accuracy of 98% is a 4% increase to each lone network.