EuSoMII Annual Meeting 2019 Book of abstracts

摘要

Short Summary: The convolution layers of a Convolutional Neural Network (CNN) include a stride parameter that dictates how big are the steps for samplingwhen scanning the input layer to run convolutional operations. The vast majority of CNNs use a fixed stride value. This work reports experiments designedto test in a systematic way the following hypothesis: "The use of variable stride in skin lesion images will lead to improved performance (when compared tothe baseline case of comparable computational complexity, i.e., fixed stride = 2)."Purpose/Objectives: The purpose of this work is to demonstrate that by changing the stride value in CNNs depending on the position of the pixel within theimage (smaller stride for the center of the image, and larger one for pixels close to the edges), an increase in processing speed in both training andrecognition phases can be achieved without sacrificing accuracy.Methods and materials: The network used for the experiments follows a typical CNN architecture for image classification tasks, whose layers include: oneblock of Convolutional 2D (the layer in which we modify the stride parameter) + Batch Normalization + ReLU + Max Pooling layers; two blocks ofConvolutional 2D + Batch Normalization + ReLU layers (with standard stride 1); a Fully connected layer, a Softmax layer, and a Classification layer.We used the HAM10000 dataset, which consists of a large collection of dermatoscopic images, divided into seven different classes: Melanocytic nevi,Melanoma, Benign keratosis-like lesions, Basal cell carcinoma, Actinic keratoses, Vascular lesions, and Dermatofibroma. 95% of the images were randomlyselected for training and 5% used for validation.Experiments were performed in four different network configurations: fixed stride of (1, 2, or 3) or variable stride. We chose accuracy as a metric ofclassification performance and also measured training time and inference time.Results: Experimental results (Table 1) have confirmed our hypothesis. In this case, the accuracy for the variable stride case was not only higher than thebaseline case (stride 2) but also the highest overall.