RISE

• @petsiukRISERandomizedInput2018
• based on a stochastic approach
• Input images are iteratively altered via random noise, and the final saliency map is composed by accumulating the partial estimations
• However, its application requires much more computational power, as it needs to run hundreds of thousands of prediction cycles
• it seems that RISE is not able to highlight regions of interest of skin lesion images with the same reliability as on pictures of real-world objects
• In the first step, it creates a segmentation mask and applies it to the dermoscopic image. Secondly, it creates a structure segmentation mask to identify the structure of the dermoscopic image. After masking, the original segmented image and some nonvisual metadata are fed into a convolutional neural network for classification